{
  "article": "https://rewirebio.io/blog/genomic-foundation-models-in-the-diagnostic-lab/",
  "source_file": "posts/genomic-foundation-models-in-the-diagnostic-lab.md",
  "generated": "2026-10-19",
  "notes": [
    "references are the numbered citations in the article, numbered by first use.",
    "Each reference is an independent source except where alternate_urls lists another address for the same work.",
    "other_links are companion reviews, our own analyses, and download or installation links; download and installation links are not evidence sources.",
    "Companion reviews cite further primary studies in addition to the anchor baseline and benchmark papers cited directly here."
  ],
  "references": [
    {
      "ref": 1,
      "key": "alavi",
      "citation": "Alavi MV. [Probing Genomic Foundation Models with Splice-Variant Perturbations](https://doi.org/10.64898/2026.10.02.756289). bioRxiv (2026).",
      "title": "Probing Genomic Foundation Models with Splice-Variant Perturbations",
      "url": "https://doi.org/10.64898/2026.10.02.756289",
      "type": "preprint",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Introduction",
          "text": "- Across the **eight diagnostic tasks**, five had no published study comparing a foundation model against the established workflow baseline, two liquid-biopsy tasks had only preliminary evidence, and in the one independent threshold-transfer test on patient RNA-confirmed splice variants, **SpliceAI** achieved the highest sensitivity (88.2% at 99.1% specificity) across the five models scored at the intended variant coordinates."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Splice variants confirmed in patient RNA (Inherited disorders):** This is the one independent test on patient RNA-confirmed variants, designed as a **threshold-transfer study** rather than a multi-gene ranking. Alavi (2026, preprint, independent of all tested model teams) tuned decision thresholds exclusively on *OPA1* substitutions (49 spliceogenic and 79 benign variants from LOVD) to minimise type I errors (maximising specificity), and then transferred those fixed thresholds without re-tuning to an independent 65-gene set of 76 RNA-confirmed splice-disrupting (spliceogenic) substitutions—spanning canonical $\\pm 1,2$ sites, near-splice intronic positions ($+3$ to $+7$, $-13$), and exonic positions—plus 107 ClinVar benign intronic substitutions drawn from 46 of those 65 genes."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Table 4. Sensitivity and specificity on 76 RNA-confirmed splice-disrupting substitutions and 107 benign intronic substitutions across 65 genes when thresholds tuned for high specificity on OPA1 are transferred without re-tuning in Alavi's independent evaluation. AlphaGenome is omitted because a 1-base coordinate offset in the benchmark script shifted every queried variant by one nucleotide.*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 12. The **Table 4** threshold-transfer results plotted as sensitivity against specificity across five models with valid variant coordinates. Upper right is better. Plotted by the author from Alavi's reported values.*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 13. The splice-variant workflow supported by current evidence: predict to prioritise, then confirm in patient RNA. Figures from **Table 4**. Diagram by the author.*"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* **SpliceAI** or **Pangolin**, followed by confirmation in patient RNA (**FRASER**). When *OPA1*-tuned high-specificity thresholds were transferred across 65 genes in the only independent RNA-confirmed splice test, SpliceAI achieved the highest sensitivity (88.2% at 99.1% specificity) ahead of four newer genomic models (**Table 4**)."
        }
      ]
    },
    {
      "ref": 2,
      "key": "alaviissue",
      "citation": "Alavi MV. [Probing-genomic-foundation-models-2026 issue #2: 1-base coordinate offset in `2c8_alphagenome_splicescorer.py` and `2c8_alphagenome_junctionscorer.py`](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2). GitHub (2026).",
      "title": "Probing-genomic-foundation-models-2026 issue #2: 1-base coordinate offset in `2c8_alphagenome_splicescorer.py` and `2c8_alphagenome_junctionscorer.py`",
      "url": "https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2",
      "type": "issue-report",
      "version": "issue #2 (2026)",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Introduction",
          "text": "- Across the **eight diagnostic tasks**, five had no published study comparing a foundation model against the established workflow baseline, two liquid-biopsy tasks had only preliminary evidence, and in the one independent threshold-transfer test on patient RNA-confirmed splice variants, **SpliceAI** achieved the highest sensitivity (88.2% at 99.1% specificity) across the five models scored at the intended variant coordinates."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Table 4. Sensitivity and specificity on 76 RNA-confirmed splice-disrupting substitutions and 107 benign intronic substitutions across 65 genes when thresholds tuned for high specificity on OPA1 are transferred without re-tuning in Alavi's independent evaluation. AlphaGenome is omitted because a 1-base coordinate offset in the benchmark script shifted every queried variant by one nucleotide.*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 12. The **Table 4** threshold-transfer results plotted as sensitivity against specificity across five models with valid variant coordinates. Upper right is better. Plotted by the author from Alavi's reported values.*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        }
      ]
    },
    {
      "ref": 3,
      "key": "zhou",
      "citation": "Zhou Z, et al. [DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome](https://arxiv.org/abs/2306.15006). ICLR 2024.",
      "title": "DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome",
      "url": "https://arxiv.org/abs/2306.15006",
      "type": "conference-paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**DNABERT-2** is the compact, fast starting point for short DNA sequences. In its authors' benchmark, it \"achieves comparable performance to the state-of-the-art model with 21x fewer parameters and approximately 92x less GPU time in pre-training\", where the comparator is the 2.5-billion-parameter Nucleotide Transformer. At 117M parameters, it replaces fixed $k$-mer tokens with byte-pair encoding (BPE) that groups frequent multi-base runs into single tokens (**Figure 4**)—cutting token length roughly fivefold and enabling \"efficient fine-tuning on most consumer GPUs\". While the architecture has no hard length cap, Peng Wei, Chong Wu, and colleagues recommend keeping inputs to about 2 kb."
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "*Table 3. Estimated cloud compute cost to pretrain five DNA foundation models from scratch, multiplying author-reported GPU-hours by 13 October 2026 Runpod secure-cloud on-demand prices (A100 80 GB \\$1.79/h, H100 \\$3.99/h; DNABERT-2's RTX 2080 Ti is priced at its faster RTX 3090 successor at \\$0.50/h; small HyenaDNA's A100 40 GB is priced at the 80 GB rate). My calculation, not what the authors paid; storage, networking, failed runs, and staff time are excluded. Sources: Nucleotide Transformer Methods; NTv3 Table 5; HyenaDNA Section 5, Appendix A.1 and Table A.2; DNABERT-2 Appendix A.3. (HyenaDNA Table A.2 lists a higher 215,000 GPU-hours for NT 2.5B; this table uses the Nucleotide Transformer paper's hardware figures.)*"
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "Look at the gap in **Table 3** and **Figure 7**: pretraining DNABERT-2 took 32 times fewer raw GPU-hours than Nucleotide Transformer 2.5B (about \\$1,300 vs. \\$154,000 at current cloud rates), yet performed comparably on the DNABERT-2 benchmark. Within the Nucleotide Transformer family itself, the 250M NT-v2 beat the 2.5B model at one-tenth the parameter count. Smarter tokenization and updated training recipes matter far more than raw parameter count—and smaller models are much cheaper for your lab to run and adapt."
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "*Figure 7. Pretraining GPU-hours from **Table 3**, derived from author-reported hardware and time, on a log scale, with my cloud cost estimates. DNABERT-2 used 32 times fewer GPU-hours than Nucleotide Transformer 2.5B and performed comparably on the DNABERT-2 benchmark.*"
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| DNABERT-2 | Fine-tunes \"on most consumer GPUs\" |"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Short sequences (up to ~2–12 kb):* **DNABERT-2** (117M) or **NT-v2** (50M–500M) using frozen embeddings, mean pooling, and a simple classifier probe (**Route 2**)."
        }
      ]
    },
    {
      "ref": 4,
      "key": "feng",
      "citation": "Feng H, et al. [Benchmarking DNA foundation models for genomic and genetic tasks](https://doi.org/10.1038/s41467-025-65823-8). Nature Communications 16, 10780 (2025).",
      "title": "Benchmarking DNA foundation models for genomic and genetic tasks",
      "url": "https://doi.org/10.1038/s41467-025-65823-8",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**DNABERT-2** is the compact, fast starting point for short DNA sequences. In its authors' benchmark, it \"achieves comparable performance to the state-of-the-art model with 21x fewer parameters and approximately 92x less GPU time in pre-training\", where the comparator is the 2.5-billion-parameter Nucleotide Transformer. At 117M parameters, it replaces fixed $k$-mer tokens with byte-pair encoding (BPE) that groups frequent multi-base runs into single tokens (**Figure 4**)—cutting token length roughly fivefold and enabling \"efficient fine-tuning on most consumer GPUs\". While the architecture has no hard length cap, Peng Wei, Chong Wu, and colleagues recommend keeping inputs to about 2 kb."
        },
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Caduceus** builds DNA's reverse-complement symmetry directly into a state-space architecture that reads both strands while scaling linearly with sequence length (0.5M to 7.7M parameters, up to 131 kb). In Feng and colleagues' independent benchmark, **Caduceus-Ph** was the strongest model overall across human genome classification tasks."
        },
        {
          "section": "Benchmarks That Actually Matter: Research vs. Clinical Reality",
          "text": "Every model paper highlights benchmarks where its own architecture excels. Before trusting a headline score, examine two independent research benchmarks—**Feng and colleagues** and **GenBench**—while keeping their scope in mind: both evaluate curated research sequence-classification datasets rather than clinical diagnostic endpoints, Feng and colleagues test frozen embeddings only (without fine-tuning or task-specific baselines), and model developers (including AlphaGenome's lead author) dispute how such multi-task suites construct and aggregate tasks into overall rankings. Where these suites are genuinely useful is in isolating design choices across shared splits—such as token pooling on frozen embeddings (confirmed by Feng and colleagues co-author Chong Wu) and how core layer families scale with input length—before you compare models against actual clinical diagnostic workflows."
        },
        {
          "section": "The Pooling Strategy Secret",
          "text": "In their 2025 *Nature Communications* benchmark comparing DNABERT-2, NT-v2, HyenaDNA, Caduceus, and GROVER with frozen embeddings, Peng Wei and Chong Wu's groups at MD Anderson found a simple, consistent rule: *\"Mean token embedding consistently and significantly improves sequence classification performance, outperforming other pooling strategies.\"* Averaging the vectors across all tokens (**mean pooling**, **Figure 8**) significantly beat summary-token and max pooling in **35 to 42 of 52 binary classification datasets**, depending on the model. Start with mean pooling—averaging the hidden states across the token dimension—and verify alternatives on your validation split if you fine-tune."
        },
        {
          "section": "The Pooling Strategy Secret",
          "text": "*Figure 8. Three ways to pool per-token vectors into one sequence vector, shown on toy numbers. With frozen embeddings, Feng and colleagues found mean pooling significantly better than the other two in 35 to 42 of 52 binary classification datasets, depending on the model. Diagram by the author.*"
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "Two practical lessons emerge from the benchmarks. First, on short-range sequence tasks, dense-attention models consistently outperform convolution-based models in GenBench, whereas HyenaDNA catches up and leads on tasks that require long-range context (such as species classification at 450 kb to 1 Mb, where HyenaDNA reached 99.5% accuracy and standard Transformers could not be trained past 32 kb on the authors' hardware). Second, do not assume a fixed sequence length where attention suddenly hits a wall: when Feng and colleagues measured single-sequence inference (batch size 1) on an A100 GPU, none of the attention models showed a sharp quadratic runtime blow-up across their tested window lengths. Always time your actual input lengths on your own GPU."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Short sequences (up to ~2–12 kb):* **DNABERT-2** (117M) or **NT-v2** (50M–500M) using frozen embeddings, mean pooling, and a simple classifier probe (**Route 2**)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Medium-to-long sequences (10 kb to 100 kb):* **Caduceus-Ph** (up to 131 kb), **NTv3**, or a **HyenaDNA** checkpoint pretrained on 2 to 4 times your input length. Verify on validation data that wider context actually improves accuracy over a 2 kb window."
        }
      ]
    },
    {
      "ref": 5,
      "key": "dallatorre",
      "citation": "Dalla-Torre H, et al. [Nucleotide Transformer: building and evaluating robust foundation models for human genomics](https://doi.org/10.1038/s41592-024-02523-z). Nature Methods 22, 287 to 297 (2025).",
      "title": "Nucleotide Transformer: building and evaluating robust foundation models for human genomics",
      "url": "https://doi.org/10.1038/s41592-024-02523-z",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Nucleotide Transformer (NT)** spans three generations. The first generation scaled up to 2.5B parameters across 3,202 human genomes and 850 genomes from other species. Crucially, bigger was not better: in the authors' own benchmark, the second-generation 250M-parameter **NT-v2** \"attained the best performance on our benchmark (average MCC of 0.769) while being ten-times smaller than the 2.5B-parameter model\" (where Matthews correlation coefficient, MCC, measures class agreement from $-1$ to $+1$). NT-v2 covers 50M to 500M parameters with a 12 kb window. The third generation, **NTv3** (December 2025, 8M to 650M parameters), is designed to run full 1 Mb windows on a single A100 or H100 GPU."
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "*Table 3. Estimated cloud compute cost to pretrain five DNA foundation models from scratch, multiplying author-reported GPU-hours by 13 October 2026 Runpod secure-cloud on-demand prices (A100 80 GB \\$1.79/h, H100 \\$3.99/h; DNABERT-2's RTX 2080 Ti is priced at its faster RTX 3090 successor at \\$0.50/h; small HyenaDNA's A100 40 GB is priced at the 80 GB rate). My calculation, not what the authors paid; storage, networking, failed runs, and staff time are excluded. Sources: Nucleotide Transformer Methods; NTv3 Table 5; HyenaDNA Section 5, Appendix A.1 and Table A.2; DNABERT-2 Appendix A.3. (HyenaDNA Table A.2 lists a higher 215,000 GPU-hours for NT 2.5B; this table uses the Nucleotide Transformer paper's hardware figures.)*"
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "Look at the gap in **Table 3** and **Figure 7**: pretraining DNABERT-2 took 32 times fewer raw GPU-hours than Nucleotide Transformer 2.5B (about \\$1,300 vs. \\$154,000 at current cloud rates), yet performed comparably on the DNABERT-2 benchmark. Within the Nucleotide Transformer family itself, the 250M NT-v2 beat the 2.5B model at one-tenth the parameter count. Smarter tokenization and updated training recipes matter far more than raw parameter count—and smaller models are much cheaper for your lab to run and adapt."
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| Nucleotide Transformer 2.5B | Weights take 9.5 GB; fine-tuning took 50 minutes on eight A100s (Methods) |"
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "Fine-tuning a compact or mid-sized model is far faster than pretraining: the Nucleotide Transformer team fine-tuned their 500M model in 20 minutes on average and their 2.5B model in 50 minutes on one node of eight A100s, and Hegde and colleagues independently fine-tuned NT 500M on roughly 89,000 expression quantitative trait locus (eQTL) variants in about 7.5 hours per five-epoch run on two NVIDIA TITAN Xp GPUs. When running large models such as Evo 2 7B for zero-shot likelihood scoring across cloud GPUs, note that low-precision `bfloat16` and `FP8` reductions are not associative across different GPU models: while two H100s on different clouds produced bit-identical *BRCA1* scores, scoring the exact same container on an L40S shifted per-variant scores by a median of $1.2 \\times 10^{-4}$ and flipped binary calls on 1.8% to 4.9% of variants across fixed thresholds—a drift eliminated by running in full 32-bit `FP32` precision or locking the pipeline to a single validated GPU product ([classification flips](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CLASSIFICATION-FLIPS.md); [full cloud operations and reproducibility study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Start with Frozen Embeddings (Route 2) When:",
          "text": "You are building a custom sequence classifier and want a fast, inexpensive test of whether a foundation model adds signal over a classical baseline. Pass your sequences through the frozen model, apply **mean pooling** across the token vectors, and fit a simple classifier (such as logistic regression or a small multi-layer perceptron) on top. Remember that this **probe** is still a supervised predictor trained on your labels—it is not zero-shot prediction. In the Nucleotide Transformer paper, trained probes on frozen embeddings matched a task-specific supervised baseline (BPNet) on 5 of 18 tasks and beat it on 8 (Results; Supplementary Table 6). Always compare your probe against a simple non-foundation-model baseline on the exact same train/validation/test split."
        },
        {
          "section": "Fine-Tune (Route 3) When:",
          "text": "In the Nucleotide Transformer benchmark, fine-tuned models matched 6 of 18 supervised baselines and beat the remaining 12 (Fig. 2a,b; Supplementary Tables 7 and 9). Within the first-generation Nucleotide Transformer models, larger checkpoints pretrained on more diverse genomes fine-tuned better—though the second-generation 250M NT-v2 subsequently beat the 2.5B model at one-tenth the size."
        },
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "- **(IA)3** (used by Nucleotide Transformer) scales inner activations with learned vectors, adding trainable parameters equal to only about **0.1% of the model's parameters** (Results and Methods, Fine-tuning)."
        },
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "*Figure 16. Which weights are trained in full fine-tuning, LoRA and (IA)3. Matrix sizes are schematic. Diagram by the author.*"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Short sequences (up to ~2–12 kb):* **DNABERT-2** (117M) or **NT-v2** (50M–500M) using frozen embeddings, mean pooling, and a simple classifier probe (**Route 2**)."
        }
      ]
    },
    {
      "ref": 6,
      "key": "ntv3",
      "citation": "Boshar S, et al. [A foundational model for joint sequence-function multi-species modeling at scale for long-range genomic prediction](https://doi.org/10.64898/2025.12.22.695963). bioRxiv (2025). Introduces NTv3.",
      "title": "A foundational model for joint sequence-function multi-species modeling at scale for long-range genomic prediction",
      "url": "https://doi.org/10.64898/2025.12.22.695963",
      "type": "preprint",
      "version": "v3",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Nucleotide Transformer (NT)** spans three generations. The first generation scaled up to 2.5B parameters across 3,202 human genomes and 850 genomes from other species. Crucially, bigger was not better: in the authors' own benchmark, the second-generation 250M-parameter **NT-v2** \"attained the best performance on our benchmark (average MCC of 0.769) while being ten-times smaller than the 2.5B-parameter model\" (where Matthews correlation coefficient, MCC, measures class agreement from $-1$ to $+1$). NT-v2 covers 50M to 500M parameters with a 12 kb window. The third generation, **NTv3** (December 2025, 8M to 650M parameters), is designed to run full 1 Mb windows on a single A100 or H100 GPU."
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "*Table 3. Estimated cloud compute cost to pretrain five DNA foundation models from scratch, multiplying author-reported GPU-hours by 13 October 2026 Runpod secure-cloud on-demand prices (A100 80 GB \\$1.79/h, H100 \\$3.99/h; DNABERT-2's RTX 2080 Ti is priced at its faster RTX 3090 successor at \\$0.50/h; small HyenaDNA's A100 40 GB is priced at the 80 GB rate). My calculation, not what the authors paid; storage, networking, failed runs, and staff time are excluded. Sources: Nucleotide Transformer Methods; NTv3 Table 5; HyenaDNA Section 5, Appendix A.1 and Table A.2; DNABERT-2 Appendix A.3. (HyenaDNA Table A.2 lists a higher 215,000 GPU-hours for NT 2.5B; this table uses the Nucleotide Transformer paper's hardware figures.)*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "| NTv3 650M (post-trained) | 85.5% | 86.9% |"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| NTv3 | Designed to run inference on 1 Mb windows on one A100 or H100; in mixed precision, the 8M model runs tens of kb on a laptop CPU |"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Medium-to-long sequences (10 kb to 100 kb):* **Caduceus-Ph** (up to 131 kb), **NTv3**, or a **HyenaDNA** checkpoint pretrained on 2 to 4 times your input length. Verify on validation data that wider context actually improves accuracy over a 2 kb window."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Very long sequences (>100 kb) or sequence generation:* **NTv3** (designed for 1 Mb on one A100/H100) or **Evo 2** (noting that Evo 2 7B needed more than one 80 GB GPU at 131 kb in standard single-GPU inference)."
        }
      ]
    },
    {
      "ref": 7,
      "key": "hyenadna",
      "citation": "Nguyen E, et al. [HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution](https://arxiv.org/abs/2306.15794). NeurIPS 2023.",
      "title": "HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution",
      "url": "https://arxiv.org/abs/2306.15794",
      "type": "conference-paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**HyenaDNA** was built for extreme sequence length. It handles \"context lengths of up to 1 million tokens at the single nucleotide-level\", \"an up to 500x increase over previous dense attention-based models\". At a sequence length of 1M tokens, it trains up to 160 times faster than a Transformer comparator. Not 1.6 times. One hundred and sixty (note that this is training speed in the authors' setup, not an inference ratio). The checkpoints are tiny, from 0.44M to 6.6M parameters."
        },
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**A caveat on HyenaDNA.** Long context does not guarantee higher accuracy on short-range biology. In independent evaluations, attention models beat HyenaDNA on short-range tasks, and HyenaDNA gained an edge only when tasks required long context. On BreCol, a 16S microbiome cancer-detection benchmark, HyenaDNA also trailed classical machine-learning methods on held-out data. When you do need long windows, pick the checkpoint carefully: the authors found (Appendix A.1) that a pretrained window 2 to 4 times longer than your downstream input works best."
        },
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "*Table 3. Estimated cloud compute cost to pretrain five DNA foundation models from scratch, multiplying author-reported GPU-hours by 13 October 2026 Runpod secure-cloud on-demand prices (A100 80 GB \\$1.79/h, H100 \\$3.99/h; DNABERT-2's RTX 2080 Ti is priced at its faster RTX 3090 successor at \\$0.50/h; small HyenaDNA's A100 40 GB is priced at the 80 GB rate). My calculation, not what the authors paid; storage, networking, failed runs, and staff time are excluded. Sources: Nucleotide Transformer Methods; NTv3 Table 5; HyenaDNA Section 5, Appendix A.1 and Table A.2; DNABERT-2 Appendix A.3. (HyenaDNA Table A.2 lists a higher 215,000 GPU-hours for NT 2.5B; this table uses the Nucleotide Transformer paper's hardware figures.)*"
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "- **Long convolutions (`HyenaDNA`):** replace dense attention with implicit long convolutions evaluated via fast Fourier transforms in $O(L \\log L)$ time."
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "*Figure 9. How dense attention, long convolution and state-space layers connect positions in a sequence, simplified. **Figure 10** shows how the resulting work grows with input length. Diagram by the author from the HyenaDNA, Caduceus and GenBench papers.*"
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "Two practical lessons emerge from the benchmarks. First, on short-range sequence tasks, dense-attention models consistently outperform convolution-based models in GenBench, whereas HyenaDNA catches up and leads on tasks that require long-range context (such as species classification at 450 kb to 1 Mb, where HyenaDNA reached 99.5% accuracy and standard Transformers could not be trained past 32 kb on the authors' hardware). Second, do not assume a fixed sequence length where attention suddenly hits a wall: when Feng and colleagues measured single-sequence inference (batch size 1) on an A100 GPU, none of the attention models showed a sharp quadratic runtime blow-up across their tested window lengths. Always time your actual input lengths on your own GPU."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Medium-to-long sequences (10 kb to 100 kb):* **Caduceus-Ph** (up to 131 kb), **NTv3**, or a **HyenaDNA** checkpoint pretrained on 2 to 4 times your input length. Verify on validation data that wider context actually improves accuracy over a 2 kb window."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "*Figure 17. Starting models for a custom sequence classifier by input length, from the decision matrix above, with the HyenaDNA checkpoint rule. Diagram by the author.*"
        },
        {
          "section": "What These Models Can't Do (Yet)",
          "text": "2. **Context limits persist in practice.** Even a 1-million-base window is only about **0.03% of a 3.2-billion-base human genome**, and in standard single-GPU Evo 2 7B runs, a single 80 GB H100 ran out of memory at 131 kb (**Figure 19**). Chromosome-scale 3D regulation remains beyond single-GPU language model windows."
        }
      ]
    },
    {
      "ref": 8,
      "key": "genbench",
      "citation": "Liu Z, et al. [GenBench: A Benchmarking Suite for Systematic Evaluation of Genomic Foundation Models](https://arxiv.org/abs/2406.01627). arXiv (2024).",
      "title": "GenBench: A Benchmarking Suite for Systematic Evaluation of Genomic Foundation Models",
      "url": "https://arxiv.org/abs/2406.01627",
      "type": "preprint",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**A caveat on HyenaDNA.** Long context does not guarantee higher accuracy on short-range biology. In independent evaluations, attention models beat HyenaDNA on short-range tasks, and HyenaDNA gained an edge only when tasks required long context. On BreCol, a 16S microbiome cancer-detection benchmark, HyenaDNA also trailed classical machine-learning methods on held-out data. When you do need long windows, pick the checkpoint carefully: the authors found (Appendix A.1) that a pretrained window 2 to 4 times longer than your downstream input works best."
        },
        {
          "section": "Benchmarks That Actually Matter: Research vs. Clinical Reality",
          "text": "Every model paper highlights benchmarks where its own architecture excels. Before trusting a headline score, examine two independent research benchmarks—**Feng and colleagues** and **GenBench**—while keeping their scope in mind: both evaluate curated research sequence-classification datasets rather than clinical diagnostic endpoints, Feng and colleagues test frozen embeddings only (without fine-tuning or task-specific baselines), and model developers (including AlphaGenome's lead author) dispute how such multi-task suites construct and aggregate tasks into overall rankings. Where these suites are genuinely useful is in isolating design choices across shared splits—such as token pooling on frozen embeddings (confirmed by Feng and colleagues co-author Chong Wu) and how core layer families scale with input length—before you compare models against actual clinical diagnostic workflows."
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "GenBench evaluates ten genomic models across 43 datasets, grouping them by how their core layers scale with sequence length $L$ (**Figure 9**):"
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "- **State-space models (`Caduceus`):** scan along the sequence in linear $O(L)$ time."
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "*Figure 9. How dense attention, long convolution and state-space layers connect positions in a sequence, simplified. **Figure 10** shows how the resulting work grows with input length. Diagram by the author from the HyenaDNA, Caduceus and GenBench papers.*"
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "Two practical lessons emerge from the benchmarks. First, on short-range sequence tasks, dense-attention models consistently outperform convolution-based models in GenBench, whereas HyenaDNA catches up and leads on tasks that require long-range context (such as species classification at 450 kb to 1 Mb, where HyenaDNA reached 99.5% accuracy and standard Transformers could not be trained past 32 kb on the authors' hardware). Second, do not assume a fixed sequence length where attention suddenly hits a wall: when Feng and colleagues measured single-sequence inference (batch size 1) on an A100 GPU, none of the attention models showed a sharp quadratic runtime blow-up across their tested window lengths. Always time your actual input lengths on your own GPU."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As GenBench highlights, protein-coding exons make up only about **1.5% of the human genome**, while the remaining **98.5% is noncoding** sequence (**Figure 11**) containing promoters, splice regions, enhancers, and structural elements. Models and baselines behave very differently across those regions—and even more differently when you move from curated research benchmarks to clinical diagnostic pipelines."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 11. The human genome by function, using GenBench's 1.5% coding figure. Diagram by the author.*"
        }
      ]
    },
    {
      "ref": 9,
      "key": "brecol",
      "citation": "Dick JM. [BreCol: Benchmarking Classical and Deep-Learning Methods for Microbiome-Based Cancer Detection](https://arxiv.org/abs/2609.27207). arXiv (2026).",
      "title": "BreCol: Benchmarking Classical and Deep-Learning Methods for Microbiome-Based Cancer Detection",
      "url": "https://arxiv.org/abs/2609.27207",
      "type": "preprint",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**A caveat on HyenaDNA.** Long context does not guarantee higher accuracy on short-range biology. In independent evaluations, attention models beat HyenaDNA on short-range tasks, and HyenaDNA gained an edge only when tasks required long context. On BreCol, a 16S microbiome cancer-detection benchmark, HyenaDNA also trailed classical machine-learning methods on held-out data. When you do need long windows, pick the checkpoint carefully: the authors found (Appendix A.1) that a pretrained window 2 to 4 times longer than your downstream input works best."
        }
      ]
    },
    {
      "ref": 10,
      "key": "caduceus",
      "citation": "Schiff Y, et al. [Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling](https://arxiv.org/abs/2403.03234). ICML 2024.",
      "title": "Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling",
      "url": "https://arxiv.org/abs/2403.03234",
      "type": "conference-paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Caduceus** builds DNA's reverse-complement symmetry directly into a state-space architecture that reads both strands while scaling linearly with sequence length (0.5M to 7.7M parameters, up to 131 kb). In Feng and colleagues' independent benchmark, **Caduceus-Ph** was the strongest model overall across human genome classification tasks."
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "- **State-space models (`Caduceus`):** scan along the sequence in linear $O(L)$ time."
        },
        {
          "section": "Attention vs. Convolution vs. State-Space Trade-offs",
          "text": "*Figure 9. How dense attention, long convolution and state-space layers connect positions in a sequence, simplified. **Figure 10** shows how the resulting work grows with input length. Diagram by the author from the HyenaDNA, Caduceus and GenBench papers.*"
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| Caduceus | Experiments ran on GPUs from an RTX 3090 to an A100 (Appendix F) |"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Medium-to-long sequences (10 kb to 100 kb):* **Caduceus-Ph** (up to 131 kb), **NTv3**, or a **HyenaDNA** checkpoint pretrained on 2 to 4 times your input length. Verify on validation data that wider context actually improves accuracy over a 2 kb window."
        }
      ]
    },
    {
      "ref": 11,
      "key": "caduceushf",
      "citation": "Kuleshov group. [Caduceus checkpoints on Hugging Face](https://huggingface.co/kuleshov-group); the 131k checkpoints have 7,725,312 parameters in their model metadata, for example [caduceus-ph_seqlen-131k_d_model-256_n_layer-16](https://huggingface.co/kuleshov-group/caduceus-ph_seqlen-131k_d_model-256_n_layer-16).",
      "title": "Caduceus checkpoints on Hugging Face",
      "url": "https://huggingface.co/kuleshov-group",
      "type": "paper",
      "version": null,
      "alternate_urls": [
        "https://huggingface.co/kuleshov-group/caduceus-ph_seqlen-131k_d_model-256_n_layer-16"
      ],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Caduceus** builds DNA's reverse-complement symmetry directly into a state-space architecture that reads both strands while scaling linearly with sequence length (0.5M to 7.7M parameters, up to 131 kb). In Feng and colleagues' independent benchmark, **Caduceus-Ph** was the strongest model overall across human genome classification tasks."
        }
      ]
    },
    {
      "ref": 12,
      "key": "evo",
      "citation": "Nguyen E, et al. [Sequence modeling and design from molecular to genome scale with Evo](https://doi.org/10.1126/science.ado9336). Science 386, eado9336 (2024).",
      "title": "Sequence modeling and design from molecular to genome scale with Evo",
      "url": "https://doi.org/10.1126/science.ado9336",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Evo and Evo 2** are large generative models built on the StripedHyena architecture (which I examine in detail below). While the original 7B **Evo** was pretrained only on prokaryotic and phage sequences (131 kb window), **Evo 2** trained on more than 8.8 trillion nucleotides across all domains of life in 7B, 20B, and 40B sizes with windows up to 1 Mb. (A 1B checkpoint with an 8 kb window also exists, but the Evo 2 paper advises that it \"should be avoided owing to overall weaker performance\".)"
        },
        {
          "section": "What Scale Actually Buys You: Inside Evo and Evo 2",
          "text": "- **Hybrid StripedHyena architecture:** To avoid the $O(L^2)$ scaling of pure dense attention while preserving the precision attention brings to local motifs, the original 7B **Evo** interleaves **29 Hyena convolutional layers with 3 rotary-attention layers (10%)** (**Figure 18**, panel B)."
        },
        {
          "section": "What Scale Actually Buys You: Inside Evo and Evo 2",
          "text": "- **From single genes to multi-gene systems:** Pretrained on 300 billion nucleotide tokens from over 80,000 bacterial and archaeal genomes plus phage and plasmid sequences (1,000 times HyenaDNA's parameter count and 100 times its training data), Evo learned how genes co-occur across whole operons and mobile elements—allowing it to generate coding-rich sequences up to 650 kb long and outperform specialised RNA language models on zero-shot noncoding RNA fitness prediction."
        },
        {
          "section": "What Scale Actually Buys You: Inside Evo and Evo 2",
          "text": "- **Scaling across all domains of life in Evo 2:** Because Evo saw only prokaryotic and phage DNA, its authors noted that its zero-shot predictions did not transfer well to human genes. **Evo 2** scaled pretraining to more than **8.8 trillion nucleotides** across bacteria, archaea, organelles, and eukaryotes (9.3 trillion tokens for the 40B model) with context windows up to 1 Mb. That cross-species, multi-domain training is what allows a single unsupervised model to score coding, noncoding, and splice-site mutations in *BRCA1* without ever seeing a clinical variant label during training—even when task-specific supervised tools like REVEL, CADD v1.7, and SpliceAI still hold the edge on standard diagnostic benchmarks."
        },
        {
          "section": "What Scale Actually Buys You: Inside Evo and Evo 2",
          "text": "*Figure 18. Evo's design and scaling study. (B) StripedHyena interleaves Hyena convolution layers with rotary attention layers. (F, G) In the authors' compute-matched comparison on DNA, StripedHyena and Hyena reached lower evaluation perplexity than Transformer++ and Mamba. Source: Nguyen E, et al., [Sequence modeling and design from molecular to genome scale with Evo](https://doi.org/10.1101/2024.02.27.582234), bioRxiv (2024), Figure 1, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/); published in Science.*"
        },
        {
          "section": "What These Models Can't Do (Yet)",
          "text": "4. **Pretraining genomes dictate what the model knows.** The original Evo was trained strictly on prokaryotes and phages, limiting its human performance; always check which organisms and genomic elements a checkpoint saw during pretraining."
        }
      ]
    },
    {
      "ref": 13,
      "key": "brixi",
      "citation": "Brixi G, et al. [Genome modelling and design across all domains of life with Evo 2](https://doi.org/10.1038/s41586-026-10176-5). Nature 652, 1349 to 1361 (2026).",
      "title": "Genome modelling and design across all domains of life with Evo 2",
      "url": "https://doi.org/10.1038/s41586-026-10176-5",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Evo and Evo 2** are large generative models built on the StripedHyena architecture (which I examine in detail below). While the original 7B **Evo** was pretrained only on prokaryotic and phage sequences (131 kb window), **Evo 2** trained on more than 8.8 trillion nucleotides across all domains of life in 7B, 20B, and 40B sizes with windows up to 1 Mb. (A 1B checkpoint with an 8 kb window also exists, but the Evo 2 paper advises that it \"should be avoided owing to overall weaker performance\".)"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "| Evo 2 40B | 71.1% | 100% |"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *To test a foundation model locally:* Run **Evo 2 7B** zero-shot likelihood scoring (**Route 1**) on a 48 GB or 80 GB GPU with 8 kb windows, or evaluate a supervised embedding probe (**Route 2**) on a held-out split ([cloud setup and cost guide](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Very long sequences (>100 kb) or sequence generation:* **NTv3** (designed for 1 Mb on one A100/H100) or **Evo 2** (noting that Evo 2 7B needed more than one 80 GB GPU at 131 kb in standard single-GPU inference)."
        },
        {
          "section": "What Scale Actually Buys You: Inside Evo and Evo 2",
          "text": "- **Scaling across all domains of life in Evo 2:** Because Evo saw only prokaryotic and phage DNA, its authors noted that its zero-shot predictions did not transfer well to human genes. **Evo 2** scaled pretraining to more than **8.8 trillion nucleotides** across bacteria, archaea, organelles, and eukaryotes (9.3 trillion tokens for the 40B model) with context windows up to 1 Mb. That cross-species, multi-domain training is what allows a single unsupervised model to score coding, noncoding, and splice-site mutations in *BRCA1* without ever seeing a clinical variant label during training—even when task-specific supervised tools like REVEL, CADD v1.7, and SpliceAI still hold the edge on standard diagnostic benchmarks."
        }
      ]
    },
    {
      "ref": 14,
      "key": "evo2release",
      "citation": "Arc Institute. [Evo 2 release v0.5.0: 20B model and hardware notes](https://github.com/ArcInstitute/evo2/releases/tag/v0.5.0). GitHub (2026).",
      "title": "Evo 2 release v0.5.0: 20B model and hardware notes",
      "url": "https://github.com/ArcInstitute/evo2/releases/tag/v0.5.0",
      "type": "release-notes",
      "version": "v0.5.0",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Evo and Evo 2** are large generative models built on the StripedHyena architecture (which I examine in detail below). While the original 7B **Evo** was pretrained only on prokaryotic and phage sequences (131 kb window), **Evo 2** trained on more than 8.8 trillion nucleotides across all domains of life in 7B, 20B, and 40B sizes with windows up to 1 Mb. (A 1B checkpoint with an 8 kb window also exists, but the Evo 2 paper advises that it \"should be avoided owing to overall weaker performance\".)"
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| Evo 2 (7B, 20B, 40B) | 7B peaked at 16.8 GiB at 8 kb and 30.4 GiB at 33 kb on one 48 GB L40S or 80 GB H100, and ran out of memory at 131 kb on one 80 GB H100 ([study repository](https://github.com/rewire-bio/evo2-cloud-operations); [cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/)). The 1B, 20B, and 40B checkpoints require Hopper FP8 support; Arc reports 46 GB on one H100 for 20B (8 kb, 0.7 s/seq) and 95 GB across two H100s for 40B (1.5 s/seq) |"
        }
      ]
    },
    {
      "ref": 15,
      "key": "evo2readme",
      "citation": "Arc Institute. [Evo 2 README: requirements](https://github.com/ArcInstitute/evo2). GitHub.",
      "title": "Evo 2 README: requirements",
      "url": "https://github.com/ArcInstitute/evo2",
      "type": "repository-documentation",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Evo and Evo 2** are large generative models built on the StripedHyena architecture (which I examine in detail below). While the original 7B **Evo** was pretrained only on prokaryotic and phage sequences (131 kb window), **Evo 2** trained on more than 8.8 trillion nucleotides across all domains of life in 7B, 20B, and 40B sizes with windows up to 1 Mb. (A 1B checkpoint with an 8 kb window also exists, but the Evo 2 paper advises that it \"should be avoided owing to overall weaker performance\".)"
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| Evo 2 (7B, 20B, 40B) | 7B peaked at 16.8 GiB at 8 kb and 30.4 GiB at 33 kb on one 48 GB L40S or 80 GB H100, and ran out of memory at 131 kb on one 80 GB H100 ([study repository](https://github.com/rewire-bio/evo2-cloud-operations); [cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/)). The 1B, 20B, and 40B checkpoints require Hopper FP8 support; Arc reports 46 GB on one H100 for 20B (8 kb, 0.7 s/seq) and 95 GB across two H100s for 40B (1.5 s/seq) |"
        }
      ]
    },
    {
      "ref": 16,
      "key": "alphagenome",
      "citation": "Avsec Ž, et al. [Advancing regulatory variant effect prediction with AlphaGenome](https://doi.org/10.1038/s41586-025-10014-0). Nature 649, 1206 to 1218 (2026).",
      "title": "Advancing regulatory variant effect prediction with AlphaGenome",
      "url": "https://doi.org/10.1038/s41586-025-10014-0",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**AlphaGenome** is a supervised sequence-to-function model rather than a self-supervised DNA language model (**Figure 6**): it predicts thousands of experimental tracks—gene expression, splicing, chromatin state, and 3D contacts—across a 1 Mb window, and its distilled model scores a variant in under a second on an H100 GPU. For ClinVar-style variant prioritisation where you want a single summary number rather than thousands of functional tracks, **AlphaGenome Atlas** precomputes the **AlphaGenome Variant Impact (AVI)** score across all possible human single-nucleotide variants. However, its free API and weights are restricted to non-commercial research (with commercial access via Google Cloud), and its terms state that predictions must not be used for clinical decision-making."
        },
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "*Figure 6. How a self-supervised DNA language model and a supervised sequence-to-function model differ in training signal and output. Diagram by the author.*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| AlphaGenome | Under 1 second per variant on an H100; the repository recommends at least an H100 |"
        }
      ]
    },
    {
      "ref": 17,
      "key": "alphagenomeatlas",
      "citation": "Cheng J, Taylor J, Nicolaisen L, et al. [AlphaGenome Atlas: in silico mutagenesis of the entire human genome](https://doi.org/10.64898/2026.09.16.26363192). medRxiv (2026). Introduces the AlphaGenome Variant Impact (AVI) score.",
      "title": "AlphaGenome Atlas: in silico mutagenesis of the entire human genome",
      "url": "https://doi.org/10.64898/2026.09.16.26363192",
      "type": "preprint",
      "version": "v1",
      "alternate_urls": [
        "https://www.medrxiv.org/content/10.64898/2026.09.16.26363192v1"
      ],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**AlphaGenome** is a supervised sequence-to-function model rather than a self-supervised DNA language model (**Figure 6**): it predicts thousands of experimental tracks—gene expression, splicing, chromatin state, and 3D contacts—across a 1 Mb window, and its distilled model scores a variant in under a second on an H100 GPU. For ClinVar-style variant prioritisation where you want a single summary number rather than thousands of functional tracks, **AlphaGenome Atlas** precomputes the **AlphaGenome Variant Impact (AVI)** score across all possible human single-nucleotide variants. However, its free API and weights are restricted to non-commercial research (with commercial access via Google Cloud), and its terms state that predictions must not be used for clinical decision-making."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Start with:* Free precomputed scores from **CADD v1.7**, **REVEL** or **AlphaMissense** (for missense), **GPN-MSA**, and **SpliceAI** (or, for non-commercial research, precomputed **AlphaGenome Variant Impact (AVI)** scores from **AlphaGenome Atlas**)."
        }
      ]
    },
    {
      "ref": 18,
      "key": "alphagenomeapi",
      "citation": "Google DeepMind. [AlphaGenome API README and terms of use](https://github.com/google-deepmind/alphagenome). GitHub.",
      "title": "AlphaGenome API README and terms of use",
      "url": "https://github.com/google-deepmind/alphagenome",
      "type": "repository-documentation",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**AlphaGenome** is a supervised sequence-to-function model rather than a self-supervised DNA language model (**Figure 6**): it predicts thousands of experimental tracks—gene expression, splicing, chromatin state, and 3D contacts—across a 1 Mb window, and its distilled model scores a variant in under a second on an H100 GPU. For ClinVar-style variant prioritisation where you want a single summary number rather than thousands of functional tracks, **AlphaGenome Atlas** precomputes the **AlphaGenome Variant Impact (AVI)** score across all possible human single-nucleotide variants. However, its free API and weights are restricted to non-commercial research (with commercial access via Google Cloud), and its terms state that predictions must not be used for clinical decision-making."
        }
      ]
    },
    {
      "ref": 19,
      "key": "alphagenomeweights",
      "citation": "Google DeepMind. [AlphaGenome model weights README](https://github.com/google-deepmind/alphagenome_research). GitHub.",
      "title": "AlphaGenome model weights README",
      "url": "https://github.com/google-deepmind/alphagenome_research",
      "type": "repository-documentation",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**AlphaGenome** is a supervised sequence-to-function model rather than a self-supervised DNA language model (**Figure 6**): it predicts thousands of experimental tracks—gene expression, splicing, chromatin state, and 3D contacts—across a 1 Mb window, and its distilled model scores a variant in under a second on an H100 GPU. For ClinVar-style variant prioritisation where you want a single summary number rather than thousands of functional tracks, **AlphaGenome Atlas** precomputes the **AlphaGenome Variant Impact (AVI)** score across all possible human single-nucleotide variants. However, its free API and weights are restricted to non-commercial research (with commercial access via Google Cloud), and its terms state that predictions must not be used for clinical decision-making."
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| AlphaGenome | Under 1 second per variant on an H100; the repository recommends at least an H100 |"
        }
      ]
    },
    {
      "ref": 20,
      "key": "genalm",
      "citation": "Fishman V, et al. [GENA-LM: a family of open-source foundational DNA language models for long sequences](https://doi.org/10.1093/nar/gkae1310). Nucleic Acids Research 53, gkae1310 (2025).",
      "title": "GENA-LM: a family of open-source foundational DNA language models for long sequences",
      "url": "https://doi.org/10.1093/nar/gkae1310",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Other models you will see.** **GENA-LM** ([GitHub](https://github.com/AIRI-Institute/GENA_LM)) accepts windows of about 4.5 kb or 36 kb. **Genos** (1.2B and 10B, up to 1 Mb) appears in the independent splicing test below; BGI subsequently released an updated checkpoint, **Genos-10B-v2** ([Hugging Face](https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2), 18 December 2025), trained with additional non-human primate and mammalian genomes and reporting improved RNA splicing prediction. **GenomeOcean** (100M to 4B) contributes H100 embedding memory figures in **Table 5**. All of these models have open weights or API access, but licences vary by checkpoint—check them before touching clinical workflows."
        }
      ]
    },
    {
      "ref": 21,
      "key": "genos",
      "citation": "Lin A, et al. [Genos: a human-centric genomic foundation model](https://doi.org/10.1093/gigascience/giaf132). GigaScience 14, giaf132 (2025).",
      "title": "Genos: a human-centric genomic foundation model",
      "url": "https://doi.org/10.1093/gigascience/giaf132",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Other models you will see.** **GENA-LM** ([GitHub](https://github.com/AIRI-Institute/GENA_LM)) accepts windows of about 4.5 kb or 36 kb. **Genos** (1.2B and 10B, up to 1 Mb) appears in the independent splicing test below; BGI subsequently released an updated checkpoint, **Genos-10B-v2** ([Hugging Face](https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2), 18 December 2025), trained with additional non-human primate and mammalian genomes and reporting improved RNA splicing prediction. **GenomeOcean** (100M to 4B) contributes H100 embedding memory figures in **Table 5**. All of these models have open weights or API access, but licences vary by checkpoint—check them before touching clinical workflows."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "| Genos 10B | 27.6% | 85.0% |"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        }
      ]
    },
    {
      "ref": 22,
      "key": "genosv2",
      "citation": "BGI-HangzhouAI. [Genos-10B-v2 model card](https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2). Hugging Face (released 18 December 2025).",
      "title": "Genos-10B-v2 model card",
      "url": "https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2",
      "type": "model-card",
      "version": "Genos-10B-v2 (2025-12-18)",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Other models you will see.** **GENA-LM** ([GitHub](https://github.com/AIRI-Institute/GENA_LM)) accepts windows of about 4.5 kb or 36 kb. **Genos** (1.2B and 10B, up to 1 Mb) appears in the independent splicing test below; BGI subsequently released an updated checkpoint, **Genos-10B-v2** ([Hugging Face](https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2), 18 December 2025), trained with additional non-human primate and mammalian genomes and reporting improved RNA splicing prediction. **GenomeOcean** (100M to 4B) contributes H100 embedding memory figures in **Table 5**. All of these models have open weights or API access, but licences vary by checkpoint—check them before touching clinical workflows."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        }
      ]
    },
    {
      "ref": 23,
      "key": "genomeocean",
      "citation": "Yu M, et al. [FP8 Inference in Genomic Foundation Models: Theoretical vs. Realized Speedups on GenomeOcean](https://doi.org/10.64898/2026.08.09.743676). bioRxiv (2026).",
      "title": "FP8 Inference in Genomic Foundation Models: Theoretical vs. Realized Speedups on GenomeOcean",
      "url": "https://doi.org/10.64898/2026.08.09.743676",
      "type": "preprint",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Which DNA Model Fits Your Task?",
          "text": "**Other models you will see.** **GENA-LM** ([GitHub](https://github.com/AIRI-Institute/GENA_LM)) accepts windows of about 4.5 kb or 36 kb. **Genos** (1.2B and 10B, up to 1 Mb) appears in the independent splicing test below; BGI subsequently released an updated checkpoint, **Genos-10B-v2** ([Hugging Face](https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2), 18 December 2025), trained with additional non-human primate and mammalian genomes and reporting improved RNA splicing prediction. **GenomeOcean** (100M to 4B) contributes H100 embedding memory figures in **Table 5**. All of these models have open weights or API access, but licences vary by checkpoint—check them before touching clinical workflows."
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| GenomeOcean | Peak 21 GB (100M), 47 GB (500M) and 72 GB (4B) on an H100 when extracting embeddings in 16-bit (BF16) precision, with 5 kb inputs for the two smaller models and 50 kb for 4B; batch size not stated (Table 3) |"
        }
      ]
    },
    {
      "ref": 24,
      "key": "runpod",
      "citation": "Runpod GPU catalogue, secure-cloud on-demand prices, read through the Runpod API on 13 October 2026. My [recorded prices and calculation](/downloads/genomic-foundation-models-in-the-diagnostic-lab/pretraining-costs.json); prices change, and the live list is on [Runpod's pricing page](https://www.runpod.io/pricing).",
      "title": "recorded prices and calculation",
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/pretraining-costs.json",
      "type": "price-record",
      "version": "read 2026-10-13",
      "alternate_urls": [
        "https://www.runpod.io/pricing"
      ],
      "claims": [
        {
          "section": "The Efficiency Revolution: Why Pretraining Scale Doesn't Pick Your Model",
          "text": "*Table 3. Estimated cloud compute cost to pretrain five DNA foundation models from scratch, multiplying author-reported GPU-hours by 13 October 2026 Runpod secure-cloud on-demand prices (A100 80 GB \\$1.79/h, H100 \\$3.99/h; DNABERT-2's RTX 2080 Ti is priced at its faster RTX 3090 successor at \\$0.50/h; small HyenaDNA's A100 40 GB is priced at the 80 GB rate). My calculation, not what the authors paid; storage, networking, failed runs, and staff time are excluded. Sources: Nucleotide Transformer Methods; NTv3 Table 5; HyenaDNA Section 5, Appendix A.1 and Table A.2; DNABERT-2 Appendix A.3. (HyenaDNA Table A.2 lists a higher 215,000 GPU-hours for NT 2.5B; this table uses the Nucleotide Transformer paper's hardware figures.)*"
        }
      ]
    },
    {
      "ref": 25,
      "key": "mutect2",
      "citation": "Benjamin D, et al. [Calling somatic SNVs and indels with Mutect2](https://doi.org/10.1101/861054). bioRxiv (2019).",
      "title": "Calling somatic SNVs and indels with Mutect2",
      "url": "https://doi.org/10.1101/861054",
      "type": "preprint",
      "version": "2019",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 26,
      "key": "deepsomatic",
      "citation": "Park J, et al. [Accurate somatic small variant discovery for multiple sequencing technologies with DeepSomatic](https://doi.org/10.1038/s41587-025-02839-x). Nature Biotechnology (2025).",
      "title": "Accurate somatic small variant discovery for multiple sequencing technologies with DeepSomatic",
      "url": "https://doi.org/10.1038/s41587-025-02839-x",
      "type": "paper",
      "version": "2025",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        }
      ]
    },
    {
      "ref": 27,
      "key": "strelka2",
      "citation": "Kim S, et al. [Strelka2: fast and accurate calling of germline and somatic variants](https://doi.org/10.1038/s41592-018-0051-x). Nature Methods 15, 591 to 594 (2018).",
      "title": "Strelka2: fast and accurate calling of germline and somatic variants",
      "url": "https://doi.org/10.1038/s41592-018-0051-x",
      "type": "paper",
      "version": "2018",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        }
      ]
    },
    {
      "ref": 28,
      "key": "hasan",
      "citation": "Hasan M. [Prospectively frozen external benchmarking reveals benchmark-context dependence in somatic variant calling](https://doi.org/10.21203/rs.3.rs-11210923/v1). Research Square preprint (2026).",
      "title": "Prospectively frozen external benchmarking reveals benchmark-context dependence in somatic variant calling",
      "url": "https://doi.org/10.21203/rs.3.rs-11210923/v1",
      "type": "preprint",
      "version": "v1",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        }
      ]
    },
    {
      "ref": 29,
      "key": "wang2020",
      "citation": "Wang M, et al. [SomaticCombiner: improving the performance of somatic variant calling based on evaluation tests and a consensus approach](https://doi.org/10.1038/s41598-020-69772-8). Scientific Reports 10, 12898 (2020).",
      "title": "SomaticCombiner: improving the performance of somatic variant calling based on evaluation tests and a consensus approach",
      "url": "https://doi.org/10.1038/s41598-020-69772-8",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        }
      ]
    },
    {
      "ref": 30,
      "key": "behera",
      "citation": "Behera S, et al. [Comprehensive genome analysis and variant detection at scale using DRAGEN](https://doi.org/10.1038/s41587-024-02382-1). Nature Biotechnology 43, 1177 to 1191 (2025).",
      "title": "Comprehensive genome analysis and variant detection at scale using DRAGEN",
      "url": "https://doi.org/10.1038/s41587-024-02382-1",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        }
      ]
    },
    {
      "ref": 31,
      "key": "dragensomatic",
      "citation": "Illumina. [DRAGEN v4.4 somatic mode documentation](https://help.connected.illumina.com/dragen/dragen-v4.4/product-guide/dragen-v4.4/dragen-dna-pipeline/small-variant-calling/somatic-mode).",
      "title": "DRAGEN v4.4 somatic mode documentation",
      "url": "https://help.connected.illumina.com/dragen/dragen-v4.4/product-guide/dragen-v4.4/dragen-dna-pipeline/small-variant-calling/somatic-mode",
      "type": "paper",
      "version": "v4.4",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour somatic SNVs and indels (Oncology):** No foundation model was compared against the workflow baseline, **Mutect2**. Among task-specific deep-learning callers, DeepSomatic matched Mutect2 and beat Strelka2 in the only independent tumour-normal benchmark that included it, and NeuSomatic beat MuTect2 on SNV F1 score in 3 of 8 whole-genome pairs tested by Wang and colleagues. Keep Mutect2 (or hardware-accelerated Illumina DRAGEN) and evaluate deep-learning callers against it locally. [Full review](/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/)."
        }
      ]
    },
    {
      "ref": 32,
      "key": "starfusion",
      "citation": "Haas BJ, et al. [Accuracy assessment of fusion transcript detection via read-mapping and de novo fusion transcript assembly-based methods](https://doi.org/10.1186/s13059-019-1842-9). Genome Biology 20, 213 (2019). Introduces STAR-Fusion.",
      "title": "Accuracy assessment of fusion transcript detection via read-mapping and de novo fusion transcript assembly-based methods",
      "url": "https://doi.org/10.1186/s13059-019-1842-9",
      "type": "paper",
      "version": "2019",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour RNA fusions (Oncology):** No foundation-model or deep-learning caller was compared against **STAR-Fusion** or **Arriba** on the clinical workflow endpoint, including in Tamura and colleagues' 12-caller clinical RNA-seq benchmark. [Full review](/blog/what-the-benchmarks-say-about-choosing-an-rna-fusion-caller/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 33,
      "key": "arriba",
      "citation": "Uhrig S, et al. [Accurate and efficient detection of gene fusions from RNA sequencing data](https://doi.org/10.1101/gr.257246.119). Genome Research 31, 448 to 460 (2021). Introduces Arriba.",
      "title": "Accurate and efficient detection of gene fusions from RNA sequencing data",
      "url": "https://doi.org/10.1101/gr.257246.119",
      "type": "paper",
      "version": "2021",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour RNA fusions (Oncology):** No foundation-model or deep-learning caller was compared against **STAR-Fusion** or **Arriba** on the clinical workflow endpoint, including in Tamura and colleagues' 12-caller clinical RNA-seq benchmark. [Full review](/blog/what-the-benchmarks-say-about-choosing-an-rna-fusion-caller/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 34,
      "key": "tamura",
      "citation": "Tamura Z, et al. [Comparison of gene fusion detection algorithms reveals frequently overlooked driver fusions in hematologic malignancies](https://doi.org/10.1038/s41698-026-01397-y). npj Precision Oncology 10, 81 (2026).",
      "title": "Comparison of gene fusion detection algorithms reveals frequently overlooked driver fusions in hematologic malignancies",
      "url": "https://doi.org/10.1038/s41698-026-01397-y",
      "type": "paper",
      "version": "2026",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Tumour RNA fusions (Oncology):** No foundation-model or deep-learning caller was compared against **STAR-Fusion** or **Arriba** on the clinical workflow endpoint, including in Tamura and colleagues' 12-caller clinical RNA-seq benchmark. [Full review](/blog/what-the-benchmarks-say-about-choosing-an-rna-fusion-caller/)."
        }
      ]
    },
    {
      "ref": 35,
      "key": "cristiano",
      "citation": "Cristiano S, et al. [Genome-wide cell-free DNA fragmentation in patients with cancer](https://doi.org/10.1038/s41586-019-1272-6). Nature 570, 385 to 389 (2019). Introduces DELFI.",
      "title": "Genome-wide cell-free DNA fragmentation in patients with cancer",
      "url": "https://doi.org/10.1038/s41586-019-1272-6",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting circulating tumour DNA (ctDNA) from fragmentomics (Oncology):** **Fragmentia-AI** (a fragment model built on DNABERT-2) reported an AUROC of 0.956 versus at most 0.935 for DELFI fragment features in its developers' study, but has not been compared against **ichorCNA**, the standard tumour-fraction baseline. Promising, not yet proven. [Full review](/blog/what-eight-studies-show-about-detecting-cancer-from-cfdna-fragmentomics/)."
        }
      ]
    },
    {
      "ref": 36,
      "key": "xufragmentia",
      "citation": "Xu Y, et al. [Toward generalizable prediction of cancer signal using a cell-free DNA language model](https://doi.org/10.1016/j.xcrm.2026.102866). Cell Reports Medicine 7, 102866 (2026). Introduces Fragmentia-AI.",
      "title": "Toward generalizable prediction of cancer signal using a cell-free DNA language model",
      "url": "https://doi.org/10.1016/j.xcrm.2026.102866",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting circulating tumour DNA (ctDNA) from fragmentomics (Oncology):** **Fragmentia-AI** (a fragment model built on DNABERT-2) reported an AUROC of 0.956 versus at most 0.935 for DELFI fragment features in its developers' study, but has not been compared against **ichorCNA**, the standard tumour-fraction baseline. Promising, not yet proven. [Full review](/blog/what-eight-studies-show-about-detecting-cancer-from-cfdna-fragmentomics/)."
        }
      ]
    },
    {
      "ref": 37,
      "key": "ichorcna",
      "citation": "Adalsteinsson VA, et al. [Scalable whole-exome sequencing of cell-free DNA reveals high concordance with metastatic tumors](https://doi.org/10.1038/s41467-017-00965-y). Nature Communications 8, 1324 (2017). Introduces ichorCNA.",
      "title": "Scalable whole-exome sequencing of cell-free DNA reveals high concordance with metastatic tumors",
      "url": "https://doi.org/10.1038/s41467-017-00965-y",
      "type": "paper",
      "version": "2017",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting circulating tumour DNA (ctDNA) from fragmentomics (Oncology):** **Fragmentia-AI** (a fragment model built on DNABERT-2) reported an AUROC of 0.956 versus at most 0.935 for DELFI fragment features in its developers' study, but has not been compared against **ichorCNA**, the standard tumour-fraction baseline. Promising, not yet proven. [Full review](/blog/what-eight-studies-show-about-detecting-cancer-from-cfdna-fragmentomics/)."
        }
      ]
    },
    {
      "ref": 38,
      "key": "methatlas",
      "citation": "Moss J, et al. [Comprehensive human cell-type methylation atlas reveals origins of circulating cell-free DNA in health and disease](https://doi.org/10.1038/s41467-018-07466-6). Nature Communications 9, 5068 (2018). Introduces MethAtlas.",
      "title": "Comprehensive human cell-type methylation atlas reveals origins of circulating cell-free DNA in health and disease",
      "url": "https://doi.org/10.1038/s41467-018-07466-6",
      "type": "paper",
      "version": "2018",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting ctDNA and tissue of origin from DNA methylation (Oncology):** Neither elastic-net detection nor constrained least-squares tissue-of-origin deconvolution (such as **MethAtlas**) has been surpassed in an unconfounded foundation-model test. Xu and Feng fine-tuned **CpGPT** (a methylation foundation model they did not develop) and reported a mean test AUROC of 0.990 against 0.942 for elastic net and 0.957 for an SVM across three ~30-sample test splits. However, their Methods state that cancer-versus-control labels from all 300 samples were used to select the candidate CpG sites *before* splitting the cohort into training and test sets—a data-leakage limitation the authors acknowledge, which prevents knowing how much of CpGPT's lead would hold if feature selection were confined strictly to each training split. [Full review](/blog/what-seven-studies-show-about-detecting-cancer-from-plasma-dna-methylation/)."
        }
      ]
    },
    {
      "ref": 39,
      "key": "cpgpt",
      "citation": "de Lima Camillo LP, et al. [CpGPT: a foundation model for DNA methylation](https://doi.org/10.1101/2024.10.24.619766). bioRxiv (2024).",
      "title": "CpGPT: a foundation model for DNA methylation",
      "url": "https://doi.org/10.1101/2024.10.24.619766",
      "type": "preprint",
      "version": "2024",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting ctDNA and tissue of origin from DNA methylation (Oncology):** Neither elastic-net detection nor constrained least-squares tissue-of-origin deconvolution (such as **MethAtlas**) has been surpassed in an unconfounded foundation-model test. Xu and Feng fine-tuned **CpGPT** (a methylation foundation model they did not develop) and reported a mean test AUROC of 0.990 against 0.942 for elastic net and 0.957 for an SVM across three ~30-sample test splits. However, their Methods state that cancer-versus-control labels from all 300 samples were used to select the candidate CpG sites *before* splitting the cohort into training and test sets—a data-leakage limitation the authors acknowledge, which prevents knowing how much of CpGPT's lead would hold if feature selection were confined strictly to each training split. [Full review](/blog/what-seven-studies-show-about-detecting-cancer-from-plasma-dna-methylation/)."
        }
      ]
    },
    {
      "ref": 40,
      "key": "xufeng",
      "citation": "Xu Y, Feng Q. [An Evaluation of DMR-Informed Fine-Tuning of Tissue Array-Pretrained CpGPT for Gastrointestinal Cancer Classification Using cfDNA Targeted Methylation](https://www.medrxiv.org/content/10.64898/2026.07.26.26358979v1.full). medRxiv (2026). Version 1, 29 July 2026.",
      "title": "An Evaluation of DMR-Informed Fine-Tuning of Tissue Array-Pretrained CpGPT for Gastrointestinal Cancer Classification Using cfDNA Targeted Methylation",
      "url": "https://www.medrxiv.org/content/10.64898/2026.07.26.26358979v1.full",
      "type": "preprint",
      "version": "v1",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting ctDNA and tissue of origin from DNA methylation (Oncology):** Neither elastic-net detection nor constrained least-squares tissue-of-origin deconvolution (such as **MethAtlas**) has been surpassed in an unconfounded foundation-model test. Xu and Feng fine-tuned **CpGPT** (a methylation foundation model they did not develop) and reported a mean test AUROC of 0.990 against 0.942 for elastic net and 0.957 for an SVM across three ~30-sample test splits. However, their Methods state that cancer-versus-control labels from all 300 samples were used to select the candidate CpG sites *before* splitting the cohort into training and test sets—a data-leakage limitation the authors acknowledge, which prevents knowing how much of CpGPT's lead would hold if feature selection were confined strictly to each training split. [Full review](/blog/what-seven-studies-show-about-detecting-cancer-from-plasma-dna-methylation/)."
        },
        {
          "section": "Fine-Tune (Route 3) When:",
          "text": "*Figure 15. Feature selection, model fitting and threshold choice belong inside the training and validation data. The leaky pattern matches the limitation acknowledged in the CpGPT methylation study. Diagram by the author.*"
        }
      ]
    },
    {
      "ref": 41,
      "key": "exomedepth",
      "citation": "Plagnol V, et al. [A robust model for read count data in exome sequencing experiments and implications for copy number variant calling](https://doi.org/10.1093/bioinformatics/bts526). Bioinformatics 28, 2747 to 2754 (2012). Introduces ExomeDepth.",
      "title": "A robust model for read count data in exome sequencing experiments and implications for copy number variant calling",
      "url": "https://doi.org/10.1093/bioinformatics/bts526",
      "type": "paper",
      "version": "2012",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting and classifying copy-number variants (Oncology and Inherited):** No foundation model was evaluated against established CNV callers (**ExomeDepth**, **CNVnator**) or clinical pathogenicity classifiers (**AnnotSV**, **ClassifyCNV**). Because established exome CNV callers and automated ACMG classifiers themselves show limited concordance and positive predictive value in independent benchmarks, read the [full CNV review](/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/) before selecting a pipeline."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 42,
      "key": "cnvnator",
      "citation": "Abyzov A, et al. [CNVnator: an approach to discover, genotype, and characterize typical and atypical CNVs from family and population genome sequencing](https://doi.org/10.1101/gr.114876.110). Genome Research 21, 974 to 984 (2011).",
      "title": "CNVnator: an approach to discover, genotype, and characterize typical and atypical CNVs from family and population genome sequencing",
      "url": "https://doi.org/10.1101/gr.114876.110",
      "type": "paper",
      "version": "2011",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting and classifying copy-number variants (Oncology and Inherited):** No foundation model was evaluated against established CNV callers (**ExomeDepth**, **CNVnator**) or clinical pathogenicity classifiers (**AnnotSV**, **ClassifyCNV**). Because established exome CNV callers and automated ACMG classifiers themselves show limited concordance and positive predictive value in independent benchmarks, read the [full CNV review](/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/) before selecting a pipeline."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 43,
      "key": "annotsv",
      "citation": "Geoffroy V, et al. [AnnotSV: an integrated tool for structural variations annotation](https://doi.org/10.1093/bioinformatics/bty304). Bioinformatics 34, 3572 to 3574 (2018).",
      "title": "AnnotSV: an integrated tool for structural variations annotation",
      "url": "https://doi.org/10.1093/bioinformatics/bty304",
      "type": "paper",
      "version": "2018",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting and classifying copy-number variants (Oncology and Inherited):** No foundation model was evaluated against established CNV callers (**ExomeDepth**, **CNVnator**) or clinical pathogenicity classifiers (**AnnotSV**, **ClassifyCNV**). Because established exome CNV callers and automated ACMG classifiers themselves show limited concordance and positive predictive value in independent benchmarks, read the [full CNV review](/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/) before selecting a pipeline."
        }
      ]
    },
    {
      "ref": 44,
      "key": "classifycnv",
      "citation": "Gurbich TA, Ilinsky VV. [ClassifyCNV: a tool for clinical annotation of copy-number variants](https://doi.org/10.1038/s41598-020-76425-3). Scientific Reports 10, 20375 (2020).",
      "title": "ClassifyCNV: a tool for clinical annotation of copy-number variants",
      "url": "https://doi.org/10.1038/s41598-020-76425-3",
      "type": "paper",
      "version": "2020",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting and classifying copy-number variants (Oncology and Inherited):** No foundation model was evaluated against established CNV callers (**ExomeDepth**, **CNVnator**) or clinical pathogenicity classifiers (**AnnotSV**, **ClassifyCNV**). Because established exome CNV callers and automated ACMG classifiers themselves show limited concordance and positive predictive value in independent benchmarks, read the [full CNV review](/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/) before selecting a pipeline."
        }
      ]
    },
    {
      "ref": 45,
      "key": "hwang",
      "citation": "Hwang J, et al. [Improving CNV Detection Performance Except for Software-Specific Problematic Regions](https://doi.org/10.3390/genes17010105). Genes 17, 105 (2026).",
      "title": "Improving CNV Detection Performance Except for Software-Specific Problematic Regions",
      "url": "https://doi.org/10.3390/genes17010105",
      "type": "paper",
      "version": "2026",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting and classifying copy-number variants (Oncology and Inherited):** No foundation model was evaluated against established CNV callers (**ExomeDepth**, **CNVnator**) or clinical pathogenicity classifiers (**AnnotSV**, **ClassifyCNV**). Because established exome CNV callers and automated ACMG classifiers themselves show limited concordance and positive predictive value in independent benchmarks, read the [full CNV review](/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/) before selecting a pipeline."
        }
      ]
    },
    {
      "ref": 46,
      "key": "zhangcnv",
      "citation": "Zhang Y, Zhang X, Chen D. [Can automated copy number variation pathogenicity prediction tools replace manual review under the ACMG framework? A clinical validation](https://doi.org/10.1016/j.gmg.2026.100096). Global Medical Genetics 100096 (2026).",
      "title": "Can automated copy number variation pathogenicity prediction tools replace manual review under the ACMG framework? A clinical validation",
      "url": "https://doi.org/10.1016/j.gmg.2026.100096",
      "type": "paper",
      "version": "2026",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Detecting and classifying copy-number variants (Oncology and Inherited):** No foundation model was evaluated against established CNV callers (**ExomeDepth**, **CNVnator**) or clinical pathogenicity classifiers (**AnnotSV**, **ClassifyCNV**). Because established exome CNV callers and automated ACMG classifiers themselves show limited concordance and positive predictive value in independent benchmarks, read the [full CNV review](/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/) before selecting a pipeline."
        }
      ]
    },
    {
      "ref": 47,
      "key": "kraken2",
      "citation": "Wood DE, Lu J, Langmead B. [Improved metagenomic analysis with Kraken 2](https://doi.org/10.1186/s13059-019-1891-0). Genome Biology 20, 257 (2019).",
      "title": "Improved metagenomic analysis with Kraken 2",
      "url": "https://doi.org/10.1186/s13059-019-1891-0",
      "type": "paper",
      "version": "2019",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **DNA pathogen identification (Infectious disease):** No foundation model was compared against clinical metagenomic workflow baselines (**Kraken2**, or **minimap2** with **MEGAN-LR**). Deep-learning taxonomic classifiers such as BERTax and MetageNN have been tested only by their developers on simulated, isolate, or mock-community reads; at the genus level, MetageNN classified more reads from species absent from its reference database than MEGAN-LR, but with lower precision. Start with minimap2 or MEGAN-LR with protein alignment, and audit your reference database carefully. [Full review](/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/)."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **RNA pathogen detection (Infectious disease):** No machine-learning comparison—foundation model or otherwise—was found against **Kraken2** or **Centrifuge**. Clinical metagenomic labs more often align non-human RNA reads to the NCBI nucleotide database (as **SURPI+** and **CZ ID** do). [Full review](/blog/rna-pathogen-detection-by-sequencing/)."
        }
      ]
    },
    {
      "ref": 48,
      "key": "minimap2",
      "citation": "Li H. [Minimap2: pairwise alignment for nucleotide sequences](https://doi.org/10.1093/bioinformatics/bty191). Bioinformatics 34, 3094 to 3100 (2018).",
      "title": "Minimap2: pairwise alignment for nucleotide sequences",
      "url": "https://doi.org/10.1093/bioinformatics/bty191",
      "type": "paper",
      "version": "2018",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **DNA pathogen identification (Infectious disease):** No foundation model was compared against clinical metagenomic workflow baselines (**Kraken2**, or **minimap2** with **MEGAN-LR**). Deep-learning taxonomic classifiers such as BERTax and MetageNN have been tested only by their developers on simulated, isolate, or mock-community reads; at the genus level, MetageNN classified more reads from species absent from its reference database than MEGAN-LR, but with lower precision. Start with minimap2 or MEGAN-LR with protein alignment, and audit your reference database carefully. [Full review](/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 49,
      "key": "meganlr",
      "citation": "Huson DH, et al. [MEGAN-LR: new algorithms allow accurate binning and easy interactive exploration of metagenomic long reads and contigs](https://doi.org/10.1186/s13062-018-0208-7). Biology Direct 13, 6 (2018).",
      "title": "MEGAN-LR: new algorithms allow accurate binning and easy interactive exploration of metagenomic long reads and contigs",
      "url": "https://doi.org/10.1186/s13062-018-0208-7",
      "type": "paper",
      "version": "2018",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **DNA pathogen identification (Infectious disease):** No foundation model was compared against clinical metagenomic workflow baselines (**Kraken2**, or **minimap2** with **MEGAN-LR**). Deep-learning taxonomic classifiers such as BERTax and MetageNN have been tested only by their developers on simulated, isolate, or mock-community reads; at the genus level, MetageNN classified more reads from species absent from its reference database than MEGAN-LR, but with lower precision. Start with minimap2 or MEGAN-LR with protein alignment, and audit your reference database carefully. [Full review](/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 50,
      "key": "portik",
      "citation": "Portik DM, Brown CT, Pierce-Ward NT. [Evaluation of taxonomic classification and profiling methods for long-read shotgun metagenomic sequencing datasets](https://doi.org/10.1186/s12859-022-05103-0). BMC Bioinformatics 23, 541 (2022).",
      "title": "Evaluation of taxonomic classification and profiling methods for long-read shotgun metagenomic sequencing datasets",
      "url": "https://doi.org/10.1186/s12859-022-05103-0",
      "type": "paper",
      "version": "2022",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **DNA pathogen identification (Infectious disease):** No foundation model was compared against clinical metagenomic workflow baselines (**Kraken2**, or **minimap2** with **MEGAN-LR**). Deep-learning taxonomic classifiers such as BERTax and MetageNN have been tested only by their developers on simulated, isolate, or mock-community reads; at the genus level, MetageNN classified more reads from species absent from its reference database than MEGAN-LR, but with lower precision. Start with minimap2 or MEGAN-LR with protein alignment, and audit your reference database carefully. [Full review](/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/)."
        }
      ]
    },
    {
      "ref": 51,
      "key": "bertax",
      "citation": "Mock F, et al. [Taxonomic classification of DNA sequences beyond sequence similarity using deep neural networks](https://doi.org/10.1073/pnas.2122636119). PNAS 119, e2122636119 (2022). Introduces BERTax.",
      "title": "Taxonomic classification of DNA sequences beyond sequence similarity using deep neural networks",
      "url": "https://doi.org/10.1073/pnas.2122636119",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **DNA pathogen identification (Infectious disease):** No foundation model was compared against clinical metagenomic workflow baselines (**Kraken2**, or **minimap2** with **MEGAN-LR**). Deep-learning taxonomic classifiers such as BERTax and MetageNN have been tested only by their developers on simulated, isolate, or mock-community reads; at the genus level, MetageNN classified more reads from species absent from its reference database than MEGAN-LR, but with lower precision. Start with minimap2 or MEGAN-LR with protein alignment, and audit your reference database carefully. [Full review](/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/)."
        }
      ]
    },
    {
      "ref": 52,
      "key": "metagenn",
      "citation": "Peres da Silva R, et al. [MetageNN: a memory-efficient neural network taxonomic classifier robust to sequencing errors and missing genomes](https://doi.org/10.1186/s12859-024-05760-3). BMC Bioinformatics 25, 153 (2024).",
      "title": "MetageNN: a memory-efficient neural network taxonomic classifier robust to sequencing errors and missing genomes",
      "url": "https://doi.org/10.1186/s12859-024-05760-3",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **DNA pathogen identification (Infectious disease):** No foundation model was compared against clinical metagenomic workflow baselines (**Kraken2**, or **minimap2** with **MEGAN-LR**). Deep-learning taxonomic classifiers such as BERTax and MetageNN have been tested only by their developers on simulated, isolate, or mock-community reads; at the genus level, MetageNN classified more reads from species absent from its reference database than MEGAN-LR, but with lower precision. Start with minimap2 or MEGAN-LR with protein alignment, and audit your reference database carefully. [Full review](/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/)."
        }
      ]
    },
    {
      "ref": 53,
      "key": "centrifuge",
      "citation": "Kim D, et al. [Centrifuge: rapid and sensitive classification of metagenomic sequences](https://doi.org/10.1101/gr.210641.116). Genome Research 26, 1721 to 1729 (2016).",
      "title": "Centrifuge: rapid and sensitive classification of metagenomic sequences",
      "url": "https://doi.org/10.1101/gr.210641.116",
      "type": "paper",
      "version": "2016",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **RNA pathogen detection (Infectious disease):** No machine-learning comparison—foundation model or otherwise—was found against **Kraken2** or **Centrifuge**. Clinical metagenomic labs more often align non-human RNA reads to the NCBI nucleotide database (as **SURPI+** and **CZ ID** do). [Full review](/blog/rna-pathogen-detection-by-sequencing/)."
        }
      ]
    },
    {
      "ref": 54,
      "key": "surpi",
      "citation": "Tan JK, et al. [Laboratory validation of a clinical metagenomic next-generation sequencing assay for respiratory virus detection and discovery](https://doi.org/10.1038/s41467-024-51470-y). Nature Communications 15, 9016 (2024). Evaluates SURPI+ on respiratory RNA specimens.",
      "title": "Laboratory validation of a clinical metagenomic next-generation sequencing assay for respiratory virus detection and discovery",
      "url": "https://doi.org/10.1038/s41467-024-51470-y",
      "type": "paper",
      "version": "2024",
      "alternate_urls": [
        "https://pmc.ncbi.nlm.nih.gov/articles/PMC11558004/"
      ],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **RNA pathogen detection (Infectious disease):** No machine-learning comparison—foundation model or otherwise—was found against **Kraken2** or **Centrifuge**. Clinical metagenomic labs more often align non-human RNA reads to the NCBI nucleotide database (as **SURPI+** and **CZ ID** do). [Full review](/blog/rna-pathogen-detection-by-sequencing/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 55,
      "key": "czid",
      "citation": "Kalantar KL, et al. [IDseq—An open source cloud-based pipeline and analysis service for metagenomic pathogen detection and monitoring](https://doi.org/10.1093/gigascience/giaa111). GigaScience 9, giaa111 (2020). Introduces IDseq (now CZ ID).",
      "title": "IDseq—An open source cloud-based pipeline and analysis service for metagenomic pathogen detection and monitoring",
      "url": "https://doi.org/10.1093/gigascience/giaa111",
      "type": "paper",
      "version": "2020",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **RNA pathogen detection (Infectious disease):** No machine-learning comparison—foundation model or otherwise—was found against **Kraken2** or **Centrifuge**. Clinical metagenomic labs more often align non-human RNA reads to the NCBI nucleotide database (as **SURPI+** and **CZ ID** do). [Full review](/blog/rna-pathogen-detection-by-sequencing/)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* Your validated diagnostic workflow (**Mutect2**, **STAR-Fusion** / **Arriba**, **ExomeDepth** / **CNVnator**, **minimap2** + **MEGAN-LR**, or **SURPI+** / **CZ ID**). None of my eight workflow reviews found a foundation model proven to replace them."
        }
      ]
    },
    {
      "ref": 56,
      "key": "marangoni",
      "citation": "Marangoni S, et al. [GeNePi: a graphics processing unit enhanced next-generation bioinformatics pipeline for whole-genome sequencing analysis](https://doi.org/10.1093/bib/bbag001). Briefings in Bioinformatics 27, bbag001 (2026).",
      "title": "GeNePi: a graphics processing unit enhanced next-generation bioinformatics pipeline for whole-genome sequencing analysis",
      "url": "https://doi.org/10.1093/bib/bbag001",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Infrastructure for large-scale genomics analysis and diagnostic testing:** When the bottleneck is primary variant calling across whole genomes rather than post-calling model scoring, published benchmarks show that GPU-accelerated pipelines (**NVIDIA Parabricks**) and FPGA/hardware servers (**Illumina DRAGEN**) cut whole-genome germline calling from 17–39 hours on CPUs down to between about 25 minutes and 2.5 hours, with close agreement to standard CPU calls in the study that compared them directly. [Full infrastructure and execution review](/blog/what-published-benchmarks-show-about-accelerating-variant-calling-and-genomic-models/)."
        }
      ]
    },
    {
      "ref": 57,
      "key": "oconnell",
      "citation": "O'Connell K, et al. [Accelerating genomic workflows using NVIDIA Parabricks](https://doi.org/10.1186/s12859-023-05292-2). BMC Bioinformatics 24, 221 (2023).",
      "title": "Accelerating genomic workflows using NVIDIA Parabricks",
      "url": "https://doi.org/10.1186/s12859-023-05292-2",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Infrastructure for large-scale genomics analysis and diagnostic testing:** When the bottleneck is primary variant calling across whole genomes rather than post-calling model scoring, published benchmarks show that GPU-accelerated pipelines (**NVIDIA Parabricks**) and FPGA/hardware servers (**Illumina DRAGEN**) cut whole-genome germline calling from 17–39 hours on CPUs down to between about 25 minutes and 2.5 hours, with close agreement to standard CPU calls in the study that compared them directly. [Full infrastructure and execution review](/blog/what-published-benchmarks-show-about-accelerating-variant-calling-and-genomic-models/)."
        }
      ]
    },
    {
      "ref": 58,
      "key": "samarakoon",
      "citation": "Samarakoon P, et al. [Benchmarking accelerated next-generation sequencing analysis pipelines](https://doi.org/10.1093/bioadv/vbaf085). Bioinformatics Advances 5, vbaf085 (2025).",
      "title": "Benchmarking accelerated next-generation sequencing analysis pipelines",
      "url": "https://doi.org/10.1093/bioadv/vbaf085",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Infrastructure for large-scale genomics analysis and diagnostic testing:** When the bottleneck is primary variant calling across whole genomes rather than post-calling model scoring, published benchmarks show that GPU-accelerated pipelines (**NVIDIA Parabricks**) and FPGA/hardware servers (**Illumina DRAGEN**) cut whole-genome germline calling from 17–39 hours on CPUs down to between about 25 minutes and 2.5 hours, with close agreement to standard CPU calls in the study that compared them directly. [Full infrastructure and execution review](/blog/what-published-benchmarks-show-about-accelerating-variant-calling-and-genomic-models/)."
        }
      ]
    },
    {
      "ref": 59,
      "key": "spliceai",
      "citation": "Jaganathan K, et al. [Predicting Splicing from Primary Sequence with Deep Learning](https://doi.org/10.1016/j.cell.2018.12.015). Cell 176, 535 to 548 (2019).",
      "title": "Predicting Splicing from Primary Sequence with Deep Learning",
      "url": "https://doi.org/10.1016/j.cell.2018.12.015",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "| SpliceAI | 88.2% | 99.1% |"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 13. The splice-variant workflow supported by current evidence: predict to prioritise, then confirm in patient RNA. Figures from **Table 4**. Diagram by the author.*"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* **SpliceAI** or **Pangolin**, followed by confirmation in patient RNA (**FRASER**). When *OPA1*-tuned high-specificity thresholds were transferred across 65 genes in the only independent RNA-confirmed splice test, SpliceAI achieved the highest sensitivity (88.2% at 99.1% specificity) ahead of four newer genomic models (**Table 4**)."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Start with:* Free precomputed scores from **CADD v1.7**, **REVEL** or **AlphaMissense** (for missense), **GPN-MSA**, and **SpliceAI** (or, for non-commercial research, precomputed **AlphaGenome Variant Impact (AVI)** scores from **AlphaGenome Atlas**)."
        }
      ]
    },
    {
      "ref": 60,
      "key": "borzoi",
      "citation": "Linder J, et al. [Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation](https://doi.org/10.1038/s41588-024-02053-6). Nature Genetics 57, 949 to 961 (2025). Introduces Borzoi.",
      "title": "Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation",
      "url": "https://doi.org/10.1038/s41588-024-02053-6",
      "type": "paper",
      "version": "2025",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "| Borzoi | 73.7% | 96.3% |"
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        }
      ]
    },
    {
      "ref": 61,
      "key": "pangolin",
      "citation": "Zeng T, Li YI. [Predicting RNA splicing from DNA sequence using Pangolin](https://doi.org/10.1186/s13059-022-02664-4). Genome Biology 23, 103 (2022).",
      "title": "Predicting RNA splicing from DNA sequence using Pangolin",
      "url": "https://doi.org/10.1186/s13059-022-02664-4",
      "type": "paper",
      "version": "2022",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 13. The splice-variant workflow supported by current evidence: predict to prioritise, then confirm in patient RNA. Figures from **Table 4**. Diagram by the author.*"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* **SpliceAI** or **Pangolin**, followed by confirmation in patient RNA (**FRASER**). When *OPA1*-tuned high-specificity thresholds were transferred across 65 genes in the only independent RNA-confirmed splice test, SpliceAI achieved the highest sensitivity (88.2% at 99.1% specificity) ahead of four newer genomic models (**Table 4**)."
        }
      ]
    },
    {
      "ref": 62,
      "key": "spliceai2",
      "citation": "Jaganathan K, et al. [A unified framework for quantitative splicing and transcript prediction](https://doi.org/10.64898/2026.10.08.757673). bioRxiv (2026). Introduces SpliceAI2. Release date from Illumina's [press release, 8 October 2026](https://www.illumina.com/company/news-center/press-releases/2026/29236.html).",
      "title": "A unified framework for quantitative splicing and transcript prediction",
      "url": "https://doi.org/10.64898/2026.10.08.757673",
      "type": "preprint",
      "version": null,
      "alternate_urls": [
        "https://www.illumina.com/company/news-center/press-releases/2026/29236.html"
      ],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        }
      ]
    },
    {
      "ref": 63,
      "key": "fraser",
      "citation": "Mertes C, et al. [Detection of aberrant splicing events in RNA-seq data using FRASER](https://doi.org/10.1038/s41467-020-20573-7). Nature Communications 12, 529 (2021).",
      "title": "Detection of aberrant splicing events in RNA-seq data using FRASER",
      "url": "https://doi.org/10.1038/s41467-020-20573-7",
      "type": "paper",
      "version": "2021",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "As **Table 4** and **Figure 12** show, transferring *OPA1*-tuned high-specificity thresholds across 65 genes exposes a trade-off between sensitivity and cross-gene generalisability: the 2019 task-specific tool **SpliceAI** retained both 88.2% sensitivity and 99.1% specificity, whereas **Evo 2 40B** kept 100% specificity at 71.1% sensitivity and post-trained **NTv3 650M** reached 85.5% sensitivity at 86.9% specificity (the initial October 2025 **Genos 10B** checkpoint scored 27.6% sensitivity, before BGI's December 2025 **Genos-10B-v2** update). Alavi also applied those same *OPA1*-tuned thresholds to 306 pathogenic *OPA1* substitutions across every mutation class (missense, nonsense, splice-site, and untranslated region): pretrained-only DNA language models flagged most pathogenic variants regardless of mechanism (Evo 2 40B: 276/306, or 90.2%; pretrained NTv3 650M: 212/306, or 69.3%), acting as general sequence-constraint sensors, whereas post-training NTv3 650M on functional tracks narrowed positive calls to 77/306 (25.2%), in line with the ~25% splicing fraction and specialist models (SpliceAI: 75/306, or 24.5%; Borzoi: 86/306, or 28.1%). Why is **AlphaGenome** omitted from **Table 4** and **Figure 12**? Alavi's AlphaGenome scoring scripts subtracted 1 from each 1-based variant coordinate before calling AlphaGenome's 1-based `Variant` API, shifting every tested position one base to the left ([issue #2](https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2)), so the 56.6% sensitivity and 80.4% specificity reported in the preprint did not evaluate the intended variants. On valid coordinates, AlphaGenome's composite splicing score reached a ClinVar splice-region area under the precision-recall curve (auPRC) of 0.57 versus 0.55 for Pangolin in its own paper. Meanwhile, Illumina released **SpliceAI2** on 8 October 2026: in its developers' preprint (11 October), SpliceAI2 beat SpliceAI, Pangolin, and AlphaGenome on a GTEx cryptic splice variant benchmark (auPRC of 0.77 vs. 0.66 for SpliceAI, 0.56 for Pangolin, and 0.54 for AlphaGenome's composite score or 0.63 with AlphaGenome's raw splice-site score) and on cryptic splice gains in blood RNA-seq from 5,435 Genomics England rare-disease participants (auPRC 0.59 vs. 0.48 for SpliceAI, 0.40 for Pangolin, and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice-site score). My [SpliceAI2 review](/blog/spliceai2/) and [database extraction](https://github.com/rewire-bio/rewire-benchmark-data/issues/127) document those developer-reported figures; for clinical work today, prioritise splice candidates with SpliceAI or Pangolin and confirm in patient RNA with FRASER (**Figure 13**; [splicing review](/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/))."
        },
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 13. The splice-variant workflow supported by current evidence: predict to prioritise, then confirm in patient RNA. Figures from **Table 4**. Diagram by the author.*"
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Use today:* **SpliceAI** or **Pangolin**, followed by confirmation in patient RNA (**FRASER**). When *OPA1*-tuned high-specificity thresholds were transferred across 65 genes in the only independent RNA-confirmed splice test, SpliceAI achieved the highest sensitivity (88.2% at 99.1% specificity) ahead of four newer genomic models (**Table 4**)."
        }
      ]
    },
    {
      "ref": 64,
      "key": "maxentscan",
      "citation": "Yeo G, Burge CB. [Maximum entropy modeling of short sequence motifs with applications to RNA splicing signals](https://doi.org/10.1089/1066527041410418). Journal of Computational Biology 11, 377 to 394 (2004). Introduces MaxEntScan.",
      "title": "Maximum entropy modeling of short sequence motifs with applications to RNA splicing signals",
      "url": "https://doi.org/10.1089/1066527041410418",
      "type": "paper",
      "version": "2004",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "*Figure 13. The splice-variant workflow supported by current evidence: predict to prioritise, then confirm in patient RNA. Figures from **Table 4**. Diagram by the author.*"
        }
      ]
    },
    {
      "ref": 65,
      "key": "pearce",
      "citation": "Pearce MT, et al. [Interpretable variant effect prediction from genomic foundation model representations](https://www.biorxiv.org/content/10.64898/2026.04.10.717844v4). bioRxiv (2026). Introduces EVEE. Version 4, 11 July 2026.",
      "title": "Interpretable variant effect prediction from genomic foundation model representations",
      "url": "https://www.biorxiv.org/content/10.64898/2026.04.10.717844v4",
      "type": "preprint",
      "version": "v4",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        }
      ]
    },
    {
      "ref": 66,
      "key": "alphamissense",
      "citation": "Cheng J, et al. [Accurate proteome-wide missense variant effect prediction with AlphaMissense](https://doi.org/10.1126/science.adg7492). Science 381, eadg7492 (2023).",
      "title": "Accurate proteome-wide missense variant effect prediction with AlphaMissense",
      "url": "https://doi.org/10.1126/science.adg7492",
      "type": "paper",
      "version": "2023",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Start with:* Free precomputed scores from **CADD v1.7**, **REVEL** or **AlphaMissense** (for missense), **GPN-MSA**, and **SpliceAI** (or, for non-commercial research, precomputed **AlphaGenome Variant Impact (AVI)** scores from **AlphaGenome Atlas**)."
        }
      ]
    },
    {
      "ref": 67,
      "key": "esm1b",
      "citation": "Brandes N, et al. [Genome-wide prediction of disease variant effects with a deep protein language model](https://doi.org/10.1038/s41588-023-01465-0). Nature Genetics 55, 1512 to 1522 (2023). Evaluates ESM-1b on human coding variants.",
      "title": "Genome-wide prediction of disease variant effects with a deep protein language model",
      "url": "https://doi.org/10.1038/s41588-023-01465-0",
      "type": "paper",
      "version": "2023",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        }
      ]
    },
    {
      "ref": 68,
      "key": "gpnmsa",
      "citation": "Benegas G, et al. [A DNA language model based on multispecies alignment predicts the effects of genome-wide variants](https://doi.org/10.1038/s41587-024-02511-w). Nature Biotechnology 43, 1960 to 1965 (2025). Introduces GPN-MSA.",
      "title": "A DNA language model based on multispecies alignment predicts the effects of genome-wide variants",
      "url": "https://doi.org/10.1038/s41587-024-02511-w",
      "type": "paper",
      "version": "2025",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Start with:* Free precomputed scores from **CADD v1.7**, **REVEL** or **AlphaMissense** (for missense), **GPN-MSA**, and **SpliceAI** (or, for non-commercial research, precomputed **AlphaGenome Variant Impact (AVI)** scores from **AlphaGenome Atlas**)."
        }
      ]
    },
    {
      "ref": 69,
      "key": "cadd",
      "citation": "Schubach M, et al. [CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions](https://doi.org/10.1093/nar/gkad989). Nucleic Acids Research 52, D1143 to D1154 (2024).",
      "title": "CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions",
      "url": "https://doi.org/10.1093/nar/gkad989",
      "type": "paper",
      "version": "v1.7",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Start with:* Free precomputed scores from **CADD v1.7**, **REVEL** or **AlphaMissense** (for missense), **GPN-MSA**, and **SpliceAI** (or, for non-commercial research, precomputed **AlphaGenome Variant Impact (AVI)** scores from **AlphaGenome Atlas**)."
        }
      ]
    },
    {
      "ref": 70,
      "key": "benegas",
      "citation": "Benegas G, et al. [Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics](https://doi.org/10.1101/2025.02.11.637758). bioRxiv (2025). Introduces TraitGym.",
      "title": "Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics",
      "url": "https://doi.org/10.1101/2025.02.11.637758",
      "type": "preprint",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        }
      ]
    },
    {
      "ref": 71,
      "key": "findlay",
      "citation": "Findlay GM, et al. [Accurate classification of BRCA1 variants with saturation genome editing](https://doi.org/10.1038/s41586-018-0461-z). Nature 562, 217 to 222 (2018).",
      "title": "Accurate classification of BRCA1 variants with saturation genome editing",
      "url": "https://doi.org/10.1038/s41586-018-0461-z",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        }
      ]
    },
    {
      "ref": 72,
      "key": "revel",
      "citation": "Ioannidis NM, et al. [REVEL: An Ensemble Method for Predicting the Pathogenicity of Rare Missense Variants](https://doi.org/10.1016/j.ajhg.2016.08.016). American Journal of Human Genetics 99, 877 to 885 (2016). The study used REVEL v1.3 scores.",
      "title": "REVEL: An Ensemble Method for Predicting the Pathogenicity of Rare Missense Variants",
      "url": "https://doi.org/10.1016/j.ajhg.2016.08.016",
      "type": "paper",
      "version": "v1.3",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
          "text": "- **Post-calling variant effect scoring (Oncology and Inherited):** Once a variant is already called, scoring its functional impact is where foundation models look most competitive. In the **EVEE** preprint (version 4, July 2026), Pearce and colleagues trained a covariance-based probe on frozen Evo 2 7B embeddings and reached an AUROC of 0.991 on ClinVar indels versus 0.980 for CADD. For sequence-to-function models, the **AlphaGenome Atlas** preprint similarly collapses AlphaGenome's multi-track predictions into a precomputed **AlphaGenome Variant Impact (AVI)** score for every possible human SNV (under non-commercial research terms). Still, several EVEE authors co-authored Evo 2, its ClinVar embedding extraction took roughly 20,000 H100 GPU-hours, and Evo 2 does not win on every variant class: in its own paper it trails protein and alignment models (**AlphaMissense**, **ESM-1b**, and **GPN-MSA**) on coding SNVs, and in the TraitGym regulatory benchmark, alignment-based models (CADD and GPN-MSA) compare favourably across Mendelian and complex disease traits. In my companion evaluation on 3,893 experimentally measured *BRCA1* single-nucleotide variants from Findlay and colleagues, zero-shot Evo 2 7B likelihood scoring beat the 2015 CADD v1.3 baseline bundled with the original dataset on noncoding variants (+0.077 AUROC), but that lead disappeared once compared against current **CADD v1.7** (0.953 vs. 0.932; paired difference +0.021, 95% confidence interval -0.003 to +0.045) and **SpliceAI** (0.953 vs. 0.963; paired difference -0.011, 95% interval -0.025 to +0.003), while **CADD v1.7** (0.892 vs. 0.840) and **REVEL** (0.867 vs. 0.785) clearly led on coding and missense variants ([results and methods](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md); [full benchmark and cloud study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        },
        {
          "section": "Practical Decision Matrix",
          "text": "- *Start with:* Free precomputed scores from **CADD v1.7**, **REVEL** or **AlphaMissense** (for missense), **GPN-MSA**, and **SpliceAI** (or, for non-commercial research, precomputed **AlphaGenome Variant Impact (AVI)** scores from **AlphaGenome Atlas**)."
        }
      ]
    },
    {
      "ref": 73,
      "key": "hyenacard",
      "citation": "LongSafari. [HyenaDNA model card](https://huggingface.co/LongSafari/hyenadna-large-1m-seqlen-hf). Hugging Face.",
      "title": "HyenaDNA model card",
      "url": "https://huggingface.co/LongSafari/hyenadna-large-1m-seqlen-hf",
      "type": "model-card",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| HyenaDNA 32k | Model card suggests a T4 (16 GB) to fine-tune and run |"
        },
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "| HyenaDNA 1M | Model card suggests an A100 80 GB to fine-tune and an A100 40 GB to run |"
        }
      ]
    },
    {
      "ref": 74,
      "key": "hegde",
      "citation": "Hegde M, et al. [Reaping the Fruits of LLM Pruning: Towards Small Language Models for Efficient Non-Coding Variant Effect Prediction](https://doi.org/10.3390/genes16111358). Genes 16, 1358 (2025).",
      "title": "Reaping the Fruits of LLM Pruning: Towards Small Language Models for Efficient Non-Coding Variant Effect Prediction",
      "url": "https://doi.org/10.3390/genes16111358",
      "type": "paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
          "text": "Fine-tuning a compact or mid-sized model is far faster than pretraining: the Nucleotide Transformer team fine-tuned their 500M model in 20 minutes on average and their 2.5B model in 50 minutes on one node of eight A100s, and Hegde and colleagues independently fine-tuned NT 500M on roughly 89,000 expression quantitative trait locus (eQTL) variants in about 7.5 hours per five-epoch run on two NVIDIA TITAN Xp GPUs. When running large models such as Evo 2 7B for zero-shot likelihood scoring across cloud GPUs, note that low-precision `bfloat16` and `FP8` reductions are not associative across different GPU models: while two H100s on different clouds produced bit-identical *BRCA1* scores, scoring the exact same container on an L40S shifted per-variant scores by a median of $1.2 \\times 10^{-4}$ and flipped binary calls on 1.8% to 4.9% of variants across fixed thresholds—a drift eliminated by running in full 32-bit `FP32` precision or locking the pipeline to a single validated GPU product ([classification flips](https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CLASSIFICATION-FLIPS.md); [full cloud operations and reproducibility study](/blog/running-evo-2-on-google-cloud-aws-and-runpod/))."
        }
      ]
    },
    {
      "ref": 75,
      "key": "ia3",
      "citation": "Liu H, et al. [Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning](https://arxiv.org/abs/2205.05638). NeurIPS 2022. Introduces (IA)3.",
      "title": "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning",
      "url": "https://arxiv.org/abs/2205.05638",
      "type": "conference-paper",
      "version": null,
      "alternate_urls": [],
      "claims": [
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "- **(IA)3** (used by Nucleotide Transformer) scales inner activations with learned vectors, adding trainable parameters equal to only about **0.1% of the model's parameters** (Results and Methods, Fine-tuning)."
        },
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "*Figure 16. Which weights are trained in full fine-tuning, LoRA and (IA)3. Matrix sizes are schematic. Diagram by the author.*"
        }
      ]
    },
    {
      "ref": 76,
      "key": "lora",
      "citation": "Hu EJ, et al. [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685). ICLR 2022.",
      "title": "LoRA: Low-Rank Adaptation of Large Language Models",
      "url": "https://arxiv.org/abs/2106.09685",
      "type": "conference-paper",
      "version": "ICLR 2022",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "- **LoRA** (built into DNABERT-2's `finetune/train.py` script via `use_lora`) injects low-rank update matrices into attention layers."
        },
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "*Figure 16. Which weights are trained in full fine-tuning, LoRA and (IA)3. Matrix sizes are schematic. Diagram by the author.*"
        }
      ]
    },
    {
      "ref": 77,
      "key": "dnabert2train",
      "citation": "MAGICS Lab. [DNABERT-2 fine-tuning script, `finetune/train.py`](https://github.com/MAGICS-LAB/DNABERT_2/blob/b3337f3958f388f7ff27ba26835c03cb83b41d5f/finetune/train.py). GitHub, commit b3337f3 (21 August 2024).",
      "title": "DNABERT-2 fine-tuning script, `finetune/train.py`",
      "url": "https://github.com/MAGICS-LAB/DNABERT_2/blob/b3337f3958f388f7ff27ba26835c03cb83b41d5f/finetune/train.py",
      "type": "source-code",
      "version": "commit b3337f3 (21 August 2024)",
      "alternate_urls": [],
      "claims": [
        {
          "section": "Parameter-Efficient Fine-Tuning: (IA)3 and LoRA",
          "text": "- **LoRA** (built into DNABERT-2's `finetune/train.py` script via `use_lora`) injects low-rank update matrices into attention layers."
        }
      ]
    }
  ],
  "other_links": [
    {
      "url": "https://github.com/MAGICS-LAB/DNABERT_2",
      "type": "model-or-code-download",
      "link_text": [
        "GitHub"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://huggingface.co/zhihan1996/DNABERT-2-117M",
      "type": "model-or-code-download",
      "link_text": [
        "Hugging Face"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/instadeepai/nucleotide-transformer",
      "type": "model-or-code-download",
      "link_text": [
        "GitHub"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://huggingface.co/InstaDeepAI/NTv3_650M_pre",
      "type": "model-or-code-download",
      "link_text": [
        "Hugging Face"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/HazyResearch/hyena-dna",
      "type": "model-or-code-download",
      "link_text": [
        "GitHub"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://huggingface.co/LongSafari",
      "type": "model-or-code-download",
      "link_text": [
        "Hugging Face"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/kuleshov-group/caduceus",
      "type": "model-or-code-download",
      "link_text": [
        "GitHub"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/ArcInstitute/evo2",
      "type": "model-or-code-download",
      "link_text": [
        "GitHub"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/google-deepmind/alphagenome",
      "type": "model-or-code-download",
      "link_text": [
        "API"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/google-deepmind/alphagenome_research",
      "type": "model-or-code-download",
      "link_text": [
        "weights"
      ],
      "sections": [
        "The Decision Tree: What Are You Actually Trying to Do?"
      ]
    },
    {
      "url": "https://github.com/AIRI-Institute/GENA_LM",
      "type": "model-or-code-download",
      "link_text": [
        "GitHub"
      ],
      "sections": [
        "Which DNA Model Fits Your Task?"
      ]
    },
    {
      "url": "https://huggingface.co/BGI-HangzhouAI/Genos-10B-v2",
      "type": "model-or-code-download",
      "link_text": [
        "Hugging Face"
      ],
      "sections": [
        "Which DNA Model Fits Your Task?"
      ]
    },
    {
      "url": "/blog/what-eight-studies-say-about-choosing-a-somatic-variant-caller/",
      "type": "companion-review",
      "link_text": [
        "Full review",
        "Choosing a somatic variant caller"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-the-benchmarks-say-about-choosing-an-rna-fusion-caller/",
      "type": "companion-review",
      "link_text": [
        "Full review",
        "Choosing an RNA fusion caller"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-eight-studies-show-about-detecting-cancer-from-cfdna-fragmentomics/",
      "type": "companion-review",
      "link_text": [
        "Full review",
        "Detecting cancer from cfDNA fragmentomics"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-seven-studies-show-about-detecting-cancer-from-plasma-dna-methylation/",
      "type": "companion-review",
      "link_text": [
        "Full review",
        "Detecting cancer from plasma DNA methylation"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-six-studies-show-about-detecting-and-classifying-copy-number-variants/",
      "type": "companion-review",
      "link_text": [
        "full CNV review",
        "Detecting and classifying copy-number variants"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-published-comparisons-show-about-dna-pathogen-identification-workflows/",
      "type": "companion-review",
      "link_text": [
        "Full review",
        "DNA pathogen identification workflows"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/rna-pathogen-detection-by-sequencing/",
      "type": "companion-review",
      "link_text": [
        "Full review",
        "RNA pathogen detection by sequencing"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-published-benchmarks-show-about-accelerating-variant-calling-and-genomic-models/",
      "type": "companion-review",
      "link_text": [
        "Full infrastructure and execution review",
        "Accelerating variant calling and genomic models"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2",
      "type": "issue-report",
      "link_text": [
        "issue #2"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows"
      ]
    },
    {
      "url": "/blog/spliceai2/",
      "type": "companion-review",
      "link_text": [
        "SpliceAI2 review",
        "SpliceAI2"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "https://github.com/rewire-bio/rewire-benchmark-data/issues/127",
      "type": "original-analysis",
      "link_text": [
        "database extraction"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/what-published-comparisons-show-about-confirming-splice-variants-in-patient-rna/",
      "type": "companion-review",
      "link_text": [
        "splicing review",
        "Confirming splice variants in patient RNA"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CURRENT-PREDICTORS.md",
      "type": "original-analysis",
      "link_text": [
        "results and methods"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Sources"
      ]
    },
    {
      "url": "/blog/running-evo-2-on-google-cloud-aws-and-runpod/",
      "type": "companion-review",
      "link_text": [
        "full benchmark and cloud study",
        "cloud study",
        "full cloud operations and reproducibility study",
        "cloud setup and cost guide",
        "cloud reproducibility study",
        "companion cloud study",
        "step-by-step Evo 2 cloud guide",
        "write-up"
      ],
      "sections": [
        "Task-Specific Performance: Coding (1.5%), Noncoding (98.5%), and Eight Diagnostic Workflows",
        "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
        "Practical Decision Matrix",
        "What These Models Can't Do (Yet)",
        "Conclusion",
        "Sources"
      ]
    },
    {
      "url": "https://github.com/rewire-bio/rewire-benchmarks/issues/42",
      "type": "original-analysis",
      "link_text": [
        "sizing protocol",
        "protocol"
      ],
      "sections": [
        "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
        "Sources"
      ]
    },
    {
      "url": "https://github.com/rewire-bio/evo2-cloud-operations",
      "type": "original-analysis",
      "link_text": [
        "study repository"
      ],
      "sections": [
        "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
        "Sources"
      ]
    },
    {
      "url": "https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/CLASSIFICATION-FLIPS.md",
      "type": "original-analysis",
      "link_text": [
        "classification flips"
      ],
      "sections": [
        "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
        "Sources"
      ]
    },
    {
      "url": "https://creativecommons.org/licenses/by/4.0/",
      "type": "license",
      "link_text": [
        "CC BY 4.0"
      ],
      "sections": [
        "What Scale Actually Buys You: Inside Evo and Evo 2"
      ]
    },
    {
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/sources.json",
      "type": "original-analysis",
      "link_text": [
        "source manifest"
      ],
      "sections": [
        "Sources"
      ]
    },
    {
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/reproduce_brca1.py",
      "type": "original-analysis",
      "link_text": [
        "reproduction script"
      ],
      "sections": [
        "Sources"
      ]
    },
    {
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/README.md",
      "type": "original-analysis",
      "link_text": [
        "instructions and data terms"
      ],
      "sections": [
        "Sources"
      ]
    },
    {
      "url": "https://github.com/rewire-bio/evo2-cloud-operations/blob/f8ae11c9a1c45505487c978b4cb7e064c4e59905/followup/RESULTS.md",
      "type": "original-analysis",
      "link_text": [
        "follow-up results"
      ],
      "sections": [
        "Sources"
      ]
    },
    {
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/pretraining-costs.json",
      "type": "original-analysis",
      "link_text": [
        "prices and inputs"
      ],
      "sections": [
        "Sources"
      ]
    },
    {
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/evo2-l4-brca1-20261016/README.md",
      "type": "original-analysis",
      "link_text": [
        "L4 follow-up",
        "measurements, scores and data terms"
      ],
      "sections": [
        "Hardware, Memory, and Fine-Tuning Requirements Across DNA Models",
        "Sources"
      ],
      "version": "Recovered result completed 2026-10-16; evidence bundle assembled 2026-10-16",
      "claims": [
        "All 3,893 variants completed at 8,192 bp, batch size one; allocated and sampled device memory are distinct.",
        "Scoring and allocation costs use distinct time boundaries and are estimates, not settled invoices.",
        "Spearman 0.9828845 is below the predefined 0.99 threshold; differing FP8 execution order prevents isolating a GPU effect.",
        "The broader model-size and CPU work is incomplete; no clean full-study reproduction was performed."
      ]
    },
    {
      "url": "/downloads/genomic-foundation-models-in-the-diagnostic-lab/evo2-l4-brca1-20261016.zip",
      "type": "download",
      "link_text": [
        "offline analysis bundle"
      ],
      "sections": [
        "Sources"
      ],
      "sha256": "918dd1942f9f32ebf7369012b73a18ac589ca638fd8e283d3043642acd7366a4",
      "note": "Preserved 16 October evidence snapshot. Recomputes point metrics without model inference; its draft/publication status is historical."
    }
  ]
}