Benchmarking DNA foundation models for genomic and genetic tasks: Supplementary Information

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 2 protocols
• 47 datasets
• 6 evidence claims

Claims

Statements this source supports, each with where in the source it was checked.

• Relevance of feng-dna-fm-20261013-protocol-gtex-wb-expression-6kb to "Select regulatory variants and genes for functional follow-up": Assessed by Feng GTEx whole-blood expression prediction, all genes, 6 kb input (evidence-expansion-dna-foundation-models-2025-5d8ca9bc: Table 4)
• Relevance of feng-dna-fm-20261013-protocol-gtex-wb-expression-long-768 to "Select regulatory variants and genes for functional follow-up": Assessed by Feng GTEx whole-blood expression prediction, 768 genes, long input (evidence-expansion-dna-foundation-models-2025-5d8ca9bc: Table 4)
• reported_finding: feng-dna-fm-20261013-protocol-gtex-wb-expression-long-768: Reported finding Among long-input models, HyenaDNA-450K was significantly better than Enformer (p < 0.01) and Enformer was similar to Caduceus-Ph. On the same 768 genes, extending the input improved HyenaDNA (average correlation 0.122 short versus 0.137 long, two-sided Wilcoxon p = 0.0005) but not Caduceus-Ph (0.124 versus 0.127, p = 0.975). (Results, 'Gene expression prediction', paragraph 2; Supplementary Table 4)
• reported_finding: feng-dna-fm-20261013-protocol-tad-attention-nt-v2: Reported finding The difference between average NT-v2 attention for TAD-centred and background sequences showed no distinct pattern: values were near zero apart from the diagonal. The authors conclude that NT-v2 self-attention does not recognise TAD boundaries without fine-tuning. The heatmap is Supplementary Fig. 2 (Supplementary Information page 11). (Results, 'TAD region recognition', paragraph 2; Supplementary Fig. 2 (Supplementary Information page 11))
• tables_not_extracted: evidence-expansion-dna-foundation-models-2025-5d8ca9bc: Tables not extracted Not extracted as results: Table 2 rows other than the two Arabidopsis promoter rows (bacterial promoters, human vs worm, mouse TFBS); Table 3 except the Human 5mC Caduceus-Ph and NT-v2 cells (human 6mA, yeast histone marks, 4mC in six species); Supplementary Tables 1-4 (pooling counts, multi-class accuracy, random forest vs XGBoost, short vs long input); Supplementary Data 3 (top genes), 4 (per-chromosome-group AUCs behind Tables 5 and 6) and 5 (HyenaDNA re-pretraining); Supplementary Tables 5-8 (hyperparameter grids, embedding dimensions) and Supplementary Fig. 1 (classifier comparison); Figures 2-4 including runtime. These are outside the human diagnostic and regulatory use cases or repeat main-table values at another aggregation. (Tables 2-3; Supplementary Tables 1-8 and Supplementary Fig. 1; Supplementary Data 3-5; Figures 2-4)
• tables_not_extracted: evidence-expansion-dna-foundation-models-2025-5d8ca9bc: Tables not extracted Not extracted as results: Cohen's d per test group (Supplementary Data 4 prints AUC only); Supplementary Tables 1-4 (pooling counts, multi-class accuracy, random forest vs XGBoost, short vs long input); Supplementary Data 3 (top genes) and 5 (HyenaDNA re-pretraining); Supplementary Tables 5-8 (hyperparameter grids, embedding dimensions) and Supplementary Fig. 1 (classifier comparison); Figures 2-4 including runtime. Tables 1-6 and Supplementary Data 4 are now fully extracted. The remaining parts are outside the use cases or repeat main-table values at another aggregation. (Supplementary Tables 1-8 and Supplementary Fig. 1; Supplementary Data 3-5; Figures 2-4)

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.