Extended Data Fig. 5: BRCA1/2 variant effect prediction and DART-eval for regulatory genomics.

(a) Zero-shot prediction of BRCA1 variant pathogenicity for coding and noncoding SNVs evaluated in aggregate, showing AUROC and AUPRC scores across models. (b) Zero-shot prediction of BRCA1 variant pathogenicity for noncoding SNVs separated according to “intronic” and “splice site” variants based on whether the variant site is located more than 8 bp away from the intron-exon boundary²⁶. (c) Zero-shot prediction of BRCA2 variant pathogenicity for coding (left), noncoding (center), and aggregated (right) SNVs, showing AUROC and AUPRC scores across models. (d) Comparison of AUROCs for the supervised classification of BRCA1 SNVs using embeddings from Evo 2 40B block 20, for pooling functions that use different window sizes around the variant site to average embeddings. AUROCs are averaged from 5-fold cross-validation on the training set, and are from the best-performing ridge regression regularization parameter for each pooling function. (e) Comparison of AUROCs for the supervised classification of BRCA1 SNVs trained with embeddings from different blocks of Evo 2 40B and Evo 1. AUROCs are averaged from 5-fold cross-validation on the training set, and are from the best combination of pooling function and ridge regression hyperparameter α for each layer. (f) DART-Eval results for zero-shot regulatory DNA tasks. Task 1 evaluates models on their ability to distinguish candidate cis-regulatory elements (cCREs) from shuffled sequences, Task 2 tests their ability to identify transcription factor (TF) motifs by distinguishing true TF binding sites from control sequences, and Task 5 predicts variant effects for chromatin-accessibility QTLs (caQTLs, African) and dynamic sequence QTLs (dsQTLs, Yoruban). Sequence likelihoods were computed under each model and used to measure classification accuracy with mean AUROC. Across these tasks, Evo 2 7B and Evo 2 40B outperformed other baselines. For Task 5, while we did not see strong signal across all models when using the zero-shot log-likelihoods, Evo 2 embeddings were predictive of noncoding variant effects.