Benchmarks

Accuracy on the PubMedQA labelled split.

Entity-level F1 for chemical mentions on BC5CDR.

Three-state secondary structure accuracy.

Blood–brain barrier penetration, ROC-AUC.

Mean AUROC across the CheXpert competition findings.

Spearman correlation between zero-shot model scores and measured fitness across the ProteinGym deep mutational scanning assays.

Share of predicted ligand poses that are within 2 Å RMSD and pass the PoseBusters physical-validity checks.

Matthews correlation across the eighteen downstream genomics tasks released with the Nucleotide Transformer.

Classification accuracy across the Genomic Benchmarks regulatory sequence datasets.

Top-k accuracy on donor and acceptor site prediction, the metric reported for SpliceAI.

Base-pair F1 for predicted RNA secondary structure.

Macro-averaged F1 for cell type assignment on held-out donors or tissues.

Correlation between predicted and measured change in expression for perturbations held out of training.

HIV replication inhibition on the MoleculeNet scaffold split, ROC-AUC.