ViruFunc Atlas v1.0: a leakage-aware benchmark and reusable evaluation resource for viral protein function annotation

ViruFunc Atlas v1.0 is a leakage-aware benchmark and reusable evaluation resource for viral protein function annotation. The release provides frozen manifests for 713,487 viral proteins, 19,149 genomes, 1,283 viral families, and 17 primary function labels. It includes family-heldout, host-heldout, default/hash, calibration, and strict-zero split resources; label ontology and functional-group mapping; feature-boundary rules; baseline prediction and metric tables; source tables for leakage audits, current-evidence triage, sequence-structure-context analyses, and the matched 160-protein Phold/PHROG-style evidence panel. The v1.0.1 resource revision adds a full-test Phold T4 annotation-refinement baseline for 66,837 family-heldout test proteins, including normalized Phold annotations, mapped-label metrics, the four-GPU run manifest, and a compact comparison summary. No new sequence data were generated. Source viral protein and genome records were derived from public NCBI/RefSeq/GenBank-linked resources. The release records accession identifiers, source databases, freeze date, retrieval scripts, and checksums. Large regenerable artifacts, including frozen pLM embeddings, model checkpoints, Foldseek/Phold databases, and predicted structure archives, are not included. The release provides identifiers, manifests, source tables, and scripts needed to rebuild these artifacts when the external databases and compute environment are available.