N

Predicted monomer structures for 292,804 unique protein sequences: 38,934 from a Gene Ontology molecular-function coupling set and 254,182 audited ancestral sequence reconstructions (312 sequences occur in both). The predictions serve as teacher labels for 1,421,494 training pairs; one structure is reused by every pair referencing that exact sequence. Sequence IDs are seq, so a prediction can always be matched back to its exact sequence. Stock open-source AlphaFold 3 with default settings — no protocol Both MSAs AlphaFold 3 builds are retained: the unpaired MSA and the paired (UniProt) MSA. The paired MSA is kept even though every input is a single chain, because AlphaFold 3 featurises it…