PolyFold: Homology Filtering Approach Alters Apparent Structure Prediction Model Performances

Length-weighted average fractional identity (WAFI) homology filtering reduces residual training-data leakage and narrows or removes Boltz-2’s performance advantage over OpenFold 3 across most partitions.

• Why it matters: Benchmarking structure prediction models is vulnerable to nonrandom failure patterns and apparent gains from insufficiently conservative homology filtering.
• What they did: PolyFold automates inference and evaluation on several thousand homology-filtered PDB structures, comparing Boltz-2 and OpenFold 3, and uses cluster analysis to link structural features to performance profiles.
• The result: Conventional homology filtering favored Boltz-2, but minimum chain fractional identity novelty filtering can inflate metrics in multichain complexes; WAFI yields a more conservative evaluation while both models still struggle with homomultimers, small molecule realism, and non-canonical amino acids.

AI summary of the abstract · may contain errors · not medical advice · read the paper.