Data and software for: Does AlphaFold 3 Replace Protein–Protein Docking? A Blind Benchmark Against Fast-Fourier-Transform Docking and an Appraisal of Interface Confidence

Abstract

Complete data and software accompanying a benchmark of template-free AlphaFold 3 co-folding against blind, clustering-reranked fast-Fourier-transform docking on 20 protein–protein complexes from Docking Benchmark 5, spanning the rigid-body, medium and difficult tiers and three interface classes, together with negative controls probing what the AlphaFold 3 interface confidence score (ipTM) measures. The archive contains: machine-readable benchmark results (per-complex interface class, difficulty tier, ipTM, pTM, ranking score, DockQ for top-ranked and best-of-five models, and DockQ under all three docking pose-selection schemes); ipTM and ranking scores for all five returned models of every AlphaFold 3 job; ipTM values for cognate complexes, same-fold wrong-partner controls and random non-cognate controls; the exact job specifications submitted to the AlphaFold Server; the unbound receptor and ligand structures supplied to MEGADOCK; all returned AlphaFold 3 models with per-model confidence data and predicted aligned error matrices; the 20 experimental reference complexes used for scoring; and the complete scoring, clustering, analysis and figure-generation code with software versions and the commands that regenerate every figure and table. Raw MEGADOCK pose files are omitted for size and regenerate deterministically from the included inputs by the documented command. Benchmark complexes derive from Docking Benchmark 5 (https://github.com/haddocking/BM5-clean) and ultimately from the Protein Data Bank. Data, structures and results are released under CC BY 4.0; the scripts under MIT.