bAIes drops its AlphaFold2-only constraint: six structure predictors yield comparable atomic-resolution ensembles of disordered proteins
Synopsis
bAIes originally drew structural information only from AlphaFold2 distograms to build atomic-resolution conformational ensembles of intrinsically disordered proteins. It now accepts six predictors, AlphaFold2, AlphaFold3, Boltz-1, Boltz-2, ESMFold and ESMFold2, and was compared on NMR and SAXS data for 17 disordered proteins. Ensembles from every predictor reach average accuracy comparable to the original AlphaFold2 version, best on local observables such as chemical shifts, RDCs and 3J couplings, and weaker on long-range and global observables such as PREs and SAXS.
Interpretation
Disordered proteins lack a single stable structure, and their function depends on dynamic conformational ensembles, while experiments supply only sparse, ensemble-averaged structural information. bAIes treats the inter-residue distance distributions produced by AI structure predictors as restraints and samples atomic-resolution ensembles by molecular dynamics. Previously bAIes used only AlphaFold2 distograms; after AlphaFold2, predictors with different architectures and training data appeared, and whether they could serve equally for disordered-protein ensemble generation had not been compared systematically. Across 3954 structure predictions on the 659-protein PeptoneDB-CS benchmark, the six predictors share 84–93% of contacts for ordered proteins but only 50–85% for disordered ones, with lower agreement on long-range contacts.
bAIes ensembles built from all six predictors reach average accuracy against experimental data comparable to the original AlphaFold2 implementation, indicating the method is insensitive to predictor choice. On the 17-protein NMR and SAXS benchmark, every predictor gives a similar improvement in chemical-shift agreement over the random-coil model; on Aβ40, PaaA2, ACTR and α-synuclein, the ESMFold-based ensembles agree with experiments better than, or within the error of, all-atom simulations with the a99SB-disp force field. On the 16 disordered proteins of PeptoneDB-Integrative, bAIes is on average the most accurate on chemical shifts together with the coil model, and bAIes-ESMFold is slightly ahead on RDCs; on SAXS and PREs the errors and standard deviations are larger.
Perspective
Validation covers 17 disordered proteins with NMR and SAXS data and 16 well-characterized disordered proteins in PeptoneDB-Integrative, with three conditionally folding proteins excluded from the ensemble comparison. The intended readers are structural-biology and computational-chemistry researchers who need atomic-resolution disordered-protein ensembles.
All bAIes ensembles agree more weakly on PREs and SAXS, consistent with predictors producing few long-range contacts for disordered proteins. Condition-dependent folding proteins are predicted as folded by every predictor, and bAIes inherits this behavior. The benchmark size does not establish that ESMFold systematically outperforms the other predictors.