Evaluation of methods for AlphaFold-based integrative modeling

AlphaFold-based integrative modeling methods (AlphaLink2, Boltz2, GRASP) satisfy chemical crosslinks in a benchmark, but often yield nearly identical structures with or without input data.

• Why it matters: It is unclear how these approaches trade off experimental constraints against AlphaFold’s learned structural priors, which matters when crosslinks are noisy or reflect multiple conformational states.
• What they did: The authors curated a non-redundant benchmark of homomeric and heteromeric protein complexes outside the AlphaFold training set and tested AlphaLink2, Boltz2, and GRASP using chemical crosslinks as input.
• The result: When 90% of crosslinks are consistent with one structure, all methods match them well, yet for most complexes predictions barely change without crosslinks, remain similar as false-positive fractions rise, and stay biased toward a single state under multi-state inputs.

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