BPS2026 – Evaluating multi-state accuracy of AI protein structure predictors

Deep learning has markedly improved protein structure prediction, yet significant challenges persist in modeling both accuracy and conformational heterogeneity. We systematically evaluated AlphaFold3, BioEmu, Boltz-2, and Chai-1 on proteins with experimentally resolved apo and holo states, including the MATE family efflux transporter PfMATE, the lysine/arginine/ornithine-binding protein (LAOBP), and the protein translocase subunit SecA. While holo-state predictions generally aligned well with experimental structures, apo-state predictions were often biased toward holo-like conformations, underrepresenting the diverse ensembles characteristic of ligand-free proteins. For PfMATE, which adopts inward- and outward-facing apo conformations, all models predicted only the outward-facing state, missing the rare inward-facing structure. Similarly, apo predictions of LAOBP and SecA collapsed toward holo or intermediate states instead of reflecting true apo diversity. These findings reveal a systematic bias: current deep learning models preferentially generate stable holo-like folds while underrepresenting the structural variability and rare conformations characteristic of ligand-free states.