#20 · Structural bio & protein design

AI structure models extended to protein conformational dynamics and ensemble sampling

The common thread is the extension of structure prediction and physics-informed modeling from single static conformations to ensembles and dynamics. Residue-residue distogram distributions from Boltz-2 are used to encode structural dynamics as representations for function prediction. Distance-restraint guidance applied to Boltz-2-type diffusion models samples conformational states and ligand dissociation pathways. bAIes couples AlphaFold2 predictions with molecular mechanics force fields to produce atomic-resolution ensembles of intrinsically disordered proteins. RNAnneal extends ensemble prediction to RNA 3D structure by combining ab initio sampling, unsupervised deep learning scoring, and interaction entropy. FrustrAI-Seq is the least structural in input, predicting local energetic frustration from sequence via protein language model embeddings, yet it targets a structural energetics quantity. What is new is that generative structure models are steered by explicit restraints and energetics to probe states beyond the predicted native fold.