#24 · Structural bio & protein design

Learned models expand protein conformational sampling and sequence-energy fitting

The shared thread is moving structural modeling from single static predictions toward ensembles, energy functions, and compact representations. ProCEDiS uses a neural surrogate to steer MSA recombination with AlphaFold2, producing diverse conformations that feed short MD runs. AFsample3 extends AlphaFold3 sampling to generate and select multiple conformational states, with improved alternate-state accuracy reported on 238 proteins. PottsMPNN argues that native sequence recovery is misaligned with design quality and energy prediction, and instead learns Potts energy functions. Structural alphabets (Q3, Q8, 3Di) are benchmarked against full TM-score structural alignment for remote homology on CATH and SCOPe, measuring how much fold information they retain. The novelty lies in evaluating objectives beyond sequence recovery and in representing conformational diversity explicitly rather than as a single fold.