#5 · ML / AI methods

Sequence- and structure-based models predict protein function and dynamics

The dominant thread is predictive modeling of protein and enzyme properties from sequence, structure, and learned embeddings, with explicit attention to whether reported gains reflect genuine generalization. Several items address out-of-distribution and leakage concerns: Aiki-XP applies leakage control to pan-bacterial relative expression prediction, O2DENet targets OOD generalization for enzymatic kinetic parameters, and PIMetaKcat adds Arrhenius constraints to few-shot temperature-dependent kcat prediction. Evaluation discipline is most explicit in MULTI-evolve, whose multimutant predictions track an additive model with no evidence of epistasis learning beyond additive baselines. A Rosetta-designed protein study uses experimental stability data and deep mutational scanning to identify steric overpacking failures and refit an energy function. Benchmarks such as GeomMotif formalise geometric motif preservation in generation, while PROTEUS and PLM-clust derive functional and plasticity signals from pretrained embeddings. The CirA Siamese graph network remains a distinct application linked to cefiderocol susceptibility.