#23 · ML / AI methods

Generative binder design pairs structure proposals with humanness and epitope evaluation

The dominant thread is computational design of protein binders and antibodies using generative and learning-based models over sequence and structure. ProteoCraft treats binder design as tree-based search, using diffusion-model structural gradients as proposals and no-gradient folding verification. GLOM-TGITO pairs a multiscale global-local fluorescence predictor with multiview landscape-guided search for offline protein design. IgScope scores humanness per chain, IMGT region, and residue, exposing CDR gaps in de novo generated antibodies. EpiRank predicts epitopes from single-chain geometry and is evaluated as partner-agnostic patch retrieval with leakage controls. What is new is evaluation layered onto generation: verification loops, region-resolved humanness, and label-limited immunogenicity benchmarks, where a format lookup beats sequence predictors. The AlphaFold pocket study adds that most predicted pockets are holo-like, which affects pocket detection and docking.