#4 · Structural bio & protein design

Benchmarks probe AlphaFold model accuracy across peptides, RNA, and antibody complexes

The dominant thread is computational assessment of predicted structural models at the level of interfaces, binding relevance, and local reliability. A systematic benchmark of AlphaFold2 and OpenFold3 on 271 protein-peptide complexes, scored with CAPRI criteria, finds AF2 more accurate and shows evidence of memorization. Engineering of AlphaFold3 MSA and template inputs targets monomer, multimer, and protein-ligand predictions, reported with TM-score and DockQ. CASP16 RNA analysis links multimer interface errors to monomer accuracy and model selection. SNAC-DB provides antibody and VHH-antigen complexes with 31-37% more structural diversity than SAbDab, with benchmarks of seven predictors. AlphaUnfold couples AlphaFold3 with short high-pressure MD to flag low-pLDDT regions as fragile. Binding free energy methods are compared on shared Amber force fields, and Martini Mapper automates Martini 3 coarse-grained model construction from SMILES.