#6 · Tools & infrastructure

Benchmarks expose setting-dependent limits of learned protein-ligand binding and docking

The dominant thread is evaluation of learned and physics-based methods for predicting protein-ligand binding under demanding conditions. Bento benchmarks 11 classical, deep learning and co-folding tools on drug-design data, and PoseBench tests deep learning docking and cofolding on apo-to-holo, multiligand and unknown-pocket settings. A separate assessment reports that Glide docking to AlphaFold models performs better with fragments than with unbound crystal structures, and that ensembles of AlphaFold models improve docking of larger ligands. Related items extend the scope to binding dynamics, with NeuralMD as a group-symmetric neural differential equation surrogate for protein-ligand molecular dynamics, and to binding-site prediction, with PRECISE and CholBindNet. What is new is the emphasis on setting-specific, benchmark-driven assessment of structure-based interaction models.