Structural biology pipeline uses differentiable density fitting to speed up ligand-pocket modeling

A team described a software redesign of the cryo-EM model-building workflow, starting from an informal conversation about the gap between raw density maps and chemically plausible binding pockets. The pipeline separates data ingestion, which validates and stores map files in a database, from a worker pool that runs a containerized pocket-finder. The pocket-finder uses a Monte Carlo fragment-placement search guided by a differentiable density-correlation loss and a graph neural network that predicts side-chain angles, trained on publicly available PDBbind structures. Reducing the sampler's proposal count lowered GPU memory use substantially with only a small increase in pose error, enabling overnight processing of hundreds of maps. The authors report a 98 percent pocket-detection success rate and about 1.3 minutes of wall time per ligand.