Abstract
<jats:p>Translating cryo-electron microscopy (cryo-EM) density maps into accurate atomic models requires refinement that maximizes map–model agreement and stereochemical validity, yet current workflows often remain laborious and expert-dependent. Here we introduce CryoNet.Refine, a physics-informed deep learning method that couples an AlphaFold 3-derived coordinate update module with a fully differentiable composite loss for fully automated atomic model refinement. We assess performance using a composite quality score (CS-score) integrating density-fit and geometry metrics. Across benchmarks spanning 225 AlphaFold 3–predicted starting models and 675 PDB-deposited structures of two deposition eras, CryoNet.Refine consistently outperforms Phenix.real_space_refine. On unrefined predicted models, CryoNet.Refine raises the mean CS-score from 0.64 to 0.91-versus 0.72 for Phenix.real_space_refine. On PDB-deposited structures-including post-2018 entries already optimized by Phenix.real_space_refine and expert manual correction—CryoNet.Refine achieves higher CS-scores in 90.9% of cases. Strikingly, it lowers the mean MolProbity score from 1.84 to 1.36, whereas Phenix.real_space_refine increases it to 2.02. Half-map cross-validation confirms that these gains reflect genuine structural signal rather than overfitting. CryoNet.Refine is available as open-source software and a web server to support reproducible, high-throughput cryo-EM refinement.</jats:p>