Abstract
<jats:p>Cryo-electron microscopy (cryo-EM) is increasingly used not only for structure determination but as a discovery tool for investigating heterogeneous and previously uncharacterized molecular assemblies. Realizing this potential hinges on identifying structures within experimental density maps-a task that grows harder as resolution declines. We developed CryoNet.Discovery, a deep-learning framework that aligns density maps and atomic structures in latent space, recasting structure identification through cross-modal retrieval. On CATH and SCOPe domain benchmarks, it achieved Top-1 accuracies of 85.7% and 84.4%, respectively, at low resolutions (6–10 Å), outperforming ModelAngelo, cryoID, and DomainFit across a wide resolution range, with the largest gains at low resolutions. Extending from domain- to single-chain identification, it scaled to 12,114 protein chains from 1,488 EMDB maps, recovering the correct fold for 89.3% of queries and revealing structurally conserved yet sequence-divergent relationships missed by sequence-based search. CryoNet.Discovery thus provides a scalable foundation for interpreting unresolved cryo-EM maps and advancing cryo-EM as an engine for structural discovery.</jats:p>