A learning-to-rank framework for large-scale identification of unknown proteins from cryo-EM and cryo-ET maps

Abstract

Cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) can yield density maps whose molecular identities are unknown. Predicted-structure databases enable proteome-scale searches, but distinguishing the correct candidates remains challenging, particularly for intermediate-resolution maps without clear side-chain features. We developed CryoAssign, an automated framework that combines large-scale structure matching with learning-to-rank prioritization. CryoAssign searches intact density maps without prior segmentation and integrates complementary measures of map-model agreement. On held-out benchmark maps, CryoAssign robustly prioritized the correct candidates. It also identified proteins or narrowed the set of candidates in previously unassigned cryo-EM maps and recovered multiple components from complex cryo-EM and cryo-ET maps. Controlled low-pass filtering showed generally robust recovery across a broad range of intermediate resolutions. CryoAssign provides a scalable approach for identifying proteins from cryo-EM and cryo-ET density maps.