Enhancing interpretability of cryo-EM maps with hybrid attention Transformers

Abstract

Cryo-electron microscopy (cryo-EM) has become a central technique for determining the structures of macromolecular complexes. However, experimental cryo-EM density maps often exhibit limited interpretability due to noise and heterogeneous, anisotropic resolution, complicating downstream atomic model construction. Here we present DEMO-EMReF, a Swin Transformer-based framework specifically designed for cryo-EM density map refinement that integrates channel attention with hybrid spatial attention. By combining local window-based attention with grid-based sparse global attention, DEMO-EMReF captures fine-grained local density features while modeling long-range, cross-region spatial dependencies, enabling more robust density refinement in heterogeneous and anisotropic map regions. DEMO-EMReF was systematically evaluated on large benchmark sets of both primary maps and half-maps spanning resolutions of 3.0-8.0 [A] and was compared against multiple state-of-the-art map enhancement methods. DEMO-EMReF consistently improves resolution-related metrics, map-model correlations, and local atomic resolvability, with particularly robust gains in challenging heterogeneous or anisotropic cases. Importantly, the refined maps facilitate more accurate and efficient automated atomic model building. Together, DEMO-EMReF provides a robust approach for enhancing cryo-EM density maps and enabling more reliable downstream studies.