Atomic Protein Structure Modeling from Cryo-EM Using Multi-Modal Deep Learning and AlphaFold3

Abstract

Cryo-electron microscopy (cryo-EM) has revolutionized structural biology by enabling near-atomic resolution visualization of protein structures. However, accurately modeling 3D atomic structures from cryo-EM density maps remains challenging, particularly for large multi-chain complexes and assemblies. Here, we present an automated pipeline that integrates multi-modal deep learning and advanced structure prediction techniques to improve model accuracy. Our approach leverages a deep learning model that combines sequence-based features from a Protein Language Model with cryo-EM density maps, enabling a richer feature representation across modalities. The deep learning-predicted voxels are utilized to build a Hidden Markov Model (HMM) and a tailored Viterbi algorithm is used to align sequences to generate an initial protein backbone structures. These backbone models then serve as templates for AlphaFold3, which refines the structures for improved accuracy. Our approach combines cryo-EM data with AlphaFold3 predictions, helping to refine and improve AlphaFold3s predicted structures. By integrating both methods, we can generate more accurate and reliable atomic models, particularly for large proteins with complex conformations.