Multimodal Protein Retrieval via Joint Representation Learning from Sequences and Cryo-EM Density Maps

Aligning protein sequences with cryo-EM density maps remains challenging due to limited paired data, structural heterogeneity, varying map resolutions, and the presence of multiple conformational states. In this work, we propose a multimodal representation learning framework that learns a shared latent space between protein sequences and cryo-EM density maps for cross-modal retrieval. Our approach combines pretrained protein sequence embeddings with a volumetric cryo-EM encoder trained using self-supervised representation learning and transfer learning. The resulting model enables bidirectional retrieval between sequences and density maps while learning biologically meaningful structural representations. Experimental results demonstrate strong retrieval performance across both sequence-to-map and map-to-sequence tasks, achieving median retrieval ranks of 2--3 within a database of 3,275 cryo-EM maps. The learned embedding space shows a clear separation between matched and unmatched sequence--map pairs and remains robust across varying cryo-EM resolutions. Additionally, the model generalizes across species, successfully retrieving conserved mouse protein structures using human sequence embeddings. Our findings demonstrate that joint latent-space learning provides a promising direction for connecting protein sequences with cryo-EM structural representations, with potential applications in structural retrieval, protein annotation, and multimodal biological representation learning.