Integrating protein sequence embeddings with structure via graph-based deep learning for the prediction of single-residue properties

Abstract

Understanding the intertwined contributions of amino acid sequence and spatial structure is essential to explain protein behaviour. Here, we introduce INFUSSE (Integrated Network Framework Unifying Structure and Sequence Embeddings), a Deep Learning framework that combines sequence embeddings, generated by a Large Language Model (LLM), with graph-based representations of protein structures, integrated through a diffusive Graph Convolutional Network (diff-GCN), to predict single-residue properties within proteins. Our approach follows two steps. First, we fine-tune LLM sequence embeddings obtained from bidirectional transformers to make predictions from protein sequence alone. Second, we combine these enriched sequence representations with a geometric graph Laplacian within diff-GCN to refine the initial predictions. This approach leads to improved predictions while allowing us to systematically disentangle the contribution of sequence and structure. We illustrate our framework by applying it to the prediction of local residue flexibility (B-factors) of antibody-antigen complexes, and show that it provides improved performance compared to current Machine Learning (ML) approaches. The addition of structural information via geometric graphs is shown to enhance predictions especially for intrinsically disordered regions, protein-protein interaction sites, and highly variable amino acid positions.