EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation

EvoStruct addresses a significant limitation in antibody complementarity-determining region (CDR) design, where equivariant graph neural network (GNN) methods suffer from vocabulary collapse, over-predicting certain amino acids like tyrosine and glycine while neglecting crucial residues. This issue stems from GNN encoders failing to capture complex amino acid distributions. By integrating evolutionary and structural priors, EvoStruct adapts protein language models to improve CDR design, potentially enhancing the discovery of functional antibodies. The approach leverages the strengths of both evolutionary and structural methods, offering a more comprehensive understanding of antibody design¹. This development has significant implications for biotechnology and pharmaceutical applications, as improved antibody design can lead to more effective treatments and therapies. The ability to accurately predict and design antibody CDRs can revolutionize the field of immunotherapy, making it a crucial area of research for practitioners and scientists.