IgBlend: Unifying 3D Structures and Sequences in Antibody Language Models

Abstract

Large language models (LLMs) trained on antibody sequences have shown significant potential in the rapidly advancing field of machine learning-assisted antibody engineering and drug discovery. However, current state-of-the-art antibody LLMs often overlook structural information, which could enable the model to more effectively learn the functional properties of antibodies by providing richer, more informative data. In response to this limitation, we introduce IgBlend, which integrates both the 3D coordinates of backbone atoms (C-alpha, N, and C) and antibody sequences. Our model is trained on a diverse dataset containing over 4 million unique structures and more than 200 million unique sequences, including heavy and light chains as well as nanobodies. We rigorously evaluate IgBlend using established benchmarks such as sequence recovery, complementarity-determining region (CDR) editing and inverse folding and demonstrate that IgBlend consistently outperforms current state-of-the-art models across all benchmarks. Furthermore, experimental validation shows that the models log probabilities correlate well with measured binding affinities.