Mixture diffusion model for multimodal antibody design

Antibody design requires modeling complementarity-determining region (CDR) loops that are highly flexible and adopt diverse conformations to achieve high-affinity antigen binding. Current diffusion-based generative models almost universally adopt unimodal distributions to parameterize sequence-structure transitions, which produce smooth conformations but constrain generation to a single conformational mode. This limitation impedes the exploration of alternative high-affinity binding conformations, particularly for challenging targets where exceptional binders may exist in low-probability regions of the conformational space. To address this, we introduce the mixture diffusion model for multimodal antibody design, a denoising diffusion probabilistic model that uses mixture density parameterizations for both positional and rotational updates. Through experiments on antibody-antigen complexes from the Structural Antibody Database (SAbDab), we find that increasing the number of mixture components improves model performance by capturing distinct canonical-like backbone conformations of CDRs. Our model achieves competitive amino acid recovery and binding-affinity-related metrics while maintaining physically consistent backbones. Through our results, we establish mixture-based diffusion modeling as a practical path toward discovering high-quality antibody conformations that remain inaccessible to conventional single-mode diffusion frameworks.