Discrete Diffusion Model for Antibody Sequence Generation

A new method for generating antibody sequences using discrete diffusion models has been introduced. The approach addresses two key limitations of existing protein language models: memorization of germline sequences and limited support for conditional generation. The authors propose discrete diffusion fine-tuning to achieve strong language modeling performance while enabling generation conditioned on any off-the-shelf classifier. They also introduce germline absorbing diffusion, a modification of the noise process that incorporates germline sequences. The work was published on arXiv as preprint 2605.06720.