Bonobo: Efficient Library-Scale Generation for De Novo Antibody Design

Recent developments in de novo antibody design show promise for generating candidates that bind to drug targets. Methods which have shown success in the lab propose designs with either a diffusion model or hallucination-based sequence optimization. They then filter these candidates with a structure prediction model. Hallucination-based methods rely on expensive backpropagation through a loss function derived from the structure prediction model to perform per-generated-sequence optimization, limiting their ability to generate designs at the scale of the libraries, which may be needed for challenging targets. We propose Bonobo, which instead trains a generative model on the structure prediction loss using a GFlowNet formulation in which we transform the loss to act as a reward function. This does not require differentiation of the structure predictor, increasing our computational efficiency and unlocking a broader class of structure-based models for usage. Crucially, our approach amortizes the cost of generation into training time, enabling Bonobo to generate library-scale sets of diverse sequences at a significantly lower cost. We show empirically that our approach can effectively model complex loss functions and generate large numbers of high-performing novel antibody sequences for a range of target proteins.