Reward-Driven Reinforcement Learning for Efficient and Low-Immunogenicity Protein Binder Design

Generative protein-design models can create binders against user-defined molecular targets, but current workflows remain so inefficient that designing binders to challenging epitopes is practically not possible. Current design pipelines operate in a design-and-filter paradigm in which only around 0.01% of generated designs pass the required filters. This limitation constrains the broader potential of protein design models; if these methods could be reliably applied to any target, they would significantly expand the range of biological problems that can be addressed, including the design of therapeutics and critical diagnostic tools.

This project will develop a general reinforcement-learning framework that can fine-tune protein-generative models towards any computable design objective. Rather than training models only to reproduce plausible protein structures, the framework uses post-generation evaluations as rewards to shift the model towards designs with desired structural, biological, or translational properties. These objectives can include foldability, interaction quality, specificity, developability, reduced predicted immunogenicity, or future experimental measurements.

As a first application, the researchers will fine-tune a state of the art protein design model to generate protein binders that are more likely to pass downstream structural and sequence-quality filters, with the longer-term goal of enabling multi-objective design that reduces predicted immunogenicity in protein design. The workflow combines target sampling, backbone generation, ProteinMPNN sequence design, independent structural evaluation, and Group Relative Policy Optimization. Preliminary Gefion experiments show that this approach can improve predicted design quality on PDB-derived targets not used during fine-tuning.

The requested 80,000 H100 GPU-hours will scale this framework across diverse targets and challenging epitopes, including polar and geometry-dependent interfaces. The team will benchmark the RL-fine-tuned model against all state-of-the-art design tools, measuring binder yield, sample efficiency, specificity, and predicted immunogenicity. The project will deliver an improved binder-design model, a reusable reward-driven fine-tuning framework, a checkpoint with de-risked immunogenicity, and a large dataset of generated, refolded, and reward-scored binders for future protein-design research.