We introduce a flexible framework for antibody sequence design that combines an infilling language model with reinforcement learning to optimize functional properties. Our approach leverages a pretrained infilling language model to generate specific antibody regions within full sequences, guided by reinforcement learning to improve desired biophysical characteristics. We implement a range of online learning strategies, exploring both vanilla REINFORCE and Proximal Policy Optimization with Kullback-Leibler (KL) regularization, and demonstrate that KL regularization is essential for maintaining a balance between score optimization and sequence plausibility. We also adapt Direct Reward Optimization to the protein domain by adding a value head to the infilling model, allowing it to learn directly from static (prompt, response, feedback) datasets using a mean-squared error objective. This formulation is particularly useful when only single-trajectory data is available, which is commonly the case for historically collected experimental assays. We evaluate both the online and offline methods across multiple antibody design tasks (including binding affinity, immunogenicity, and expression) and show that our framework improves alignment with measured biophysical properties while outperforming likelihood-only baselines. This integrated online/offline approach enables functionally driven antibody design and provides a scalable toolkit for therapeutic sequence engineering. Code and data are available at https: //github. com/LLNL/proteinₜuneᵣl.
Lee et al. (Tue,) studied this question.