Aligning protein generative models with experimental fitness via Direct Preference Optimization

Generative models trained on unlabeled protein datasets have demonstrated a remarkable ability to predict some biological functions without any task-specific training data. However, this capability does not extend to all relevant functions and, in many cases, the unsupervised model still underperforms task-specific, supervised baselines. We hypothesize that this is due to a fundamental "alignment gap" in which the rules learned during unsupervised training are not guaranteed to be related to the function of interest. Here, we demonstrate how to provide protein generative models with useful task-specific information without losing the rich, general knowledge learned during pretraining. Using an optimization task called Direct Preference Optimization (DPO), we align a structure-conditioned language model to generate stable protein sequences by encouraging the model to prefer stabilizing over destabilizing variants given a protein backbone structure. Our resulting model, ProteinDPO, is the first structure-conditioned language model preference-optimized to experimental data. ProteinDPO achieves competitive stability prediction and consistently outperforms both unsupervised and finetuned versions of the model. Notably, the aligned model also performs well in domains beyond its training data to enable absolute stability prediction of large proteins and binding affinity prediction of multi-chain complexes, while also enabling single-step stabilization of diverse backbones. These results indicate that ProteinDPO has learned generalizable information from its biophysical alignment data.