Aligning protein-generative models to experimental fitness with ProteinDPO

TL;DR - ProteinDPO uses direct preference optimization to align a structure-conditioned protein language model with experimental biophysical fitness. It delivers stability predictions competitive with task-specific models while outperforming unsupervised and fine-tuned baselines.

• Applies DPO to protein generation using experimental fitness information.
• Aligns an unsupervised, structure-conditioned language model with biophysical properties.
• Demonstrates that preference optimization can improve protein stability prediction.
• Consistently surpasses the model’s unsupervised and conventionally fine-tuned versions.