ProteinDPO aligns protein-generative models to experimental fitness using DPO

DeryaTR_ · x · 2026-08-15

Published in Nature Methods, ProteinDPO uses Direct Preference Optimization (DPO) to align a structure-conditioned protein language model with experimental fitness. The approach preserves general pretraining knowledge while preferentially generating stable protein sequences. ProteinDPO achieves stability prediction competitive with specialized models, outperforms unsupervised and fine-tuned baselines, and generalizes to stabilize and improve binding affinity predictions for large multichain proteins.

Related event: ProteinDPO Applies LLM Alignment Technique to Protein Design(4 posts)→

Original post →

More from Research

Research channel →