Protein language models learn patterns from natural protein sequences. That can help them propose sequences that look plausible, but it does not automatically optimize the particular job a researcher cares about, such as brighter fluorescence or stronger binding. The new method adds measured laboratory performance to the training loop. ^[1]

The researchers call the method Reinforcement Learning from eXperimental Feedback, or RLXF. They first test protein variants, use those measurements to teach a reward model what better performance looks like, and then steer a protein language model toward new candidates. Selected candidates still have to return to the laboratory for testing. ^{[1] [2]}

The paper applies this cycle to five different protein families. Its accompanying preprint reports useful low-data results with tens to hundreds of measured variants, but the amount and quality of data needed will vary by protein and objective. ^{[1] [2]}

For CreiLOV, a fluorescent protein that works without oxygen, the aligned models proposed variants with stronger cellular fluorescence. The authors describe the best variants as the brightest CreiLOV variants reported so far. That comparison belongs to this protein family and these assays; it is not a general claim that the method will improve every protein. ^[1]

The preprint also shows why testing matters: simply combining individually promising mutations could fail because mutations can change one another's effects. The feedback loop is meant to navigate those interactions, not replace experiments with predictions. ^[2]

This is evidence that measured experiments can guide an AI model toward stronger candidates in several laboratory protein-engineering tasks. It does not report a therapy, a study in people or a real-world health outcome. Independent teams will need to test how reliably the method transfers to other proteins, laboratories and goals. ^{[1] [2]}