MAXWELL: Calibrating the probabilistic outputs of protein language models to the mutation-induced stability change landscape

MAXWELL calibrates ProteinMPNN’s probability outputs to predict mutation-induced stability changes, outperforming ThermoMPNN on a benchmark of experimentally measured effects.

• Why it matters: Protein engineering needs cheaper ways to find stabilizing mutations than large experimental screens, but current language-model stability methods often underuse the amino-acid probability distributions they learn.
• What they did: MAXWELL is a post-training calibration method that learns from the mutation-induced stability change landscape, then generates mutational landscapes for individual substitutions; it was applied to ProteinMPNN and benchmarked against ThermoMPNN.
• The result: In the DhaA dehalogenase DhaA system, MAXWELL-designed ten single-point mutations yielded seven stabilizing variants (70%), with G171W showing the largest measured {Delta}T of 4.91.

AI summary of the abstract · may contain errors · not medical advice · read the paper.