Accurate Prediction of Protein Tertiary and Quaternary Stability Using Fine-Tuned Protein Language Models and Free Energy Perturbation.

Methods such as AlphaFold have revolutionized protein structure prediction, making quantitative prediction of the thermodynamic stability of individual proteins and their complexes one of the next frontiers in computational protein modeling. Here, we develop methods for using protein language models (PLMs) with protein mutational datasets related to protein tertiary and quaternary stability. First, we demonstrate that fine-tuning of a ProtT5 PLM enables accurate prediction of the largest protein mutant stability dataset available. Next, we show that mutational impacts on protein function can be captured by fine-tuning PLMs, using green fluorescent protein (GFP) brightness as a readout of folding and stability. In our final case study, we observe that PLMs can also be extended to protein complexes by identifying mutations that are stabilizing or destabilizing. Finally, we confirmed that state-of-the-art simulation methods (free energy perturbation) can refine the accuracy of predictions made by PLMs. This study highlights the versatility of PLMs and demonstrates their application towards the prediction of protein and complex stability.