Proteins carry out a wide range of functions informed by their structure, serving as biocatalysts, transporters, fluorophores, and more. Optimizing protein function has broad applications in research, medicine, and industry. However, progress is often limited by the slow process of designing and testing mutants. Computational tools offer a powerful solution to this problem, accelerating protein optimization and enabling more efficient discovery. We developed a neural network model combined with a gradient-ascent-based optimization protocol to design variants of green fluorescent protein (GFP). As an intermediate check, we used evolutionary scale modeling 2 (ESM), a large protein language model, to evaluate the predicted variants before experimental validation. Our optimizations have yielded two GFP mutants that outperform the industry standard, GFP superfolder, in terms of fluorescence. Moreover, the protocol was able to predict mutations that enabled the recovery of fluorescence in dark GFP mutants while producing a significant increase in their thermostability. This optimization protocol highlights a powerful approach for functional optimization of proteins, with wide-ranging implications for protein engineering.
Longfritz et al. (Sun,) studied this question.