Probing the Stability of Designed Protein

Naturally occurring proteins have evolved to perform specific biochemical functions, and their amino acid sequences are optimized for those functions rather than for stability. Enhancing protein stability without compromising its function is highly desirable for biotechnological applications, such as vaccine design. Computational protein design algorithms offer a solution by making targeted amino acid changes to the native protein sequence, potentially enhancing stability without the need for the lengthy, labor-intensive, and costly experimental processes typically involved in stabilizing proteins. There are various categories of computational protein design algorithms, and it remains unclear which approach is the most effective. In this study, we compare two prominent categories of algorithms: (a) deep-learning-based methods and (b) physics-based methods. We apply both approaches to a large set of monomeric natural proteins, redesigning their sequences to estimate the stability of the modified proteins. Our results demonstrate that, in terms of enhancing stability relative to the original native proteins, deep-learning-based methods outperform physics-based methods.