Computational Modeling of Protein Stability: Quantitative Analysis Reveals Solutions to Pervasive Problems

Abstract

Highlights d Experimentally, mutations predicted to stabilize are near neutral on average d Stability predictors favor mutations that increase stability but decrease solubility d Predictor performance is quantified well by the Matthews correlation coefficient d Multi-mutants reach stability targets with higher probability than single mutants

Predicting protein stability and solubility changes upon mutations: data perspective

“…Indeed, many predictors were reported to demonstrate a similar bias: mutations are usually correctly predicted as destabilizing, but those predicted stabilizing on average turn out to be neutral during experimental validation. [28] Apart from ProTherm data, some predictors were tested on 42 mutations of the DNA binding domain of the tumor suppressor protein p53, [60] and the performance of several predictors was recently evaluated on two newly collected datasets: 96 single-point mutants of guanylate kinase [61] and 51 mutants of β-glucosidase. [33] Several teams performed an additional independent literature search, revealing the promising prospects of seeing improved protein stability predictors in the near future.…”

“…Many promising ML-based predictors have been published for either task. [1,[28][29][30] However, they all seem to be testing a similar limit to the prediction accuracy, e. g. the root mean square error of around 1 kcal/mol for stability predictions. [1] Moreover, independent experimental validation in subsequent studies often reveals modest performance.…”

“…On the other hand, protein stabilization achieved by means of protein engineering was often reported to come at the cost of reducing protein solubility. [17,28] For instance, stabilization strategies frequently suggest surface mutations that increase hydrophobicity, and while such muta-tions do often increase stability, [35] they tend to have a detrimental effect on solubility. The flexibilities of side chains and whole protein regions have been reported to guide the engineering of both stability [36] and solubility.…”

“…However, they all seem to be testing a similar limit to the prediction accuracy, e. g. the root mean square error of around 1 kcal/mol for stability predictions [1] . Moreover, independent experimental validation in subsequent studies often reveals modest performance [28,31–33] . Several explanations can be provided, one of which is the limited data size and quality available for training: data quality and abundance are critical for ML algorithms as they ultimately aim to identify and generalize patterns in the training data.…”

“…On the one hand, unstable proteins tend to aggregate and are prone to faster degradation by proteolysis, [34] producing a negative signal in solubility assays. On the other hand, protein stabilization achieved by means of protein engineering was often reported to come at the cost of reducing protein solubility [17,28] . For instance, stabilization strategies frequently suggest surface mutations that increase hydrophobicity, and while such mutations do often increase stability, [35] they tend to have a detrimental effect on solubility.…”