Abstract
Proteins play a critical role in biology and biopharma due to their specificity and minimal side effects. Predicting the effects of mutations on protein stability is vital but experimentally challenging. Deep learning offers an efficient solution to this problem. In the present work, we introduced ProstaNet, a deep learning framework that predicts stability changes resulting from single- and multiple-point mutations using geometric vector perceptrons-graph neural network for 3-dimensional feature processing. For training ProstaNet, we meticulously crafted ProstaDB, a comprehensive and pristine thermodynamics repository, including 3,784 single-point mutations and 1,642 multiple-point mutations. We also created thermodynamic looping for enlarging the limited data size of multiple-point mutation and applied an innovative clustering method to generate a standard testing set of multiple-point mutation. Besides, we identified residue scoring as the most important encoding method in protein properties prediction. With these innovations, ProstaNet accurately predicts thermostability changes for both single-point and multiple-point mutations without showing any bias. ProstaNet achieves an accuracy of 0.75, outperforming existing methods for single-point mutation prediction, including ThermoMPNN (0.63), PoPMuSiC(sym) (0.66), MUPRO (0.52), and FoldX (0.71). ProstaNet also achieves a 1.3-fold increase in accuracy compared to FoldX for multiple-point mutation predictions. Validated by experiment, 4 out of 5 single-point mutation predictions (80%) and all multiple-point mutation predictions (100%) for HuJ3 mutants were accurate, demonstrating the potential benefits of ProstaNet for protein engineering and drug development.