A Dynamics-Informed Approach to Prediction of Protein Melting Temperature via Graph Neural Networks

A practical design of biomedical materials requires knowledge of their thermal properties. Melting temperature has a direct influence on molecular stability, functionality, and performance in the case of protein engineering. For example, biosensors and enzymes will only operate as intended within a limited thermal range. As such, the prediction of the thermostability of proteins is a crucial factor with implications in various scientific disciplines and technical applications. This task requires an understanding of the hierarchical organization of proteins, which manifests in a dynamically coupled cellular system. The introduction of machine learning methods to biology brought about astounding breakthroughs, such as the prediction of protein structures by AlphaFold. It is a promising approach to undertaking convoluted relationships in protein science. In this study, we propose a machine learning method for the computation of the melting temperature of proteins, taking into account the amino acid sequence, protein structure, and dynamics. Graphs (as in graph theory) are chosen as our data representation due to their rotational invariance and their intuitive mapping to molecular structures. The sequential, structural, and dynamical information are represented as multigraphs, with residues as its nodes, and various node features and edge connections derived from each of these properties. To process the data, a graph neural network architecture that makes use of message passing layers was designed to accommodate the multiple types of connections. Protein structures were computed by AlphaFold and the dynamics were computed based on the torsional network model (TNM) for training. Hence, the learned features and parameters can be readily applied to protein sequences without known experimental structure, satisfying the goal of aiding the prediction of design proteins. We are also able to identify key domains that contribute to the thermal stability/instability of proteins by computing the graph regression activation map, which is based on the partial derivative of the predicted value with respect to features on the input nodes. Our method gives insight into the mechanism of natural proteins and provides critical information regarding design proteins.