Enhancing protein aggregation prediction: a unified analysis leveraging graph convolutional networks and active learning

Abstract

Protein aggregation (PA) is a critical phenomenon associated with Alzheimer's and Parkinson's disease. Recent studies have suggested that factors like aggregation-prone regions (APRs) and β-strand interactions are crucial in understanding such behavior. While experimental methods have provided valuable insights, there has been a shift towards computational strategies, particularly machine learning, for their efficacy and speed. The challenge, however, lies in effectively incorporating structural information into these models. This study constructs a Graph Convolutional Network (GCN) to predict PA scores with the expanded and refined Protein Data Bank (PDB) and AlphaFold2.0 dataset. We employed AGGRESCAN3D 2.0 to calculate PA propensity and to enhance the dataset, we systematically separated multi polypeptide chains within PDB data into single polypeptide chains, removing redundancy. This effort resulted in a dataset comprising 302 032 unique PDB entries. Subsequently, we compared sequence similarity and obtained 22 774 Homo sapiens data from AlphaFold2.0. Using this expanded and refined dataset, the trained GCN model for PA prediction achieves a remarkable coefficient of determination (R²) score of 0.9849 and a low mean absolute error (MAE) of 0.0381. Furthermore, the efficacy of the active learning process was demonstrated through its rapid identification of proteins with high PA propensity. Consequently, the active learning approach achieved an MAE of 0.0291 in expected improvement, surpassing other methods. It identified 99% of the target proteins by exploring merely 29% of the entire search space. This improved GCN model demonstrates promise in selecting proteins susceptible to PA, advancing protein science. This work contributes to the development of efficient computational tools for PA prediction, with potential applications in disease diagnosis and therapy.