GENERATING SIGNALING PATHWAYS MAP FOR CARDIAC FIBROSIS FROM LITERATURE USING DEEP LEARNING

Abstract

Most diseases involve entire networks of interacting proteins rather than single proteins. Diseases typically happen because of mutations in protein sequences, or if a specific protein amount is too high or too low in a cell, or because of an issue in the interaction between the proteins in the cell if they are overacting or don’t interact at all while they should; understanding which protein-protein interactions (PPIs) are broken is therefore critical for understanding disease mechanisms and enabling targeted medicine creation. However, identifying PPI interactions using the traditional experimental methods in the labs remains computationally challenging, expensive and time-consuming. This research uses deep learning techniques that is faster and less cost to identify the PPI, it proposes a GNN architecture that uses 3D protein structures predicted using AlphaFold2 to improve PPI predictions by using pre-trained protein embeddings from D-SCRIPT. Graphs of protein interactions were built in which nodes correspond to residues, where each presented by a D-SCRIPT embedding, and edges denote proximity based on predicted intra-protein and inter-protein contacts. A learnable linear projection layer reduces dimensionality to 256, and then a 3-layer Graph Convolutional Network (GCN) encoder with residual connections and Layer normalization. A multi-head self-attention pooling mechanism generates representations at the protein level, which then feed into a 3-layer MLP classifier to make predictions. The model was trained and tested on the human PPI dataset from the STRING database, which consisted of about 150,000 protein pairs across ~15,000 unique proteins. The model implemented with a strict data leakage avoidance yielded a 96.22% accuracy, 98.81% precision, 95.37% recall, and 98.80% ROC-AUC on the held-out test set of 15,000 pairs. An ablation study was performed; comparing how well the model performed when using real AlphaFold contacts versus random synthetic contacts, leading to signify the quantification of 3D structural information. The random contacts baseline recorded an accuracy rate of 93.4%. The use of real structural contacts, therefore, enhanced the model’s accuracy rate to 96.15%, thereby validating our contention that real structural contacts do provide critical information to improve PPI prediction. On the other hand, the model performs significantly better than a purely D-SCRIPT-based model (72.33% accuracy).