Enzyme catalytic residue prediction using deep learning methods

• Main Author: Guan, Jia Sheng
• Other Authors: Mu Yuguang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Science::Biological sciences
• Online Access: https://hdl.handle.net/10356/171862
• Institution: Nanyang Technological University

Identification of catalytic residues in enzymes have important applications ranging from drug discovery to protein engineering. However, locating catalytic residues in laboratory is time consuming and costly. Through high throughput computational methods, potential catalytic residues could be elucidated. While many models trained to predict catalytic residues were published, there are still unexplored combinations of model features and data preparation methods. In this project, graph neural network (GNN) and multi-layer perceptron (MLP) models were constructed to predict catalytic residues. The choice of edge weight equation was discovered to have huge impact on GNN model performance. Embeddings from a large protein language model, Evolutionary Scale Modeling 2 (ESM-2), were experimented and found suitable as features for MLP and GNN models, rivaling many published models in performance. Atchley factors as features were investigated but results hinted that the information might have already been included in the ESM-2 embeddings. To address knowledge gap, structural information of entire protein complex was considered as GNN model feature but found no benefits as compared to using only monomer structures as in published models. To resolve class imbalance issue, down-sampling of non-catalytic to catalytic residues to a 10:1 ratio was tested but it did not improve models’ performances.