Modelling Covalent Inhibitors of Protein Kinases Using Deep Learning Techniques

Résumé fourni par la source

Protein kinases play an important role in cellular activities like cell division, growth, metabolism, etc, implicating diseases, particularly cancer. Covalent inhibitors are promising inhibitors due to their higher specificity and permanent binding with the protein. Predicting a covalent inhibitor with specific kinase proteins is a challenge. This paper compares different methods for classifying covalent and non-covalent inhibitors of P0DTD1 (replicase polyprotein 1ab), including traditional fingerprint-based models, Graph Neural Networks (GNNs), and Large Language Models (LLMs). Various methods to handle the class imbalance of 365 molecules in covalent and 5892 molecules in non-covalent class, with upsampling, class weights, and transfer learning were experimented. The results indicate that LLM with upsampling and transfer learning has the highest performance (F1 scores of 0.95+), but GNN architectures with relevant class balancing techniques showed concrete results despite dataset imbalance. The experiments provides valuable insight for development of computational methods in drug discovery, particularly in identifying potential covalent inhibitors.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.