Utilizing Deep Learning to Predict the Potency of Beta-Lactamase Inhibitors

Drug-resistant bacteria pose a significant global health threat, driving the need for innovative antibiotic development. The efficacy of these antibiotics is evaluated through biological potency assays that measure their ability to elicit targeted responses. Beta-lactam bacteria produce beta-lactamases, enzymes that hydrolyze the beta-lactam ring, rendering the antibiotics ineffective. To counteract this mechanism, beta-lactamase inhibitors play a pivotal role by preventing the enzymatic degradation of beta-lactam antibiotics. This study focuses on predicting the chemical compositions and molecular motifs that characterize active beta-lactamase inhibitors. Using k-means clustering, active small-molecule beta-lactamase inhibitors were categorized based on their unique molecular structures. A graph-based modeling approach was employed to represent these molecular structures, then leveraging Graph Attention Networks (GAT) to identify and predict substructural features associated with each distinct cluster. Molecular graph representations served as inputs for the GAT model, enabling precise classification of compounds into distinct clusters. The GAT model's performance in multiclass classification was benchmarked against traditional approaches, demonstrating superior accuracy in identifying key substructures that differentiate active beta-lactamase inhibitors. Additionally, the attention mechanism within the GAT model facilitated the identification of specific molecular motifs by focusing on relevant structural features during the learning process. The findings highlight the effectiveness of the graph-based approach in advancing the understanding and prediction of active betalactamase inhibitors, with implications for drug discovery and combating antibiotic resistance.