Detection of new candidate compounds against four antibiotic targets using explainable artificial intelligence by molecular fingerprints

Abstract

Antibiotic resistance is a threat that renders bacteria ineffective against antibiotics and makes it difficult to treat infections. Therefore, finding new target compounds is essential in discovering and developing new antibiotics. In this study, we developed an artificial intelligence algorithm that can predict and explain the pIC50 values for four antibiotic targets (Penicillin Binding Proteins (PB), β-Lactamase (BL), DNA Gyrase (DG), and Dihydrofolate Reductase(DR)). The algorithm uses molecular fingerprints of the molecules to predict the pIC50 values using the random forest regression method. We created the algorithm in a transparent and interpretable way. We used permutation feature importance (PFI) and Shapley explanations methods to identify the different molecular fingerprints that have the most influence on the pIC50 values. The results obtained from these methods show that different molecular fingerprints are essential for different antibiotic targets. According to the permutation importance results, KRFPC1646 (number of hydrogen bond donors of the compound) for BL and DR targets; 579 (a substructure with 5 bonded radius around the atom) for DG target; SubFPC182 (number of aromatic rings in the molecule) for PB target, are important fingerprints. With explainable artificial intelligence (XAI) (SHAP), KRFPC1646 (the number of hydrogen bond donors of the compound) for BL; KRFPC4274 (C1CCCCC1) for DR; 401 (C1CCCCC1) for DG; SubFPC182 (number of aromatic rings in the molecule) were determined as important fingerprints for PB. These results demonstrate the effectiveness and potential of using molecular fingerprints with explainable artificial intelligence to find new antibiotic candidates.