Screening of Antibacterial Compounds With Novel Structure From The FDA Approved Drugs Using Machine Learning Methods

Abstract

Background: Due to the lack of new antibiotics in recent years, bacterial resistance has increasingly become a serious problem globally. The aim of this study is to construct an antibacterial compound predictor using machine learning methods to screen potential antibacterial drugs.Methods: Active and inactive antibacterial compounds were acquired from the ChEMBL and PubChem database, which were used to construct benchmark datasets. The antibacterial compound predictor is constructed using the support vector machine (SVM), random forest (RF), and multi-layer perception (MLP) methods. We predicted the antibacterial activity of the Food and Drug Administration (FDA) approved drugs in the DrugBank database and screened novel antibacterial drugs through structural similarity analysis.Results: In the initial screen process, the results suggested that the benchmark dataset based on FP2 molecular fingerprints, along with the SVM, RF, and MLP methods showed excellent prediction performance (mean AUC > 0.9 for all models). Using the combination of these three models, a total of 957 potential antibacterial drugs were predicted. Most of the predicted drugs showed low structural similarities compared with the FDA approved antibacterial drugs. We finally screened 9 predicted antibacterial drugs with novel structures including 2 anti-tumor drugs (cyclophosphamide and ifosfamide), 2 ophthalmic drugs (apraclonidine and echothiophate) and 5 anesthetics (desflurane, enflurane, isoflurane, methoxyflurane, and sevoflurane).Conclusions: This study provides a new insight for predicting antibacterial compounds with novel structures by using FDA approved drugs. The predicted compounds with novel structures are worthy of further experimental verification in the future.

Machine Learning in Antibacterial Drug Design

“…The authors noted that the deep-learning method was not very powerful because of low data availability. Li et al approached in a more general manner using more data points from the ChEMBL database ( Li et al, 2021 ). The group collected a library of 2708 active antibacterial compounds (IC 50 cut-off of 10 μM) and 78,620 inactive compounds and proceeded to calculate fingerprints (FP2, FP3, FP4, DLFP, MACCS, ECFP2, ECFP4, ECFP6, FCFP2, FCFP4, and FCFP6) and vector representations (mol2vec, SMILES2Vec, FP2VEC software; Jaeger et al, 2018 ; Öztürk et al, 2018 ; Jeon and Kim, 2019 ).…”

“…( Romero-Molina et al, 2019 ). Focusing on antibacterial drug design, Li et al reported SVM model development from the fingerprint-featurized ChEMBL database in order to identify novel antibacterial compounds ( Li et al, 2021 ). SVM model applications in antibacterial design and antibacterial drug resistance research were reviewed by Serafim et al (2020 ).…”

“…Upon reviewing literature on novel antibacterial design supported by machine learning, RF models were found to be one of the most commonly applied for classification, regression, and other tasks and represent a performance and computationally lean approach. In this review, a number of RF applications are presented, for small molecules, peptides ( Bhadra et al, 2018 ), natural product–based antibacterial design, and studying antibacterial drug resistance ( Dias et al, 2019 ; Maltarollo et al, 2019 ; Shi et al, 2020 ; Li et al, 2021 ; Wani et al, 2021 ). A good example of underlying supervised learning DT method was reported by Suay-Garcia et al (2020) .…”