Article Preview
TopIntroduction
In the last decades, machine learning (ML) techniques have been considered important tools in the drug design and discovery pipeline in integration with other computational approaches and synthetic methodologies (Schneider, 2019; Jiménez-Luna et al., 2020; Rohall et al., 2020; Mouchlis et al., 2021). Cheminformatic studies using deep and transfer learning are a recent trend in drug design to predict biological activities, pharmacokinetics profiles, and toxicity of potential lead compounds (Cai et al., 2020; Chen et al., 2018; Elton et al., 2019; Serafim et al., 2021). Graph neural networks (GNN), a type of deep learning technique, are reported as robust algorithms for drug design-related tasks (Duvenaud et al., 2015; Yang et al., 2019).
As an example of the applications of ML in drug design, Durrant and Amaro (2015) and Serafim et al. (2020) disclosed their use in the discovery of new antibacterial agents, with applications in structure-activity relationship (SAR) studies, virtual screenings, and genome sequencing (Durrant & Amaro, 2015; Serafim et al., 2020). We illustrate the use of GNN in drug design by Stokes and co-workers (2020) work. The authors applied a graph convolutional-based neural network (GCN) algorithm for drug repurposing in antibiotic therapy. They selected halicin, a very distinctive chemical structure compared to known antibacterials, where the compound showed potent broad-spectrum antibacterial activity both in in vitro and murine models against both standard and clinical isolated strains (Stokes et al., 2020).
With the emergence of antibiotic-resistant bacteria, novel targets are necessary to avoid resistance mechanisms from the strains (Blair et al., 2015). In this scenario, enoyl-[acyl-carrier-protein] reductase (FabI) is spotlighted as an emergent target to develop new antibacterial agents. This enzyme takes part in the fatty acid metabolism, where the inhibition of the catalysis leads to a disruption in the cell membrane. Since the fatty acid metabolism in mammals occurs in a quite different mechanism, FabI is an interesting target in antibacterial chemotherapy (Balemans et al., 2010; Janßen & Steinbüchel, 2014; H. Lu & Tonge, 2008).
Prior works explored ML techniques to predict FabI inhibitors with great robustness and accuracy, such as holographic structure–activity relationships (HQSAR), decision tree (DT), random forest (RF), multilayer perceptron (MLP) and support vector machine (SVM) (Kronenberger et al., 2017; Maltarollo, 2019). GCNs can extract features directly from the raw input data to generate graphs that represent the chemical structures. It is more advantageous when compared to machine learning algorithms applying traditional fingerprints. These fingerprint extraction methods can miss important information due to unknown bias, where the graph convolutions can encode hidden representations to provide richer information that can impact on the model predictiveness (Kojima et al., 2020; Sun et al., 2020). In this work, we applied a graph convolutional network algorithm to generate classification models for potential Staphylococcus aureus FabI (SaFabI) inhibitors.