Predicting the molecular structure relationship and the biological activity of DPP-4 inhibitor using deep neural network with catboost method as feature selection

Dipeptidyl peptidase 4 (DPP-4) are drug targets for type-2 diabetes mellitus (T2DM). The enzyme dipeptidyl peptidase 4 (DPP-4) can catalyze the decrease in the hormone incretin peptide, especially peptide-1, such as glucagon-like peptide-1 (GLP-1) and the hormone gastric inhibitory peptide (GIP), which results in decreased insulin synthesis. Inhibitors of DPP-4 are promising drug targets for T2DM because they are able to block the work of the DPP-4 enzyme by inhibiting the action of the hormones GLP-1 and GIP. Unfortunately, DPP-4 inhibitors have some adverse effects, such as nausea, headache, nasopharyngitis, and skin reactions. So, the medical field are still expecting new DPP-4 inhibitors with minimal effects. In this study, there are 1773 structures of DPP-4 inhibitors with 1185 active compounds and 588 inactive compounds extracted using topological fingerprints as descriptors. As there is a class imbalance in the dataset, there needs to be an oversampling technique applied, we have decided to use the SMOTE technique. The deep neural network (DNN) method is proposed as a method of classifying DPP-4 inhibitors and is optimized using Adam's optimizer and dropout regularization technique. In addition, we introduce CatBoost as a feature selection method. As a result, the DNN method combined with ECFP-6 and using feature selection with the proportion of the importance value of the feature at 90% produces the highest MCC value that is 0.810, and the sensitivity, specificity, and accuracy values being 0.927, 0.881, and 0.906, respectively.