Abstract
Phosphatidylinositol 3-kinases (PI3Ks) play a crucial role in human metabolism, and their dysregulation contributes to the development of several metabolic disorders, including cancer. Despite advances in experimental high-throughput screening, discovering new therapeutic agents remains challenging and costly. In this study, we developed PI3K-Seeker, a web server based on a two-stage prediction process to find new PI3K inhibitors. The first stage eliminates nonbinders, while the second refines the selection, leaving only molecules with a high probability of being potent inhibitors. Models were trained using the XGBoost algorithm and PubChem fingerprints extracted from distinct datasets. In the first stage of classification, the model showed impressive metrics (MCC: 0.917, AUC-ROC: 0.993, and ACC: 0.917). In the second stage, the data enhancement, the model trained also performed exceptionally well (MCC: 0.939, AUC-ROC: 0.956, and ACC: 0.994). The PI3K-Seeker is a user-friendly web server suitable for a large set of compounds, available at http://www.ufrgs.br/labec/pi3k-seeker/.