AI-assisted discovery of potent FGFR1 inhibitors via virtual screening and in silico analysis.

Fibroblast growth factor receptor 1 (FGFR1) is recognized as an oncogene that fosters tumor development, playing a vital role in cancer progression. This has established it as a promising target for cancer drug development. However, existing FGFR1 inhibitors are often limited by drug resistance and lack of specificity, emphasizing the need for more selective and potent alternatives. To address this challenge, the present study employed an AI-driven virtual screening approach, integrating molecular docking (MD) and molecular dynamics simulations (MDS) to discover novel FGFR1 inhibitors. A voting classifier integrating three machine learning classifiers was utilized to screen 10 million compounds from the eMolecules database, leading to 44 promising candidates with a prediction probability exceeding 80%. MD identified compound with PubChem Compound Identifier (CID) 165426608 (-10.8 kcal/mol) as the highest-scoring ligand, while compounds with CID 145940129 (-9.8 kcal/mol), CID 131910163 (-9.4 kcal/mol), CID 155915988 (-9.2 kcal/mol), and CID 132423733 (-9.1 kcal/mol), exhibited binding affinities comparable to or slightly lower than that of the native ligand (-10.4 kcal/mol). MDS further revealed that all these compounds, except CID 131910163, maintained structural stability with time. Thermodynamic stability assessment confirmed the spontaneity and feasibility of their complex formation reactions with negative ΔGBFE values ranging from -21.87 to -12.76 kcal/mol. Decomposition of binding free energy change further provided key stabilizing residues. The heatmaps and histograms of the interaction over the full 200 ns simulation period highlighted the prominent interaction profiles. Structural similarity analysis of the four MDS-stable compounds displayed the dice similarity scores of 0.200000 to 0.452830 with known FGFR1 inhibitors. Additionally, the pIC50 prediction using a voting regressor indicated promising pIC50 values (7.07 to 7.47), highlighting their potential as hit candidates for further structural optimization and therapeutic development. Further, this study underscores the efficiency of machine learning-based virtual screening and in silico analysis as a cost-effective and reliable strategy for accelerating hit drug discovery from large datasets, even with limited resources and time.