Identification of FLT3-ITD inhibitors through AI-guided virtual screening, biochemical validation, DFT analysis, and molecular dynamics simulation.
TL;DR
An AI-assisted discovery workflow integrating graph neural network (GNN) modeling, virtual screening, biochemical validation, and mechanistic analysis to identify novel FLT3-ITD inhibitors demonstrates that combining AI-based prioritization with structure-based screening enables efficient discovery of structurally diverse FLT3-ITD inhibitors.
Abstract
FLT3-ITD is a major therapeutic target in acute myeloid leukemia (AML). Herein, we developed an AI-assisted discovery workflow integrating graph neural network (GNN) modeling, virtual screening, biochemical validation, and mechanistic analysis to identify novel FLT3-ITD inhibitors. A curated dataset of 3779 compounds from ChEMBL was used to construct three GNN models, with the message-passing neural network (MPNN) showing optimal performance (accuracy = 0.89, ROC AUC = 0.90). Guided by MPNN, 5297 compounds retrieved from the Topscience database were prioritized and filtered via docking and MM/GBSA calculations, yielding 12 candidates (F1-F12) for experimental testing. Biochemical assays identified three active hits, F2 (extended multi-aryl amide/urea scaffold, IC50 = 7.68 ± 0.63 μM), F4 (compact heteroaromatic scaffold, IC50 = 0.29 ± 0.13 μM), and F7 (heteroaryl amide scaffold with polar substituent, IC50 = 2.15 ± 0.28 μM). Mechanistic analyses revealed that F4 exhibits the most favorable profile, with a low HOMO-LUMO energy gap (3.27 eV), stable dynamic behavior, and a dense hydrogen-bond interaction. These results demonstrate that combining AI-based prioritization with structure-based screening enables efficient discovery of structurally diverse FLT3-ITD inhibitors, with F4 representing a promising hit-to‑lead candidate for further optimization. These findings provide both new candidates and a feasible computational-experimental paradigm for the future development of FLT3-targeted therapeutics.