Abstract
Dual-specificity tyrosine-regulated kinase 1A (DYRK1A) is a promising therapeutic target for pancreatic ductal adenocarcinoma (PDAC). Herein, we developed an integrated AI and structure-based pipeline featuring a Serial PNA-Transformer graph neural network, which achieved a test AUC of 0.8901. Multistage screening of 21,738 compounds prioritized 232 candidates across 10 chemical clusters. Enzymatic assays confirmed three hits with IC(50) values <500 nM; notably, CX-6258 (IC(50) = 473.7 nM) exhibited potent antiproliferative activity in MIA PaCa-2 and Panc-1 cell lines with low micromolar potencies (IC(50) = 0.679 and 1.148 μM, respectively). Selectivity profiling confirmed the potency of CX-6258 against DYRK1A/B with a favorable window over other CMGC kinases. Crucially, siRNA-mediated knockdown and overexpression assays demonstrated that its cytotoxicity is strictly DYRK1A-dependent. Molecular dynamics revealed a stable binding mode characterized by a unique Arg250-mediated electrostatic driving force. These findings underscore the utility of our AI-driven framework in accelerating the identification and mechanistic validation of potent therapeutic leads.