Deep learning-assisted discovery of a potent and cell-active inhibitor of RNA N⁶-methyladenosine recognition protein YTHDC2.

Abstract

YTHDC2, a unique YTH-domain-containing protein that recognizes N6-methyladenosine (m⁶A) on RNA, plays critical roles in diverse pathological processes and represents a promising therapeutic target. Despite its potential, no potent small-molecule inhibitors have been reported to date. To bridge this gap, we develop EPMolGen, a deep learning-based molecular generative model that explicitly incorporates the electrostatic features of receptor proteins. The model achieves state-of-the-art performance in dry-lab validations. Using EPMolGen, we identify H3, a YTHDC2 inhibitor with an IC₅₀ of 16.84 μM. Subsequent structural optimization of H3 yields DC2-C1, a highly potent compound with an IC₅₀ of 0.168 μM against YTHDC2 and selectivity over other YTH-domain proteins. In cellular assays, DC2-C1 effectively targets YTHDC2. Notably, DC2-C1 treatment substantially reduces the expression levels of multiple target mRNAs of YTHDC2, leading to phenotypic suppression of related cells. Overall, this study highlights the great potential of deep learning in drug discovery and provides a promising lead compound for drug development targeting YTHDC2.