Abstract
<title>Abstract</title> <p>Colorectal cancer (CRC) remains a major cause of cancer mortality worldwide, underscoring the need for selective cell-cycle-targeted therapies. Cyclin-dependent kinase 2 (CDK2), a central regulator of the G1/S transition, represents an attractive yet challenging therapeutic target due to high structural homology across the CDK family. Here, we present an integrated generative artificial intelligence (AI)–to–experiment framework that enables de novo discovery of a CDK2-biased inhibitor by learning molecular conformational energy landscapes rather than relying on static structure-based screening alone. Starting from a known CDK2 active scaffold, a hybrid variational autoencoder-generative adversarial network-reinforcement learning (VAE–GAN–RL) architecture generated over 2.3 million chemically diverse molecules, which were dynamically triaged using deep learning-based docking and molecular dynamics simulations. This approach identified CDKCRC-EMBS as a lead compound exhibiting enhanced dynamic stabilization within the CDK2 ATP-binding pocket. CDKCRC-EMBS was synthesized via a concise five-step route and structurally validated by NMR, high-resolution mass spectrometry, and HPLC. Biochemical assays demonstrated potent inhibition of CDK2/cyclin E (IC₅₀ = 1.92 ± 0.12 µM; Kᴅ = 1.35 µM) accompanied by protein thermal stabilization. In colorectal cancer cell models, CDKCRC-EMBS induced cell-cycle arrest, suppressed Rb phosphorylation, and promoted apoptosis, while sparing normal colonic epithelial cells. Collectively, this study establishes a scalable AI-driven discovery paradigm that integrates conformational learning with experimental validation and identifies CDKCRC-EMBS as a promising, CDK2-biased inhibitor for colorectal cancer therapy. The novel compound CDKCRC-EMBS, identified in this research, is listed in the NCBI PubChem database under CID 9829487 and SID 522761491.</p>