基于云端的配体引导虚拟筛选结合深度学习增强的对接技术，成功识别出一种微摩尔级别的CXCR4拮抗剂。

Abstract

We report a cloud-executed, ligand-guided virtual screening workflow for practical and rapid experimental testing of CXCR4 antagonists. Using the Rush platform, an Enamine purchasable library (274,092 compounds) was standardized and pre-filtered to remove assay-interfering substructures, yielding 264,953 molecules for screening. A set of 18,378 compounds was identified through AMD070-based similarity filtering (Tanimoto ≥ 0.2, ECFP4). These candidates were then docked into the IT1t-bound CXCR4 structure using Gnina. A CNN pose-score cutoff (> 0.8) was applied as a pose-confidence filter, and the retained compounds were ranked by docking affinity (kcal/mol). The top 50 were triaged by pharmacokinetic-focused in silico assessment (SwissADME) to prioritize lead-like profiles and exclude structural alerts, resulting in 9 purchasable candidates for biological testing. Flow cytometry-based 12G5 competitive binding in Jurkat cells identified CUEN-837 as the initial hit (IC 50  = 8.50 μM, 52.31% inhibition at 10 μM). CUEN-837 also inhibited CXCL12-driven chemotaxis (32.8% inhibition at 10 μM). It reduced the viability of three colorectal cancer organoid strains in a dose-dependent manner (IC 50 7.97–11.16 µM), providing supportive phenotypic evidence of activity in patient-derived organoid models. Overall, this end-to-end workflow efficiently compresses large libraries into a confirmed micromolar CXCR4 hit while lowering computational and infrastructure barriers for early-stage discovery.