人工智能与虚拟筛选识别抑制转谷氨酰胺酶2的小分子用于乳腺癌研究与治疗

Artificial intelligence-assisted virtual screening enables identification of chemically diverse bioactive small molecules. We targeted transglutaminase 2 (TG2), a multifunctional protein associated with mesenchymal-like features in breast cancer. Approximately 10 million compounds were screened against the catalytic site of open TG2 using the AtomNet structure-based deep-learning model. The top 30,000 candidates underwent physicochemical and structural filtering, diversity clustering, and manual selection, yielding 84 compounds for experimental evaluation. Fourteen induced apoptosis in MCF-7, eight retained activity in MDA-MB-436, and three in MDA-MB-231 cells, also reducing transamidase activity in cellular lysates. Chemical-space and fingerprint analyses placed C10, C12, and G11 outside the principal clusters of known active TG2 inhibitors, identifying structurally differentiated, nonpeptidic scaffolds. Docking into open and intermediate TG2 conformations consistently ranked G11 as the most favorable candidate, with predicted binding near the catalytic core. In a 5-(biotinamido)pentylamine (5-BAP-based assay, G11 inhibited purified recombinant human TG2 with an IC50 of 7.64 ± 2.87 μM, while an independent dimethyl casein-dansyl cadaverine assay confirmed concentration-dependent inhibition and yielded an estimated Ki of 44 ± 11 μM, assuming competitive inhibition with respect to the substrate. Finally, G11 did not significantly affect the viability of HEK293 cells, which express negligible or very low levels of TG2.