Background Coronary wall shear stress (WSS) derived from computational fluid dynamics (CFD) provides mechanistic insight and prognostic information, but its clinical translation is hindered by modeling complexity and computation time. We evaluated a geometric deep learning framework based on gauge-equivariant mesh graph convolutional network (GEM-GCN) to estimate coronary WSS directly in geometries reconstructed from coronary angiography in real-world patients. Methods A total of 1078 coronary arteries from 748 patients were reconstructed from invasive angiography. Time-averaged WSS computed by transient CFD served as reference labels for GEM-GCN training and testing. Two experiments were conducted: (i) random splitting of the full dataset with 10-fold cross-validation, and (ii) a clinical split , to assess whether GEM-GCN–derived WSS preserved the ability to predict myocardial infarction (MI) compared with CFD-derived WSS. Results GEM-GCN produced patient-specific WSS maps in < 5s per vessel. GEM-GCN slightly underestimated lesion- and vessel-averaged WSS in the random split, with absolute and percentage errors equal to 0.48 [0.26–0.78] Pa and 23.6 [14.8–42.6]%, respectively. High spatial agreement was found for high-WSS regions (Dice distance 0.88 [0.81–0.92]). Similar performance was observed in the clinical split (absolute error 0.65 [0.41–1.12] Pa; Dice distance 0.84 [0.71–0.90]). After normalization by vessel-averaged WSS, the correlation between GEM-GCN-derived and CFD lesion-averaged WSS improved from R = 0.67 to R = 0.89 (p < 0.0001). Lesion-averaged WSS and the lesion-to-vessel WSS ratio achieved comparable MI prediction performance for CFD and GEM-GCN. Conclusions Geometric deep learning enables fast, CFD-free coronary WSS estimation from routine angiography, supporting its potential for large-scale, real-world risk stratification.