Decoding allosteric communication via Riemannian manifold geometry

Allosteric regulation, in which binding at one protein site reshapes activity at a distant site, coordinates nearly every major cellular signal, yet predicting allosteric sites from structure alone remains unsolved. Here we show that Gauge Allo, a geometric deep learning framework that treats the protein conformational landscape as a Riemannian manifold and learns features independent of the arbitrary coordinate frame assigned to each residue, substantially advances allosteric site prediction. Three geometric principles drive the approach: gauge-equivariant convolution that transports features along low-energy conformational pathways; Ricci curvature-weighted attention that focuses on the flexible hinges through which allosteric signals propagate; and holonomy features that encode the integrated geometric phase around closed loops on the manifold, capturing long-range communication beyond the reach of local message passing. On four benchmarks, GaugeAllo achieves an area under the precision-recall curve of 0.91 ± 0.02 for allosteric site detection, top-10 recall of 0.84 ± 0.03 for cryptic pocket discovery, accuracy of 0.87 ± 0.02 for drug-resistance mutation classification, and a chemical shift error of 0.43 ± 0.04 ppm for NMR chemical shift prediction.