June 2026
Model Performance Series
Astra AI on GPCRs
GPCRs are the largest drug-target family — and the class where computational characterization fails most often: buried pockets, conformationally plastic bundles, thermostability that takes a detergent screen just to measure. This report shows how Orbion's Astra AI Suite characterizes them from sequence alone — function, topology, PTM sites, binding pockets, and thermostability — measured against public experimental ground truth across 2,615 receptors.
0.97 AUROC — Transmembrane Topology vs. UniProt
64% Ligand Recall on Co-crystal Pockets
82% Directional Accuracy — ΔTm
0.97 AUROC on Transmembrane Topology
Per-residue transmembrane prediction agrees with UniProt-annotated segments at AUROC 0.97, uniformly across every GPCR sub-family — from opsins to adhesion receptors. The model resolves the membrane-spanning core that defines the class without a structure, giving construct designers a reliable topology map before the first experiment.
F1 Up to 0.94 on PTM Predictions
Across all 39 PTM classes, the suite flags per-residue modification sites at two operating points — high-precision for confident wet-lab handoff, high-recall for hypothesis generation. Strongest on the disulfide bonds and N-linked glycosylation that govern receptor folding and surface delivery (F1 up to 0.94), with the regulatory phosphorylation map alongside.
82% Directional Accuracy on Mutations
Thermostabilizing a GPCR for structure is the field's classic bottleneck. The suite predicts the melting-temperature shift of each point mutation and ranks a candidate panel — with 82% directional accuracy on the high-impact mutations. Predictions inside the assay's own ±2 °C noise band (shaded in the plot) aren't actionable, and we show that plainly; the value is filtering the strong destabilizers before a thermal-shift screen.
Have a question?
Whether you want to understand our models, test a specific protein, or just ask how Orbion applies to your research — send us your question and we'll get back to you. Or, if you'd prefer a live walkthrough, book a demo.
