Research

Model Performance Series

June 2026

Astra AI on GPCRs

GPCRs are the most-drugged protein family in medicine — and among the hardest to characterize computationally. This report shows how the Astra suite reads them from sequence alone: function, topology, PTM sites, binding pockets, and thermostability.

Çağlar Bozkurt, Aniruddh Goteti · Orbion GmbH · Benchmarked on 2,615 GPCRs

0.97
AUROC · Transmembrane Topology
64%
Ligand Recall on Co-Crystal Pockets
82%
Directional Accuracy on ΔTm

What We Found

Three Results That Matter

Each section of the whitepaper reports the headline performance for one prediction area, its known weaknesses, and where we recommend it for production use.

01

0.97 AUROC on Transmembrane Topology

Per-residue predictions agree with UniProt annotations across the GPCR subfamilies, from the class-A rhodopsin-like receptors through the adhesion and frizzled classes.

02

F1 Up to 0.94 on PTM Sites

Modification sites are flagged across 39 classes, with two operating points for precision or recall. Strongest on disulfide bonds and N-linked glycosylation.

03

82% Directional Accuracy on Mutations

On thermostabilizing mutations the model calls the direction of effect right most of the time — enough to filter destabilizers before they reach the bench.

Every number above is reproduced in the whitepaper against public references — Swiss-Prot annotation, PDB co-crystal contacts, and curated experimental thermal-shift data. We report where the models are weak as plainly as where they are strong.

Read the Whitepaper