A class A-trained GPCR activation-state classifier transfers to class B1, with attention mass in GLP-1R, GCGR and GIPR spiking at arginine 6.37x37, the TM6 member of the cytoplasmic polar network
- 1Deep EigenMatics, Inc.
Abstract
Receptor activation state is the structural variable that most GPCR drug discovery is implicitly trying to control. Using class A GPCR structures only, we trained the Receptor Class-Activation-Map Localizer (RCL), a classifier that distinguishes active from inactive GPCR structures, and evaluated it on class B1, a family it never saw. Pooled class B1 performance was AUROC 0.9532 (95% CI 0.9201 to 0.9863, n = 146). On GLP-1R, held out entirely, it reached AUROC 1.0000 across 43 structures (37 active, 6 inactive). Attention in GLP-1R concentrates at the cytoplasmic end of TM6, a region also named in the pre-registered specification. Against a permutation null, the intracellular third of that helix is enriched 6.78 to 11.30-fold, against 2.93 to 4.01-fold for the helix taken in full. Within that third a single position dominates. R348, generic number 6.37x37, carries the attention maximum in 42 of 43 GLP-1R structures, active and inactive alike and apo structures included, holding a median of 53.58 times the attention mass a uniform distribution would place on a single residue. The same position carries the maximum in 17 of 18 GCGR and 6 of 9 GIPR structures, where it is R346 and R338 respectively; it is an arginine in all three receptors. That position was independently mutated in GLP-1R and reported to modulate receptor activation a decade before this model existed. Before the model was trained we specified and committed a single pass criterion, that attention mass would concentrate in the cytoplasmic third. It was met: enrichment 2.07-fold at 6X18 and 1.94-fold at 6XOX, p < 0.0001 in both. This is a result about activation state rather than about pathway preference; no structure in training carried a pathway label. It recovers known activation biology at the level of a single residue: the position carrying the maximum is the TM6 member of the cytoplasmic polar network, a feature independently described as holding the receptor base closed and as breaking when the receptor couples to a G protein. That a model trained without positional supervision places its attention maximum on a characterized activation residue, in a receptor class it never saw, is what makes these maps a usable input into structure-aware biased-ligand discovery.
Related
- A thioester-thiol exchange reaction in Salmonella Invasion acyl carrier protein transfers the 4PP-bound acyl chain to a distal cysteine
- GW182 Proteins Directly Recruit Cytoplasmic Deadenylase Complexes to miRNA Targets
- Preface to the special issue “Cytoplasmic Genome Regulation: Genomes and Regulatory Mechanisms of Chloroplasts and Mitochondria”
- Cytoplasmic Lipid Membrane Biogenesis in Prokaryotes: From Soluble Lipid Building Blocks to a Dynamic and Functional Boundary Layer
- Cytoplasmic recapping commences the translation of a micropeptide, MARS from MALAT1 after the removal of an inhibitory RNA G-quadruplex
BibTeX
@misc{odaibo2026class,
title = {{A class A-trained GPCR activation-state classifier transfers to class B1, with attention mass in GLP-1R, GCGR and GIPR spiking at arginine 6.37x37, the TM6 member of the cytoplasmic polar network}},
author = {Stephen G. Odaibo},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010433/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010433/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - A class A-trained GPCR activation-state classifier transfers to class B1, with attention mass in GLP-1R, GCGR and GIPR spiking at arginine 6.37x37, the TM6 member of the cytoplasmic polar network
AU - Stephen G. Odaibo
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010433/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010433/v1
AB - Receptor activation state is the structural variable that most GPCR drug discovery is implicitly trying to control. Using class A GPCR structures only, we trained the Receptor Class-Activation-Map Localizer (RCL), a classifier that distinguishes active from inactive GPCR structures, and evaluated it on class B1, a family it never saw. Pooled class B1 performance was AUROC 0.9532 (95% CI 0.9201 to 0.9863, n = 146). On GLP-1R, held out entirely, it reached AUROC 1.0000 across 43 structures (37 active, 6 inactive). Attention in GLP-1R concentrates at the cytoplasmic end of TM6, a region also named in the pre-registered specification. Against a permutation null, the intracellular third of that helix is enriched 6.78 to 11.30-fold, against 2.93 to 4.01-fold for the helix taken in full. Within that third a single position dominates. R348, generic number 6.37x37, carries the attention maximum in 42 of 43 GLP-1R structures, active and inactive alike and apo structures included, holding a median of 53.58 times the attention mass a uniform distribution would place on a single residue. The same position carries the maximum in 17 of 18 GCGR and 6 of 9 GIPR structures, where it is R346 and R338 respectively; it is an arginine in all three receptors. That position was independently mutated in GLP-1R and reported to modulate receptor activation a decade before this model existed. Before the model was trained we specified and committed a single pass criterion, that attention mass would concentrate in the cytoplasmic third. It was met: enrichment 2.07-fold at 6X18 and 1.94-fold at 6XOX, p < 0.0001 in both. This is a result about activation state rather than about pathway preference; no structure in training carried a pathway label. It recovers known activation biology at the level of a single residue: the position carrying the maximum is the TM6 member of the cytoplasmic polar network, a feature independently described as holding the receptor base closed and as breaking when the receptor couples to a G protein. That a model trained without positional supervision places its attention maximum on a characterized activation residue, in a receptor class it never saw, is what makes these maps a usable input into structure-aware biased-ligand discovery.
ER -