What does the GPCR Foundation Model predict?

It does not predict an affinity. It answers a two-class comparison. The potency model takes one target and two ligands and returns which ligand is more potent. The selectivity model takes one ligand and two targets and returns at which target it is more potent. Each answer carries a prediction strength from 0.5 to 1.0.

Answering every comparison, potency is 0.77 accurate over 762,493 held-out comparisons across 235 targets, and selectivity is 0.80 over 24,741 comparisons across 211 targets. Accuracy rises with prediction strength: potency reaches 0.94 above a strength of 0.70 and selectivity reaches 0.93.

Which targets does it cover?

The potency model was fitted on 259 targets and the selectivity model on 271, out of 400 targets the software lists. Every published accuracy covers comparisons between targets in the fitted set. The site marks which targets were fitted and shows each one's own held-out accuracy, and a target that was not fitted cannot be scored.

How does it relate to the Kinase Foundation Model?

It uses the same two comparator layouts and the same featurization as the Kinase Foundation Model at kinasefoundationmodel.com, applied to G protein-coupled receptors instead of protein kinases.

Given a target and two ligands, which one is more potent

Two models over G protein-coupled receptors. Neither predicts an affinity. Each answers one two-class question with an ordering and a prediction strength, rather than a number on the potency scale.

Models built 12 September 2026 from a training set assembled the same day, out of ChEMBL together with Eidogen-curated data. Every accuracy on this site was measured on that build. Methods.

Potency · LSL

Ligand, sequence, ligand. One target, two ligands. Which ligand is more potent, with the strength of that call.

Selectivity · SLS

Sequence, ligand, sequence. One ligand, two targets. Which target it is more potent at.

Two comparison operators

Both models are structure-free and behave as comparison operators. The whole comparison is one row, and the order is the question: for potency it reads ligand A, sequence, ligand B; for selectivity it reads sequence A, ligand, sequence B. The protein enters only as a sequence embedding.

Prioritization is a sequence of comparisons rather than a list of predicted numbers. It is also why the data goes further than the measurement count suggests: n measurements at one target support n(n−1)/2 comparisons.

A worked potency comparison. Two ligands measured at the dopamine D2 receptor, a substituted benzamide at pKi 11.52 and an aryloxy-alkylamine at pKi 3.00. Each ligand becomes a 1,024-bit Morgan count fingerprint plus 14 descriptors; the target sequence becomes 480 numbers through ESM2. All three blocks enter a random forest as a single row, which returns which ligand is more potent. — **The potency comparison, end to end.** Both ligands and the target enter as one row, in the order ligand A, sequence, ligand B. Nothing is scored on its own and subtracted. The two compounds shown are real training measurements at the dopamine D2 receptor, 8.5 log units apart.

A, measured pK_i 11.52 at DRD2
`CCN1CCC[C@H]1CNC(=O)c1cc(Br)cc(OC)c1OC`

B, measured pK_i 3.00 at DRD2
`COc1cccc(CCc2ccccc2OCCN2CCc3ccccc3C2)c1`

Both structures paste into the box above with DRD2 selected.

A worked selectivity comparison. One ligand, a 4-anilidopiperidine, measured at two opioid receptors: pKi 10.89 at the mu target and 6.91 at the kappa target. Each target sequence becomes 480 numbers through ESM2 and the ligand becomes a 1,024-bit Morgan count fingerprint plus 14 descriptors. All three blocks enter a random forest as a single row, which returns which target the compound prefers. — **The selectivity comparison, end to end.** The ligand sits in the middle and the two targets are what is being compared, so the row reads sequence A, ligand, sequence B. The compound shown is a real training measurement, roughly 4 log units mu-selective over kappa.

**The ligand**, measured pK_i 10.89 at OPRM1 and 6.91 at OPRK1
`CCC(=O)N(c1ccccc1)[C@H]1CCN(C[C@H](O)c2ccccc2)C[C@H]1C`

**A kappa-selective counter-example**, measured pK_i 6.27 at OPRM1 and 10.21 at OPRK1
`Oc1cccc(CCN(CCc2ccccc2)CC2CCCCC2)c1`

Either structure pastes into the box above with OPRM1 and OPRK1 selected. They rank opposite ways.

What it scores

Both numbers below are held-out accuracy over comparisons the models never saw during fitting, measured on the 12 September 2026 build.

Potency

0.77

over 762,493 held-out comparisons, 235 targets

Selectivity

0.80

over 24,741 held-out comparisons, 211 targets

Those are the figures for answering every comparison. The models also return a strength, and accuracy rises with it. The ranking pages act on the strength and explain it.

Read the strength, not the ranking

Every answer carries one number, its strength, from 0.5, a coin flip, to 1.0. It is not the chance that a given answer is right. It is a ranking signal whose reliability has been measured, and these are the measured values.

Potency

Strength at or above	Share of predictions kept	Accuracy
0.50, answer everything	100.0%	0.77
0.60	57.1%	0.88
0.70	29.1%	0.94
0.80	12.4%	0.98
0.90	3.3%	0.99

Selectivity

Strength at or above	Share of predictions kept	Accuracy
0.50, answer everything	100.0%	0.80
0.60	77.0%	0.87
0.70	55.2%	0.93
0.80	36.5%	0.97
0.90	17.7%	0.99

Where it is weak

Accuracy is an average over targets and the spread behind it is wide.

Model	Targets scored	Median	With 50+ comparisons	Median of those	Of those, below 0.60
Potency	235	0.79	196	0.79	0
Selectivity	211	0.83	104	0.82	0

Closely related targets are still the hard ones. Across the 196 potency targets carrying at least 50 held-out comparisons the range runs from 0.61 to 1.00, and no target in that group now falls below 0.60. The muscarinic targets, M1 through M5, were the weakest group on the previous build and now score between 0.74 and 0.82 for potency and between 0.75 and 0.78 for selectivity.

A deep dive, beside a broad one

This site is a deep dive into one family. The roster is G protein-coupled receptors and nothing else, the data is ChEMBL together with Eidogen-curated GPCR data, and the models are free to be tuned to that family because they will never be asked about anything else.

The Family Foundation Model is the broad instrument, and it is built the other way round: two separately fitted models over 34 protein families, kinases, GPCRs, proteases, transporters, ion channels and the epigenetic readers, writers and erasers, with one compound-preference model ranking compounds at any of them and one target-preference model comparing targets drawn from families that never share an assay panel. It is trained on ChEMBL 37 alone, which is what lets its weights be downloaded freely.

The third site, the Kinase Foundation Model, is the other deep dive: protein kinases, with primary measurements from the Eidogen-Sertanty Kinase Knowledgebase and ChEMBL used to cross-validate rather than to train.

Model	Roster	Primary data
Family Foundation Model	34 families in two models	ChEMBL 37 alone
GPCR Foundation Model this site	G protein-coupled receptors	ChEMBL plus Eidogen-curated GPCR data
Kinase Foundation Model	protein kinases	the Kinase Knowledgebase, with ChEMBL used to cross-validate

All three run the same two comparator layouts, so the questions have the same shape and the answers read the same way. What differs is the trade: a deep dive has more measurements per target and a narrow enough roster to model on its own terms, so within its family it is the sharper instrument, while the cross-family models give that up and buy reach. Use this site when every target you care about is a GPCR. Use the cross-family models when the comparison itself crosses the boundary, which is what off-target triage and repurposing usually require.