What does the GPCR Foundation Model predict?

It does not predict an affinity. It answers a two-class comparison. The potency model takes one receptor and two ligands and returns the probability that the first ligand is more potent. The selectivity model takes one ligand and two receptors and returns the probability that the first receptor binds it more tightly.

Answering every comparison, potency is 0.640 accurate over 413,181 held-out comparisons across 182 receptors, and selectivity is 0.694 over 15,802 comparisons across 170 receptors. Accuracy rises with prediction strength: potency reaches 0.867 above a strength of 0.70 and selectivity reaches 0.841.

Which receptors does it cover?

The potency model was fitted on 251 receptors and the selectivity model on 255, out of 796 receptor names the software recognises. Every published accuracy covers comparisons between receptors in the fitted set. The site marks which receptors were fitted and shows each one's own held-out accuracy.

How does it relate to the Kinase Foundation Model?

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

Given a receptor and two ligands, which one binds better

Two models over G protein-coupled receptors. Neither predicts an affinity you can quote. Each answers a two-class comparison, and that is the whole product.

Preview build, 11 September 2026. The models here were retrained on the full measurement set today. The ranking tools are being connected to them now; until that finishes the Rank buttons say so rather than failing quietly. Every number on this site is measured from the retrained models.

Potency · LSL

Ligand, sequence, ligand. One receptor, two ligands. Which ligand is more potent. Returns the probability that the first one wins.

Selectivity · SLS

Sequence, ligand, sequence. One ligand, two receptors. Which receptor binds it more tightly.

How a comparison is put

Each model takes one row with three blocks in it. Which block sits in the middle is the whole difference between them.

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 receptor sequence becomes 480 numbers through ESM2. All three blocks enter a random forest as a single row, which returns which ligand binds tighter. — **The potency comparison, end to end.** Both ligands and the receptor 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`

Paste both into the box above with DRD2 selected to reproduce this comparison.

A worked selectivity comparison. One ligand, a 4-anilidopiperidine, measured at two opioid receptors: pKi 10.89 at the mu receptor and 6.91 at the kappa receptor. Each receptor 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 receptor the compound prefers. — **The selectivity comparison, end to end.** The ligand sits in the middle and the two receptors 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`

Paste either into the box above with OPRM1 and OPRK1 selected. The model should send them opposite ways.

What it scores

Both numbers below are held-out accuracy over comparisons the models never saw during fitting, measured on the 11 September 2026 build.

Potency

0.640

over 413,181 held-out comparisons, 182 receptors

Selectivity

0.694

over 15,802 held-out comparisons, 170 receptors

Those are the figures for answering every comparison. The models also report a strength, and accuracy rises steeply with it. That is the number the ranking pages act on, and it is explained on each of them.

Read the strength, not the ranking

Strength is the larger of the two returned probabilities, so it runs from 0.5, a coin flip, to 1.0. It is not the probability 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%	0.640
0.60	26.5%	0.793
0.70	6.9%	0.867
0.80	1.6%	0.887
0.90	0.4%	0.869

Potency turns over above 0.80, falling from 0.887 to 0.869 on 1,537 comparisons, so the pages do not offer a cutoff past 0.80.

Selectivity

Strength at or above	Share of predictions kept	Accuracy
0.50, answer everything	100%	0.694
0.60	61.6%	0.776
0.70	32.8%	0.841
0.80	14.7%	0.915
0.90	4.4%	0.968

Where it is weak

Accuracy is an average over receptors, and the spread behind it is wide. This is the part of the model most worth knowing before acting on a result.

Model	Receptors scored	Median	With 50+ comparisons	Median of those	Of those, below 0.60
Potency	182	0.638	137	0.619	57
Selectivity	170	0.688	75	0.675	21

Closely related receptors are the hard ones, which is what you would expect. Muscarinic receptors sit near chance for selectivity, because M1 through M5 differ by little that a sequence embedding sees. Opioid receptors reach 0.91.

The models know the receptors they were fitted on, and only those. Potency was fitted on 251 receptors and selectivity on 255. Every accuracy quoted here covers comparisons between receptors in that set. Nothing on this site measures how the models behave on a receptor they have never seen, and no claim is made about it. The ranking pages mark which receptors are fitted.

What it is not

It is a comparator. It has no opinion on whether either ligand is active at all, it does not return an affinity, and a ranking of two inactive compounds is still a ranking. Read it as a triage tool over a set you already have reason to care about.