# GPCR Foundation Model > Two machine-learning comparators over human G protein-coupled receptors. > Given one target and two ligands, which ligand is more potent. Given one > ligand and two targets, at which target it is more potent. Neither model > predicts an affinity; each answers a two-class comparison and reports a > prediction strength whose reliability has been measured. Built by > Eidogen-Sertanty, Inc. Principal author Steven M. Muskal, ORCID > 0000-0002-3487-270X (https://orcid.org/0000-0002-3487-270X). ## What it is Reverse and forward questions about GPCR pharmacology are usually answered with a predicted number. These models answer a comparison instead, because a comparison is what a ranking needs and it is what the training data supports. - **LSL, potency.** Ligand A, target sequence, ligand B, in one row. Returns which ligand is the more potent at that target, with a prediction strength. - **SLS, selectivity.** Target A, ligand, target B, in one row. Returns at which target the ligand is more potent, with a prediction strength. Both enter every comparison in both argument orders and average, so comparing A with B and B with A return probabilities that sum to exactly 1. ## The measured numbers Build of 12 September 2026. The holdout is 10 percent of compounds, stratified across targets, so a compound appears on one side only. It is not a temporal split. - Potency: 0.77 accuracy over 762,493 held-out comparisons, 235 targets scored - Selectivity: 0.80 accuracy over 24,741 held-out comparisons, 211 targets scored - Targets fitted: 259 for potency, 271 for selectivity, out of 400 targets listed - Training set: 389,013 measurements on 225,676 compounds across 284 targets - The two served models are fitted on every measurement with no holdout; the accuracies above come from a second pair built by the identical pipeline on the fitting side of the split - Ligand block: Morgan count fingerprint, radius 2, 1,024 bits, plus 14 descriptors - Target block: ESM2 esm2_t12_35M_UR50D, mean pooled over residues, 480 dimensions - Row widths: 2,556 for potency, 1,998 for selectivity - Model: random forest, 300 trees Prediction strength is max(p, 1 - p) of the returned probability p, so it runs from 0.5 to 1.0. It is not the chance that an answer is right. Measured accuracy by band: - Potency, at or above 0.60: 0.88 on 57.1 percent of comparisons; 0.70: 0.94; 0.80: 0.98; 0.90: 0.99 - Selectivity, at or above 0.70: 0.93 on 55.2 percent of comparisons; 0.80: 0.97; 0.90: 0.99 Both curves rise all the way to the 0.90 band, so every cutoff through 0.90 is offered. On the previous build potency fell back above 0.80; it no longer does. ## What it does not do It is a comparator, not a potency predictor. It has no opinion on whether either compound is active at all. Accuracy covers comparisons between targets the model was fitted on; nothing here measures behavior on a target it has never seen, and no such claim is made. Per-target accuracy varies widely, and closely related targets are the hard case. ## Data Measurements from ChEMBL together with Eidogen-curated data. Endpoints are IC50, Ki, Kd, EC50 and Kb, so "more potent" spans binding affinity and functional potency. Activity is derived from value and unit rather than read from a precomputed column. Duplicates collapse by median, never by most potent. ## Pages - [Overview](https://gpcrfoundationmodel.com/): what the models answer and what they score - [Rank compounds against a target](https://gpcrfoundationmodel.com/rank-lsl.html): draw or paste structures, get a ranking - [Compare two targets](https://gpcrfoundationmodel.com/rank-sls.html): one ligand, two targets - [Methods](https://gpcrfoundationmodel.com/methods.html): data, filtering, featurization, hyperparameters - [Targets](https://gpcrfoundationmodel.com/targets.html): every target with its counts and its own accuracy - [Limitations](https://gpcrfoundationmodel.com/limitations.html): where the models are weakest ## Related work by the same group This site is a deep dive into one family: G protein-coupled receptors only, built on ChEMBL together with Eidogen-curated GPCR data. Two sibling sites sit either side of it. - [Family Foundation Model](https://familyfoundationmodel.com/): the broad instrument. Both comparator layouts over 34 protein families at once, embedded in the same two forests, so one compound preference model ranks compounds at any of them and one target preference model compares targets drawn from different families. Trained on ChEMBL 37 alone, so its weights are downloadable. Use it when the comparison itself crosses a family boundary. - [Kinase Foundation Model](https://kinasefoundationmodel.com/): the other deep dive. The same two comparator layouts over protein kinases, with primary measurements from the Eidogen-Sertanty Kinase Knowledgebase and ChEMBL used to cross-validate rather than to train. It is the model family this one was built from, so start there for the kinase equivalent of any question here. All three run the same two layouts, so the questions have the same shape. What differs is the trade: a deep dive has more measurements per target and a narrow roster to model on its own terms; the cross-family models give that up for reach. - [PharmCast](https://pharmcast.ai/): predicts a three-dimensional pharmacophore fingerprint, 10,549 bits wide, directly from a SMILES string without generating conformers. - [Reverse Screen](https://reversescreen.ai/): given a molecule, which protein targets it might interact with, by pharmacophore retrieval across every co-crystallized ligand in the Protein Data Bank. - [Eidogen-Sertanty](https://eidogen-sertanty.com/): curated kinase structure activity data and the models trained on it. ## Citation Muskal, S. M. GPCR Foundation Model: pairwise comparators for potency and selectivity across G protein-coupled receptors. In preparation, 2026.