Which target a compound prefers, and which compound a target prefers

Two preference models across 34 protein families. One asks which of two compounds a target prefers, the other which of two targets a compound prefers.

The two models

Both models answer a question about preference. They differ in whose preference is being asked about, and that is the whole distinction.

The compound-preference model holds one target fixed and asks which of two compounds that target prefers. That is the question you ask of a hit list or a congeneric series.

The target-preference model holds one compound fixed and asks which of two targets that compound prefers. The two targets can come from different families, which is the off-target and repurposing question, or from one family, which is the selectivity question a chemist asks inside a series. The two are measured separately and every figure below says which one it describes.

In both cases the answer is an ordering with one number attached, its strength, which runs from 0.5 to 1.0. The strength is its own confidence: the further it sits from 0.5, the more often the ordering turns out right. Every page here prints the strength with the measured accuracy of its band.

	Compound preference	Target preference
You supply	one target, two compounds	one compound, two targets, from different families or from one
It answers	which compound the target prefers	which target the compound prefers
Used for	ordering a hit list or a series at one target	off-target triage and repurposing, and selectivity inside one family

Compound preference

The preference comparison. Two compounds,
molibresib and inobrodib, are put to one target, the histone acetyltransferase
p300. Each compound becomes a 1,024-bit Morgan count fingerprint plus 14
descriptors and the target sequence becomes 480 numbers through ESM2, and all
three blocks enter a random forest as one row. — **Compound preference.** One target held fixed and two compounds, here Molibresib and Inobrodib at Histone acetyltransferase p300. Both readings are Kd, so the comparison is endpoint matched. p300 carries two family labels, **reader** and **writer**: its acetyltransferase domain writes acetyl marks onto lysines and its bromodomain reads acetyl marks already there, so one protein does both jobs. How it was built and what it scores →
Molibresib, to paste into the tool

Inobrodib, to paste into the tool

Target: `Q09472`, the Histone acetyltransferase p300. Paste both compounds, pick that target, and you should get the same answer.

Accuracy

0.71

over 65,725 held-out comparisons, both compounds unseen

Targets

2,079

across 32 protein families

At strength ≥ 0.70

0.92

on 17% of comparisons

At strength ≥ 0.90

0.98

on 1.2% of comparisons

Strength at or above	0.50	0.60	0.70	0.80	0.90
Accuracy	0.71	0.84	0.92	0.96	0.98
Share of comparisons kept	100%	43%	17%	6.1%	1.2%

Rank compounds at a target → Full compound-preference results →

Target preference

A cross-family comparison. Olanzapine is put
to two targets from different protein families: the histamine H1 receptor, drawn
as a seven-helix bundle in a membrane, and hERG, drawn as four subunits around a
central pore. 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 is more potent at. — **Target preference.** Olanzapine put to the histamine H1 receptor, a G protein-coupled receptor, and to hERG, a voltage-gated ion channel. Both are real measurements from the training data: pKi **8.50** at H1 against **4.44** at hERG, so H1 is 4.1 log units the more potent of the two. The model prefers **H1 at strength 0.88**, which is the measured order. The target pictures are schematics of each protein's architecture, not structures.
Olanzapine, to paste into the tool

Targets: `P35367` for H1 and `Q12809` for hERG. Paste the compound, pick those two, and you should get the same answer.

Two targets from different families

Accuracy

0.75

over 8,689 held-out comparisons, every ligand unseen

Targets

2,279

across 34 protein families

At strength ≥ 0.70

0.89

on 49% of comparisons

At strength ≥ 0.90

0.97

on 21% of comparisons

Strength at or above	0.50	0.60	0.70	0.80	0.90
Accuracy	0.75	0.82	0.89	0.94	0.97
Share of comparisons kept	100%	71%	49%	34%	21%

Two different targets from one family

Accuracy

0.78

over 32,738 held-out comparisons, every ligand unseen

Targets

2,279

across 34 protein families

At strength ≥ 0.70

0.91

on 56% of comparisons

At strength ≥ 0.90

0.98

on 21% of comparisons

Strength at or above	0.50	0.60	0.70	0.80	0.90
Accuracy	0.78	0.85	0.91	0.95	0.98
Share of comparisons kept	100%	76%	56%	38%	21%

The two kinds of comparison are fitted together and measured apart, and they are never averaged into one number. Read each figure with the question it answers.

What the accuracy is worth

An accuracy means something only beside what you could get without the model. Two weaker answers are measured on the same held-out comparisons: the family labels alone, and the same model with the ligand blinded so only the two sequences remain.

Answering with	Across families	Within one family
A family prior, the family labels alone	0.66	0.50
The same model with the ligand blinded	0.71	0.67
The model	0.75	0.78

Inside one family a family prior is chance by construction, which is the point of that column: every point above 0.50 comes from the comparison itself. Against its own sequence-only twin the model gains 11.2 points within a family and 4.4 points across them, so the compound carries more of the call inside a family than across one.

Rank targets for a compound → Full target-preference results →

What "prefers" means here, and why that word

The training data carries five endpoints, and they do not all measure the same thing. Ki and Kd are binding constants. EC50 is a functional response. Kb is an affinity estimated from functional data. IC50 is a concentration that halves an effect, which depending on the assay can be either. Calling every answer "binds more tightly" would claim a mechanism that most of this data does not measure, and even "more potent" sits awkwardly over Ki and Kd, since potency and affinity are separate ideas in pharmacology.

So the model is not told which of those it is looking at, and neither are you. Comparisons are matched by endpoint, so every one of them is between two readings of the same kind: two Ki values, or two EC50 values, never one of each. What comes back is the ordering of that pair. "Prefers" is the word that covers all five without overstating any of them: it says the compound engages one target more than the other, on whatever scale the two readings were taken, and it stops there.

Read it as a ranking between the two things you asked about. What that engagement costs or buys biologically is a separate question, and this model has no access to it.

The paper. Two Comparators May Be All We Need, Steven M. Muskal. bioRxiv, first posted 21 September 2026 (DOI, always the current version: 10.64898/2026.09.15.751854).
The overview, in video. Two Comparisons May Be All We Need: A Foundation Model for Drug Discovery That Ranks, Not Predicts.

Compound preference built 13 September 2026, target preference 21 September 2026, both from ChEMBL 37 alone. The full results behind these numbers are on the methods page.

Family deep dives, beside this one

These two models are the broad instrument: two separately fitted models over one roster of 34 protein families, so one preference model ranks compounds at a kinase, a protease or an ion channel without being told which, and one target-preference model ranks two targets for a compound whether they sit in different families or in the same one.

Two sibling sites do the opposite, and deliberately. Each takes one family and goes deep, on data assembled for that family rather than on a single public release.

Model	Roster	Primary data
Family Foundation Model this site	34 families in two models	ChEMBL 37 alone, which is what lets the weights be downloaded freely
GPCR Foundation Model	G protein-coupled receptors	ChEMBL plus Eidogen-curated GPCR data
Kinase Foundation Model	protein kinases	the Eidogen-Sertanty Kinase Knowledgebase, with ChEMBL used to cross-validate rather than to train

Both siblings run the same two layouts this site runs, one that ranks compounds at a fixed target and one that ranks targets for a fixed compound, so the shapes of the questions match and the answers read the same way. What differs is the trade. A deep dive has more measurements per target and a roster narrow enough to model on its own terms, so within its family it is the sharper instrument. These models give up that depth and buy reach: one roster, one pair of forests, and a question you can ask across families rather than inside one.

They answer different questions and neither replaces the other. A deep dive is the sharper instrument when every target you care about sits inside its family. These reach wider: one roster, one pair of forests, and a panel that can mix families freely, which is what off-target triage and repurposing usually require.

The GPCR deep dive → The kinase deep dive →