Given one compound and two targets from different families, which binds it more tightly

One model, one question. It returns a probability, never an affinity and never a predicted Ki.

Model built 13 September 2026 on 89,888 cross-family comparisons over 22,588 ligands and 1,879 targets, from ChEMBL 37 alone. Methods.

A cross-family comparison. One
compound is put to two targets from different protein families, EGFR, a kinase,
and CYP3A4, a cytochrome P450. 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 binds the compound more tightly.

What it is for

Off-target triage and repurposing. Give it a compound and a set of targets drawn from different families, and it ranks which the compound leans toward. The sibling models each compare within one family; this one crosses the boundary.

What it scores

Held-out accuracy over comparisons the model never saw, on a compound-disjoint split: a ligand lands on one side only, and no ligand appears on both.

Accuracy

0.750

over 8,689 held-out comparisons on 2,195 ligands

Targets

1,879

across 38 protein families

The two baselines, beside the headline

A cross-family comparator can look respectable while knowing nothing about the compound, because some families simply bind more tightly than others. So the honest question is not what the model scores, it is how much better it does than knowing only which two families are involved.

What is being asked	Accuracy
The model	0.750
Always pick whichever family usually wins that pairing	0.654
The same forest with the ligand removed	0.710

The compound itself contributes about four points on top of target identity. That margin held across three independently built versions of this model, at 3.9, 4.5 and 4.1 points. It is real, it is reproducible, and it is smaller than the headline number on its own suggests.

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. Accuracy rises with it and coverage stays usable.

Strength at or above	Share of comparisons kept	Accuracy
0.50, answer everything	100.0%	0.750
0.60	74.4%	0.815
0.70	52.0%	0.877
0.80	35.9%	0.933
0.90	22.1%	0.966

A near-tie is a near-tie

Accuracy depends on how far apart the two measured values really are.

True separation	Comparisons	Accuracy
Under half a log	2,283	0.573
Half a log to one log	1,766	0.701
One to two logs	2,359	0.797
Beyond two logs	2,281	0.915

Rank targets for a compound →