One model. 478 kinases.
Sequence and structure in, ranking out.

Per-target models know only about compounds — they can say nothing about a kinase you have no data for, and nothing about a resistance mutation, because changing one amino acid does not change the compound. This model reads the protein as well as the ligand, so a single trained object covers the panel.

Read the full results → Rank compounds coming soon

Screening performance: sorting compounds by potency

Two compounds, one kinase, which binds harder. A coin flip is 50%.

77.2% Correct on 2,965,273 held-out ligand pairs across 413 kinases. Matthews correlation 0.544.

88.7% When the two ligands genuinely differ by more than tenfold in potency, on 1,581,818 pairs. MCC 0.773. This is the call a chemist actually makes.

98.4% On the tenth of calls carrying the largest model margins, 296,528 pairs, held-out KKB compounds. The margin is a monotone transform of the score difference, not a calibrated probability: larger margins are associated with higher accuracy, which is what makes a threshold useful.

How the model works: an ABL1 kinase sequence is encoded by ESM2 into 480 numbers, each compound into a fingerprint plus size and atom counts; a random forest scores each compound against the kinase and the difference of the two scores decides which binds tighter. — The model is given the kinase sequence only — no structure, no docking, no binding-site definition. The ABL1 domain is drawn here as crystallised in PDB 1IEP purely to show where the two compounds bind.

478kinases in the panel

417,507training measurements

3.6×top-decile enrichment on compounds from a different source

How it works

A compound is described by a Morgan count fingerprint (radius 2, 1,024 bits) plus fourteen composition descriptors. A kinase is described by its amino-acid sequence, passed through ESM2-t12-35M and mean-pooled to 480 dimensions. The two are concatenated into a 1,518-dimensional vector and scored by a random forest trained on 417,507 KKB potency measurements.

The score is a predicted pIC50 on the model’s own scale. It is designed to be used as an ordering, not as a calibrated potency. Two compounds are compared by the difference of their scores, which makes the comparison exactly antisymmetric — swapping the two compounds flips the answer and never changes its magnitude.

What it was measured to do

Regime	Pairwise accuracy	Top-decile enrichment
Held-out compounds, same corpus	0.80	5.5×
Compounds from a different source (ChEMBL)	0.67	3.6×
Compound pairs separated by >1 log	0.77	—

The second row is the one to plan against. Same-corpus hold-out overstates deployment performance by roughly 0.3 in rank correlation, because literature-derived corpora share compound series.

Where it works, target by target

Performance is not uniform, so an average across the panel would be misleading. These are held-out ChEMBL compounds — structures the model never saw — scored per target. Agreement is the rank correlation with the measured order; decisive pairs are those whose measured potencies differ by more than tenfold.

Reliable

Order these with confidence. Agreement of 0.6 or better, and the top of the list is strongly enriched in genuinely potent compounds.

Target	Compounds	Agreement with measured order	Decisive pairs correct	Enrichment in top tenth
AURKA	221	0.83	92%	2.7×
ABL1	496	0.77	87%	3.8×
MTOR	169	0.77	89%	4.7×
SRC	370	0.71	88%	5.7×
CDK2	293	0.68	86%	4.5×
PIM1	162	0.67	82%	5.1×
BRAF	217	0.62	83%	4.5×

Useful for triage

Clearly better than chance and worth using to prioritise, but the ordering is loose enough that the top of the list should be confirmed rather than trusted.

Target	Compounds	Agreement with measured order	Decisive pairs correct	Enrichment in top tenth
CDK1	137	0.60	81%	4.9×
FGFR1	477	0.59	82%	3.1×
ERBB2	394	0.55	77%	2.8×
MET	467	0.50	79%	4.7×
CDK9	442	0.47	77%	3.9×
JAK3	377	0.46	73%	5.2×
MAPK14	145	0.46	74%	4.4×
SYK	363	0.45	79%	5.9×
PIK3CG	454	0.43	72%	3.6×
FLT1	132	0.43	73%	4.7×
TYK2	487	0.42	71%	3.4×

Full per-target results for all 413 evaluated kinases, including those that perform at or below chance, are in the model results report.

Limitations, stated plainly

Quote it per target, never in aggregate. External per-target rank correlation runs from 0.09 to 0.83. A minority of targets perform at or below chance. An average across the panel hides that spread, so the tool reports each selected kinase separately rather than pooling them.

Absolute potency does not transfer; ordering does. Only 3.9% of the training data is inactive, because the literature reports exact numbers for compounds that bind and vague ones for compounds that do not. The model therefore predicts roughly two log units too potent on chemistry outside its corpus. Use it to rank, not to estimate an IC50.

It does not yet generalise to kinases it has never seen. Ranking compounds for a kinase outside the 478-member panel measured at 0.48–0.52 — indistinguishable from chance. You may paste any sequence into the tool, and it will be scored, but anything outside the panel is labelled as unvalidated and should be treated that way.

Two further results worth knowing. On compound pairs whose ranking flips between two kinases — the selectivity question — ligand information alone scores exactly 0.500 while the full model reaches 0.712, which is the clearest evidence that the protein half is carrying real signal. Against that, a simple baseline that looks up a compound’s average potency in the training set scores 0.731 on the shared benchmark versus the model’s 0.723, so on the headline metric the model does not yet beat remembering which compounds are generally potent. Both numbers are in the full results.

Reports

Model results — architecture, benchmarks, ablations, per-target breakdown
Co-complex dataset — the structural companion set
Provenance — where every training measurement came from
Tripartite agent architecture — how this work was carried out

Reports are regenerated as the model changes; these links always point at the current version.

One model. 478 kinases.Sequence and structure in, ranking out.