KKB Kinase Model Performance — Q2-2026

Per-kinase activity models (active = IC50 ≤ 1 µM) trained on the KKB Q2-2026 release. One Random Forest per kinase target with ≥100 unique compounds and ≥10% inactive; 2,059 features (2,048-bit Morgan + 11 physicochemical); stratified 80/20 random split.

391

kinase models trained

0.907

median ROC-AUC

0.882

mean ROC-AUC (±0.089)

0.947

mean AUC, well-powered (95 at ≥2,000 cpds)

839,943

compound–target pairs

How to read this. The well-powered subset (95 kinases with ≥2,000 compounds) averages 0.947 AUC — consistent with prior kinase modeling. The lower overall mean reflects the long tail of small or low-dynamic-range targets. A handful of tiny-data models show AUC near 1.0 on a small random test fold and should be read with caution (use --scaffold for a stricter estimate on those).

ROC curves — all 391 models

Each faint curve is one kinase model, colored by training-set size; the bold red line is the mean of all 391 models. Well-powered models (bright) hug the top-left corner; the modest curves are small-data targets.

Performance vs data size

Bands: 205 kinases at AUC ≥ 0.90 · 126 at 0.80–0.90 · 60 below 0.80. Performance tracks data size: mean AUC climbs from 0.882 (<500 compounds) to 0.947 (≥2,000 compounds).

Strongest models (well-powered)

Kinase	Compounds	ROC-AUC	Accuracy	% inactive
TTK	4,404	0.984	0.951	13%
AKT3	3,406	0.984	0.946	63%
IKBKE	2,405	0.980	0.950	19%
FGFR3	5,963	0.977	0.943	16%
PLK1	5,422	0.977	0.917	33%
MET	15,062	0.975	0.924	26%
ZAP70	2,076	0.973	0.930	79%
RIPK1	3,682	0.972	0.924	46%
PTK2	5,541	0.971	0.918	29%
SYK	14,594	0.970	0.926	18%
ITK	4,485	0.970	0.910	49%
ACVR1	2,070	0.969	0.927	18%
FGFR1	11,798	0.969	0.911	46%
FLT4	3,134	0.969	0.890	68%
MAPKAPK2	5,371	0.969	0.916	49%

Weakest models (candidates for more data / review)

Kinase	Compounds	ROC-AUC	Accuracy	% inactive
PASK	119	0.495	0.792	77%
CDK17	119	0.500	0.667	66%
ACVR2A	115	0.526	0.826	82%
TIE1	104	0.567	0.667	61%
EPHA3	152	0.568	0.645	66%
STK10	255	0.571	0.549	44%
EIF2AK2	115	0.627	0.696	74%
EPHB1	141	0.641	0.621	62%
CASR	156	0.643	0.875	12%
TGFBR2	280	0.643	0.661	60%

Provenance. Produced by the internal AI kinase modeling pipeline. Trained on KKB-SAR only. Generated 2026-07-22 23:53:43 UTC.