KKB Kinase Model Performance

Kinase Knowledge Base Q3-2026-unclipped · generated 2026-10-07 03:35:06 UTC

Every KKB release ships with two machine-learning models per kinase target, built directly from the curated SAR in that release. Together they answer the two questions most often asked of a knowledge base: is this compound active at this kinase, and how potent is it likely to be.

The two model families

Activity classifiers predict the probability that a compound is active at a given kinase, where active means IC₅₀ of 1 µM or better. 390 targets qualify, with a median ROC-AUC of 0.916. Across the 97 best-studied targets — those with at least 2,000 measured compounds — the median rises to 0.965.

Potency regressors predict pIC₅₀ directly, so they distinguish a compound expected at 100 nM from one expected at 10 µM rather than calling both active. 316 targets have enough exact measurements to support one, with a median error of 0.47 log units. Averaged across all of them, 63% of predictions land within half a log unit of the measured value and 88% within one log unit.

Performance scales with the depth of the data

Both families improve as a target accumulates measured compounds, which is the central argument for a deeply curated knowledge base: the models are only as good as the SAR beneath them, and KKB's depth on well-studied kinases is what makes these accuracies achievable.

Independent validation

Internal metrics are out-of-bag estimates on KKB itself. To test the models against chemistry they had never seen, every model was additionally evaluated on 59,226 published kinase measurements from 28 well-studied targets that are absent from KKB — compounds no model was trained on, from sources independent of the knowledge base.

Median potency error of 0.84 log units on entirely unseen chemistry means a typical prediction lands within roughly a factor of 7 of the measured IC₅₀. Performance is strongest for compounds structurally related to what a target already has in KKB and falls away for novel scaffolds, which is expected of any ligand-based method and is the reason each prediction ships with a similarity score.

Best-performing well-studied targets

The best-scoring 12 of the 97 targets with at least 2,000 measured compounds. Sparse targets can post unstable near-perfect scores and are excluded for that reason. Click a column heading to re-sort; click a kinase to see its out-of-bag performance.

All internal statistics are out-of-bag: each forest is trained on 100% of its target's data, and every compound is scored only by the trees that did not see it. ROC-AUC uses those probabilities; accuracy is the share called correctly at 0.5. A dash in the independent column means the target is outside the fixed 28-target panel used for independent validation, which is held constant between releases so that number stays comparable; it is not a statement about that model's performance.

Every model in this release

All 390 activity classifiers and their matching potency regressors, largest target first. Type to filter, click any heading to re-sort. Kinases shown as links open their out-of-bag performance — 214 of the 390 have one. A model needs roughly 500 measured compounds before a per-target plot carries real information; below that the figure would imply more precision than the data supports, so the numbers are listed without one.

Independent columns are populated for the 28-target validation panel only. A dash means the target is outside that panel, not that the model performed poorly. Targets with few compounds have less stable statistics, so read their numbers with more caution.

×

Method

One random forest per kinase target meeting quality thresholds: at least 100 unique compounds, at least 10% inactive, and at least 10 examples of each class. Compounds are represented by 2,059 features — a 2,048-bit Morgan fingerprint at radius 2 plus 11 physicochemical descriptors. Classifier and regressor are trained on the identical curated cohort for each target and each is trained once, on 100% of it. Every accuracy reported here is out-of-bag: a compound is scored only by the trees of its forest that did not see it, so no data is set aside. Potency is normalised on fixed bounds so predictions are directly comparable between kinases.

Censored measurements are handled explicitly rather than treated as exact values: a bound is used where it decides the activity class and set aside where it does not.

Models are rebuilt from each quarterly KKB release. Before rebuilding, the previous release's models are tested against the compounds newly added in the current release — a forward-looking check on chemistry that did not exist when those models were fitted. Eidogen-Sertanty · Kinase Knowledge Base Q3-2026-unclipped.