What a kinase selectivity panel can teach an AI model, and where it misleads

The human genome encodes 518 protein kinases, the count Manning and colleagues reported in Science in 2002, and many of them share a closely related ATP binding pocket. That similarity is why selectivity is so hard to design and why profiling panels exist. A compound that looks clean against its target can still inhibit dozens of related kinases, and the off-target pattern often decides whether a series moves forward.

Panel data is attractive for machine learning because it is dense and structured. One compound produces hundreds of measurements under a single protocol, which is very different from the scattered potency values most discovery datasets contain. Panel data can make AI genuinely useful in kinase programs, mainly for deciding what to make and profile next, as long as the model is kept honest about what the panel actually measured.

Large panels showed how much selectivity varies

Davis and colleagues published a landmark study in Nature Biotechnology in 2011 that profiled 72 kinase inhibitors against 442 kinases with a competition binding assay. The results showed wide variation in selectivity across the inhibitors, approved drugs included, and the dataset became one of the standard benchmarks for computational selectivity models. Its value came from consistency. The same assay format ran across the whole matrix, so differences between cells in the table reflected the compounds and kinases instead of the lab.

Most internal datasets are messier. Programs profile some compounds at a single concentration as percent inhibition, follow up the hits with dose response, and switch panels or vendors as the project matures. A model that treats a 1 µM percent inhibition value, a binding Kd and an activity IC50 as interchangeable will learn the mix of formats as much as the chemistry.

Assay format and ATP concentration are features whether you model them or not

Binding assays and activity assays answer different questions. A binding assay reports how tightly a compound occupies a kinase, while an activity assay reports how much it slows phosphorylation, and for ATP-competitive compounds that depends on the ATP concentration in the well. Running every kinase at its own Km for ATP makes potencies more comparable across the panel. Running at a fixed high ATP concentration gets closer to cellular conditions and lowers the apparent potency of competitive inhibitors.

Both choices are reasonable. Trouble starts when results from both are pooled without a column saying which was used. The safer approach keeps format, ATP concentration and readout as explicit fields, trains separate models or includes them as inputs, and reports predictions in the same terms as the assay they will be checked against.

The best use is choosing the next experiment

Prediction across the entire kinome for a brand new scaffold remains unreliable, especially for kinases with little data. The stronger use is inside a program, where a model trained on the team's own panel results suggests which new analogs are likely to keep the selectivity profile and which are likely to pick up a known liability. Active learning formalizes this by ranking candidates partly on expected value and partly on how much each result would teach the model, while the chemists decide which ones to make.

This changes how profiling budgets are spent. Instead of running every analog through the full panel, teams can run a focused subset plus the kinases the model is least sure about, and send the full panel only for compounds that are serious candidates. The model earns its keep by reducing wasted measurements, and the medicinal chemists keep the decision about which molecules are worth making.

Negative results and coverage gaps matter as much as hits

Panel data has a quiet advantage over literature data because it records inactivity. Public datasets lean heavily toward active compounds since inactive results are rarely published, and models trained on them overestimate how often anything works. A team's own panels contain thousands of honest negatives, and keeping them with their tested concentrations is one of the most valuable things a data group can do.

Coverage is the other blind spot. Kinases outside the panel are simply unmeasured, and a model should say so instead of extrapolating. Reporting predictions with a confidence estimate and an explicit flag for kinases with little training data keeps the output honest, and it shows the team where an extra measurement would be most informative.

Starting with what you already have

Most kinase programs already hold enough panel data to start, scattered across vendor reports and spreadsheets. The first step is consolidating it with formats and conditions intact, then testing a simple model on the program's own history to see whether it would have ranked past compounds sensibly. QueryNow builds that consolidation and the first model with discovery teams, in their environment and on their data. Tell us the workflow you want gone, we build it in your environment in two weeks, and you pay $10,000 only after it meets the acceptance criteria you signed off on. Start at /build.