# Kinase Foundation Model

> Two sequence-conditioned models that rank kinase inhibitor potency and
> selectivity. Trained on the Kinase Knowledgebase by Eidogen-Sertanty, Inc.
> These are random forests over molecular fingerprints and protein sequence
> embeddings, not language models. Principal author Steven M. Muskal, ORCID
> 0000-0002-3487-270X (https://orcid.org/0000-0002-3487-270X).

Each model answers one comparative question and returns a ranking with a
confidence, not an absolute potency.

- Potency: given one kinase sequence and two compounds, which compound that
  kinase prefers, that is which is more potent there.
- Selectivity: given one compound and two kinase sequences, which kinase the
  compound prefers.

A compound enters as a 1,024-bit Morgan count fingerprint plus 14 descriptors,
a kinase enters as 480 numbers from an ESM2 sequence embedding, and the forest
returns which of the pair wins, with a prediction strength from 0.5 to 1.0.

## Pages

- [Kinase Foundation Model](https://kinasefoundationmodel.com/): what the models are, how they are built and how they are tested
- [Overview](https://kinasefoundationmodel.com/overview.html): the two current models
- [Rank compounds by potency](https://kinasefoundationmodel.com/rank.html): run the potency model
- [Download the models](https://kinasefoundationmodel.com/models.html): run them locally, offline
- [Potency report](https://kinasefoundationmodel.com/reports/LigASeqLigB_v2_potency.html): input layout, training and test construction, accuracy by potency gap and by novelty
- [Selectivity report](https://kinasefoundationmodel.com/reports/SeqALigSeqB_v2_selectivity.html): the worked ABL1, dasatinib and GSK3B case, accuracy by separation and by confidence
- [Limitations](https://kinasefoundationmodel.com/limitations.html)

## Related

- [Family Foundation Model](https://familyfoundationmodel.com/): both comparator layouts over one roster spanning 34 protein families at once, embedded in the same two forests. One says which of two compounds a target prefers, the other which of two targets from different families a compound prefers. ChEMBL 37 alone, so its weights are downloadable. Breadth across families. Video overview of the approach all three models share: https://www.youtube.com/watch?v=VWMWP5hi3uA
- [GPCR Foundation Model](https://gpcrfoundationmodel.com/): the same two comparator layouts over G protein-coupled receptors, built on ChEMBL plus Eidogen-curated GPCR data. A deep dive into one family.

These kinase models are the other deep dive: their primary measurements come from the Kinase Knowledgebase, with ChEMBL used to cross-validate rather than to train. The two family sites buy depth in one family; the cross-family site buys breadth across 34.

- [Eidogen-Sertanty](https://eidogen-sertanty.com/): the company
- [Kinase Knowledgebase](https://eidogen-sertanty.com/kinasekbmarvin.php): the curated kinase structure activity data the models are trained on
- [Source and install](https://github.com/smuskal/KFM)
- [Steven M. Muskal](https://www.stevenmuskal.com/): principal author. The Kinase Foundation Model sits with the Family and GPCR models on one card at https://www.stevenmuskal.com/projects.html#foundation-models
- [Reverse Screen](https://reversescreen.ai/): given a molecule, the protein targets it might interact with. 27,797 co-crystal ligands from the Protein Data Bank indexed by predicted three-dimensional pharmacophore fingerprint, so docking runs only on the sites retrieval returns. Shown on this site under "What the crystal structures say". Paper: Muskal, S. M.; Nicola, G. ReverseScreen.ai: Pharmacophore-Guided Reverse Screening Across the Growing Co-Complex Proteome. bioRxiv 2026, doi:10.64898/2026.09.09.750461 (https://www.biorxiv.org/content/10.64898/2026.09.09.750461v1). Video: https://www.youtube.com/watch?v=mFRBfK6iv_w
- [PharmCast](https://pharmcast.ai/): the pharmacophore fingerprint predictor Reverse Screen is built on. Muskal, S. M.; McGregor, M. J. PharmCast: rapid generation of three-dimensional pharmacophore fingerprints from two-dimensional structure without conformer generation. bioRxiv 2026, doi:10.64898/2026.09.02.748999 (https://doi.org/10.64898/2026.09.02.748999).

## How to read a score

Each model returns a comparison between two things, on a scale from 0.5
undecided to 1.0 certain. Read it as an ordering within one target, and read
each target on its own scale, since the scales are not calibrated against each
other.