The two version 2 models side by side. Left: which of two compounds binds this kinase more tightly - ligand A, the kinase sequence and ligand B enter one random forest in that fixed order, each ligand as a 1,024-bit Morgan count fingerprint plus 14 descriptors and the sequence as 480 ESM2 numbers, returning a confidence bar from A binds tighter to B binds tighter. Right: which of two proteins binds this ligand more tightly - sequence A, the ligand and sequence B in fixed order, returning a bar from protein A has greater affinity to protein B has greater affinity.

Version 2: potency ranking (left) and selectivity (right). Click the diagrams to visit kinasefoundationmodel.com

Kinase Foundation ModelNew!

Two models, one question each. Give the first a kinase sequence and two compounds and it says which binds more tightly. Give the second one compound and two kinases and it says which kinase the compound prefers, the selectivity question. Both read the kinase’s amino-acid sequence itself, with no structure, no docking and no binding-site definition, and both are trained on the Kinase Knowledgebase.

Tested on ChEMBL measurements the models never saw: 69.3% over 1,836,100 held-out potency comparisons across 477 targets, and 75.3% over 3,137,588 selectivity comparisons. Acting only on the confident calls, at prediction strength 0.70 and above, raises those to 87.8% and 92.3%. Version 1, the original scorer across 478 kinases, remains published alongside them.

Explore the Kinase Foundation Model →

Kinase Knowledgebase: growth to more than 3 million biological activity data points

Four-panel performance summary of the KKB per-kinase activity classifiers: ROC curves grouped by training-set size, the distribution of ROC-AUC across 392 kinases, accuracy rising with the number of measured compounds, and mean AUC by data-size band.

Per-kinase model performance across 392 kinases

Kinase Knowledgebase (KKB)

Currently the Kinase Knowledgebase Q2 2026 Release includes the following data:

Journal articles and patents: 10,326
Number of Biological Activity Data Points: 3,336,903
Number of unique kinase molecules with annotated assay data: 509,691
Number of all unique kinase molecules from patents and articles (with or without bio-activity data): 854,437
Number of unique kinase targets with assay data: 579
Number of annotated assay protocols: 105,916
Machine-learning models built from this release’s SAR: activity classifiers across 392 kinase targets, median ROC-AUC 0.91 (0.95 for well-studied kinases) – view model performance report

To show what this depth of curation supports, we build machine-learning models from the KKB SAR and report how they perform: an activity classifier for each of 392 kinase targets and a potency regressor for 319 of them. The panel at left summarises classifier accuracy: ROC curves grouped by how much data each kinase has, the spread of ROC-AUC, and how accuracy rises with the depth of measured chemistry.

Accuracy is highest for compounds chemically related to what a target already has in KKB and declines for novel scaffolds, so every prediction is reported with a similarity score against the model’s own training set. Performance is also measured against published data absent from the knowledgebase.

Search the Kinase Knowledgebase →

The ChIP search cycle in five stages: a gene holds a reaction sequence plus the exact building blocks for each reactant slot; enumeration runs every block combination through the sequence; your model filters, featurises and scores every product; elitist truncation keeps the best genes; crossover and mutation produce the next generation, and the cycle repeats.

How the search works. Click the diagram for the full method.

ChIP™ de Novo Design

Design molecules you can actually make. ChIP searches reaction space, not molecule space: it evolves synthetic protocols (a sequence of validated reactions plus the specific catalogued building blocks entering each one) and scores the molecules those protocols produce with your predictive model. Every design that comes out carries an executable synthesis from purchasable starting materials.

85 validated reaction transforms drawing on 853,409 catalogued building blocks, across 154 reactant slots. Potency, selectivity, an ADME or physicochemical property: any per-molecule model can drive the search, and every campaign is paired with a matched random control.

How ChIP works → Random forest potency example → Non-obvious me too example →

Scatter plot of predicted against reference pharmacophore similarity for 25,165 pairs of molecules, colored by their two-dimensional similarity, green for the most dissimilar pairs through red for the least. The points track the diagonal in every band, showing the surrogate is not re-deriving 2D similarity.

Predicted against reference pharmacophore similarity, colored by 2D similarity, PharmCast version 10

Turbocharged Pharmacophoric SimilarityNew!

Compare, rank and cluster molecules by the three-dimensional features they present, from flat structure alone. The comparison is made on our pharmacophoric fingerprints, the 10,549-bit three-point PolyPharmPrint™ encoding; what is new is how fast the answer arrives.

A full pharmacophoric comparison of two molecules takes 0.584 milliseconds against 5.71 seconds for the reference calculation, and agrees with it at Pearson 0.980. Ranking a reference against the whole 4.65 million compound screening collection takes 29.3 minutes, where the conventional route needs 3,697 core hours. In a retrieval test the true nearest neighbor is in the surrogate's top 100 for 89.3% of queries and its top 500 for 97.5%.

How it works → See the pipeline →

Kinase Contact MapsNew!

What the ligand is actually doing, measured from the deposited coordinates: hydrogen bonds with a real donor to hydrogen to acceptor angle wherever the heavy atoms fix the hydrogen, halogen bonds, salt bridges, pi stacking, cation to pi and metal coordination. Every compound shown is the same molecule as a ligand in a Protein Data Bank kinase structure and has Kinase Knowledgebase activity against that same target; where the two records describe its 3D form with different precision, the figure says so. A different structure each time you load this page.

Contact map for MAPK1 with ligand UO5 from PDB entry 7NR3, IC50 5.78 nM in the Kinase Knowledgebase, showing the measured interactions on the flat structure

MAPK1 Mitogen-activated protein kinase 1 · PDB 7NR3 · ligand UO5 · IC50 5.78 nM · 3 KKB SAR points against this target

How these are measured, with this structure →

Screen a molecule at reversescreen.ai →

Every co-crystal ligand in the Protein Data Bank, embedded by pharmacophore fingerprint. Drag to rotate.

Reverse ScreenNew

Forward screening asks which of many ligands bind one target. Reverse Screen asks the opposite: given a molecule, which proteins might it interact with. That is the question that arises once a compound exists, and the only question available for a molecule a generative method proposes, because a designed structure has no measured target at all.

Docking one molecule into every characterized site answers it directly and takes 40.5 hours. Instead, every co-crystal ligand in the Protein Data Bank is indexed by the three-dimensional pharmacophore fingerprint it presents, predicted from two-dimensional structure by PharmCast, together with the UniProt accessions it was solved against. A query is fingerprinted in 4 milliseconds and compared against all 27,797 indexed ligands in 40 milliseconds, and the proteins its nearest neighbors were crystallized against come back as the candidate pool. AutoDock Vina then docks only into those.

Across 3,000 held-out molecules, pooling the five most similar indexed ligands gives 21.1 candidate proteins and contains the molecule's own known target 48.8 percent of the time. Pooling twenty-five gives 104.5 proteins and 60.8 percent.

A worked example. Screening orforglipron returns matrix metalloproteinase 9 at rank 5 of 48, through the deposited ligand E40 in 4WZV at a pharmacophore similarity of 0.803. The two molecules share no scaffold, which is the point: the fingerprint compares the features a molecule presents in three dimensions rather than the way it is drawn. Docking into that site gives a top pose at −12.1 kcal/mol, and all eight poses land in the same groove the deposited ligand occupies.

See the eight poses in three dimensions →

Try Reverse Screen →

Two-panel diagram. Panel one: a compound, with its assay, primary target and potency, is reduced by a one-way SHA-256 hash to an irreversible encoded fingerprint, and the original structure is kept private. Panel two: two parties each hold a private list of encoded fingerprints, compare them, and learn only the percentage that overlaps.

Illustrative example. Figures shown are for explanation only and are not results from our datasets.

Dataset Overlap AnalysisNew

How much of a dataset do you already have? Our toolkit answers that for any two SAR datasets, in any target family, without either side revealing a structure. Each party encodes its own data locally into irreversible SHA-256 fingerprints; only fingerprints are compared, and only the overlap figure is shared.

How it works → Download the toolkit →

The Oncology Knowledgebase interface, showing a curated compound record beside a panel reading over 127,000 biological activity data points across 1,100 unique oncology targets.

Oncology Knowledgebase (OKB)

Currently the Oncology Knowledgebase Q2 2023 Release includes the following data:

Journal articles and patents: 977
Number of Biological Activity Data Points: 146,206
Number of molecules from patents and articles: 66,198
Number of unique oncology targets with assay data: 1,158
Number of annotated assay protocols: 5,614
Number of disease models: 137

Learn More →

Advancing Drug Discovery Through AI-Powered Solutions

Our Products