Drug Discovery Software & AI Solutions

Leading provider of pharmaceutical research tools

Our Products

Reverse Screen: a molecule’s likely targets →

ZGW and AMVP →

FSP and the real loop VIW →

Potency ladder: fourteen protein sites, each co-crystal ligand placed by its measured potency

Bound is not necessarily potent →

Kinase contact map: a ligand's measured interactions drawn from the crystal structure

Kinase contact maps →

Our Solutions

Family Foundation Model: the compound-preference model ranking two compounds at one target, and the target-preference model comparing two targets from different families for one compound

Family Foundation Model New!

The broad instrument. Two separately fitted models over one roster spanning 34 protein families. One says which of two compounds a target prefers, the other says which of two targets from different families a compound prefers. Trained on ChEMBL 37 alone, so both are freely downloadable. 0.71 on 65,725 held-out comparisons over 2,079 targets, and 0.75 on 8,689 over 1,879 targets. The GPCR and kinase models below are family deep dives beside it.

Explore the model →

▶ Watch the overview

GPCR Foundation Model: the potency model ranking two compounds against one G protein-coupled receptor, and the selectivity model comparing two targets for one compound

GPCR Foundation Model New!

Two models over G protein-coupled receptors, one question each. Give the first a target sequence and two compounds and it ranks which is more potent. Give the second one compound and two targets and it says which the compound prefers. No structure or docking needed. A deep dive into one family, on ChEMBL together with Eidogen-curated GPCR data, 389,013 measurements across 284 targets, and a measured accuracy published for every target with held-out data.

Explore the model →

Kinase Foundation Model version 2: the potency model ranking two compounds against one kinase, and the selectivity model ranking two kinases for one compound

Kinase Foundation Model New!

Two models, one question each. Give the first a kinase sequence and two compounds and it ranks which is more potent. Give the second one compound and two kinases and it says which the compound prefers, the selectivity question. No structure or docking needed. A deep dive into one family: primary measurements from the Eidogen-Sertanty Kinase Knowledgebase, with ChEMBL used to cross-validate rather than to train.

Explore the model →

The reverse screening index in three dimensions: 27,797 co-crystal ligands from the Protein Data Bank, each a point placed by the pharmacophore fingerprint it presents and colored by target class, with orforglipron and daraxonrasib marked

Reverse Screen New!

Given a molecule, the proteins it might interact with. Every co-crystal ligand in the Protein Data Bank is indexed by the pharmacophore fingerprint it presents, so a query returns candidate targets in 52 milliseconds instead of the 40.5 hours it takes to dock into every site. AutoDock Vina then docks only into what retrieval returns.

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, two molecules that share no scaffold. Docking into that site gives a top pose at −12.1 kcal/mol.

Try Reverse Screen →

▶ Watch the overview

Kinase Knowledgebase (KKB)

Comprehensive database for kinase research: 3.3M curated bioactivity data points across 579 kinase targets (Q2 2026 release)

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

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. Accuracy is highest for compounds chemically related to what a target already has in KKB and declines for novel scaffolds, so every prediction carries a similarity score against the model's own training set.

Model performance report →
Explore KKB →

The ChIP search cycle in five stages. 1 GENE: a reaction sequence plus the exact building blocks for each reactant slot, one to three steps deep, drawn as chains of colored shapes. 2 ENUMERATE: every block combination is run through the sequence in both orientations, fanning out to many product chains. 3 SCORE: your model filters, featurizes and predicts for every product, and the scored cloud narrows through a funnel from better to best. 4 SELECT: elitist truncation keeps the best genes, selection acting on genes rather than on molecules. 5 VARY: crossover swaps a step's blocks or splices two routes to produce the next generation.

Swipe the diagram → · tap it for the full method

ChIP™ de Novo Design

Design molecules you can actually make. ChIP searches reaction space rather than molecule space: it evolves synthetic protocols from 85 validated reactions and 853,409 cataloged building blocks, scores what they produce with your own model, and returns every design with an executable synthesis from purchasable starting materials. Score on pharmacophore similarity instead and it designs the non-obvious “me-too”: high pharmacophore similarity, low 2D similarity.

Two rotating three-dimensional overlays side by side, the ChIP design in orange over Orforglipron in blue. Left, aligned by molecular shape; right, aligned by pharmacophore feature overlap. Captioned: not obvious, me too.

Swipe the card → · the ChIP design in orange over Orforglipron in blue, aligned two ways.

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

Scatter plot of predicted against real pharmacophore similarity for pairs of molecules, colored by their two-dimensional similarity, green for very dissimilar through red for least dissimilar. The points track the diagonal at every color, showing the surrogate is not re-deriving 2D similarity.

Turbocharged Pharmacophoric Similarity New!

PharmCast and PharmSim™. 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. A full comparison of two molecules takes 0.584 milliseconds against 5.7 seconds for the real calculation, and agrees with it at Pearson 0.969. One reference against a million candidates drops from 33 days to 5 minutes. In a retrieval test the true nearest neighbor is in the surrogate's top 100 for 95.3% of queries.

How it works →

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 Analysis New

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 →