About

What we do

An EGFR structure Eidogen-Sertanty has built drug discovery data and software since 2003. The work has always come down to one question. Someone has to decide which molecule to make next, and that decision is only as good as the data behind it and the models trained on that data. More recently the same thinking has gone into proactive wellness, in the two Food Health mobile apps described below.

We have worked with the top biopharmaceutical companies over the last quarter century, on cancer, infection, arthritis and pain. The knowledgebases we curate by hand are still the foundation. Everything we have built since sits on top of them.

The data

The Kinase Knowledgebase holds 3,336,903 measured bioactivity points across 579 kinase targets, read out of the published and patent literature by people rather than by a parser, and released every quarter. The Oncology Knowledgebase applies the same curation to oncology targets. Hand curation is slow and expensive. We have never found a substitute that holds up.

Every release ships with models built from it, so the data arrives with a measured baseline rather than a promise. You can also see what the training corpora contain before you trust anything trained on them.

The models

Our foundation models answer comparative questions, because that is the question a project actually asks. Which of these two compounds should I make first, and which of these two proteins will it hit. Each model takes one row, and the order of that row is the question. Ligand, sequence, ligand asks which compound that protein prefers. Sequence, ligand, sequence asks which protein prefers that compound. We call them LSL and SLS. Nothing is docked, no structure is needed, and no binding site has to be defined.

A model returns a ranking and a prediction strength, not an absolute IC50. The strength is a ranking signal whose reliability we measured on held-out data, so you can choose the cutoff you are willing to act at. The name does not mean a language model.

There are three, and they differ in what they cover and in the data behind them. Two are family deep dives and one is the broad instrument.

The Family Foundation Model is the broad instrument: two separately fitted models over one roster of 34 protein families, from kinases and GPCRs to proteases, transporters, ion channels and the epigenetic readers, writers and erasers. It is trained on ChEMBL 37 alone, which is why both sets of weights are freely downloadable. The compound-preference model is right 0.71 of the time over 65,725 held-out comparisons at any of 2,079 targets, and the target-preference model 0.75 over 8,689 held-out comparisons between two families, and 0.78 over 32,738 within one family, at any of 2,279.

The Kinase Foundation Model is the deep dive with the deepest data behind it, trained on the Kinase Knowledgebase with ChEMBL used to cross-validate rather than to train. On measurements it never saw it is right 0.69 of the time on potency comparisons and 0.75 on selectivity, and acting only on the confident calls raises those to 0.88 and 0.92.

The GPCR Foundation Model applies the same two layouts to G protein-coupled receptors, trained on ChEMBL together with Eidogen-curated data. It scores 0.77 on potency and 0.80 on selectivity, rising to 0.94 and 0.93 on the confident calls, and every target with held-out data carries its own published accuracy. Folding our own curated set into the September 2026 rebuild, and holding out compounds rather than dates, moved those figures up sharply from 0.64 and 0.69, which is the clearest argument we have for why hand curation still matters.

PharmCast predicts a full three dimensional pharmacophore fingerprint straight from a flat structure, with no conformer generation at any point. A complete comparison of two molecules takes 0.584 milliseconds against 5.7 seconds for the explicit calculation, which is what makes the comparison usable across a whole screening collection instead of a shortlist. The method is published: Muskal and McGregor, bioRxiv 2026, doi: 10.64898/2026.09.02.748999.

Design you can actually order

ChIP searches reaction space rather than molecule space. It evolves synthetic protocols, 85 validated reaction transforms drawing on 853,409 cataloged building blocks across 154 reactant slots, and scores the molecules those protocols produce with your own predictive model. Every design that comes out carries an executable synthesis from purchasable starting materials. A design you cannot make is not a design. A worked campaign is here, in full, including the routes.

Reverse Screen

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 it is the only question available for a molecule a generative method has just proposed, because a designed structure has no measured target at all.

Docking one compound into every characterized site answers it directly and takes 40.5 hours. Instead every co-crystal ligand in the Protein Data Bank, 27,797 of them, is indexed by the pharmacophore fingerprint PharmCast predicts from flat structure. A query is fingerprinted in 4 milliseconds and compared against the whole index in 40, and the docking then runs only on what retrieval returns.

What the structures show

We publish the structural evidence rather than describing it. Kinase contact maps draw what each knowledgebase compound is doing in its pocket, measured from the deposited coordinates. Bound is not necessarily potent puts every co-crystal ligand of a protein in one frame with its measured potency, and inside a single pocket those values span up to a millionfold. That figure is the reason we do not treat a structure as an answer.

The same fingerprints run backwards as well. Reverse peptide mimetics finds short peptides that present the same pharmacophore as a bound ligand, and loops that mimic drugs shows the ones that occur as real loops in real protein structures, in the conformation the loop actually adopts.

Proactive wellness

Drug discovery, data and software are most of what we do. Proactive wellness is the rest, and it rests on the same idea: a decision is only as good as the data behind it, and a meal is a decision. Food Health is a mobile app for iPhone and Android that turns a meal photo or a typed note into calories, protein, carbs, fat and fiber in seconds, syncs with Apple Health, and shows the trends over time. Food Health TxD pairs the same carb estimation with continuous glucose readings on one timeline, so someone with type 1, type 2 or prediabetes can spot the meals that spike them. Both keep data on the device by default. Neither is a medical device.

Working together

Our DirectDesign™ service puts these technologies on one project at a time, for directed de novo design of lead compounds and target focused libraries. Before any of that, dataset overlap analysis tells you how much of our collection you already have and how much is genuinely new to you, using encoded fingerprints so neither side exposes its structures. We would rather you knew that before you licensed anything.

Talk to us if you want to put any of it on your own problem, or start with the products page to see the whole set.