# Eidogen-Sertanty > Eidogen-Sertanty, Inc. builds curated structure activity data and the models > trained on it: the Kinase Knowledgebase, the Family, GPCR and Kinase Foundation > Models, Reverse Screen, PharmCast and ChIP, together with the Food Health and > Food Health TxD mobile apps for proactive wellness. Founded and led by Steven M. Muskal, > ORCID 0000-0002-3487-270X (https://orcid.org/0000-0002-3487-270X), whose > personal site is https://www.stevenmuskal.com/. ## Data - [Kinase Knowledgebase](https://eidogen-sertanty.com/kinasekbmarvin.php): licensed, hand curated kinase structure activity and synthesis data extracted from the published and patent literature, searchable by structure, target or keyword. Quarterly releases. - [Oncology Knowledgebase](https://eidogen-sertanty.com/oncologykbmarvin.php): the same curation applied to oncology targets. - [Dataset overlap analysis](https://eidogen-sertanty.com/overlap.php): compare a private compound set against a reference set without exposing either, using irreversible encoded fingerprints. ## Models and software - [PharmCast](https://pharmcast.ai/): predicts a 10,549-bit three-dimensional pharmacophore fingerprint from a SMILES string, so a pharmacophoric similarity runs without generating conformers. A full comparison of two molecules takes 0.584 milliseconds against 5.7 seconds for the explicit calculation. Method page: [how it works](https://eidogen-sertanty.com/pfp.php). Peer citation: 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). - [Foundation models, all three](https://eidogen-sertanty.com/foundationmodels.php): comparator models that rank rather than score. Give one a protein sequence and two compounds and it says which the protein prefers. Give the other one compound and two sequences and it says which target the compound prefers. One family spanning model and two family deep dives, in the order Family, GPCR, Kinase. Video overview of the approach: https://www.youtube.com/watch?v=VWMWP5hi3uA - [Family Foundation Model](https://familyfoundationmodel.com/): compound preference and target preference across 34 protein families, from kinases and GPCRs to proteases, ion channels and the epigenetic readers, writers and erasers. Held-out accuracy 0.71 over 65,725 compound preference comparisons; for target preference, 0.75 over 8,689 comparisons between two families and 0.78 over 32,738 within one family, rising to 0.92, 0.89 and 0.91 at strength 0.70 and above. The two kinds of target comparison are measured separately and never pooled. Trained on ChEMBL 37 alone, weights freely downloadable. Either model can be extended with a user's own measurements on their own machine. Code and extension tools: https://github.com/smuskal/ffm. Paper: Muskal, S. M. Two Comparators May Be All We Need. bioRxiv 2026, doi:10.64898/2026.09.15.751854 (https://doi.org/10.64898/2026.09.15.751854). Video overview: https://www.youtube.com/watch?v=VWMWP5hi3uA - [GPCR Foundation Model](https://gpcrfoundationmodel.com/): potency and selectivity rankings across human G protein-coupled receptors, built on ChEMBL together with Eidogen-curated data. Held-out accuracy 0.77 over 762,493 potency comparisons on 235 targets and 0.80 over 24,741 selectivity comparisons on 211 targets, rising to 0.94 and 0.93 at strength 0.70 and above. - [Kinase Foundation Model](https://kinasefoundationmodel.com/): potency and selectivity rankings across the kinome, trained on the Kinase Knowledgebase with ChEMBL used to test rather than to train. Held-out accuracy 0.69 over 1,836,100 potency comparisons on 477 targets and 0.75 over 3,137,588 selectivity comparisons, rising to 0.88 and 0.92 at strength 0.70 and above. Code: https://github.com/smuskal/KFM - [cHIA](https://eidogen-sertanty.com/chia.php): percent human intestinal absorption predicted from molecular structure. The six descriptors published in 1998 for 86 compounds rested on one division of those compounds; repeating the division 250 times puts 236 of 250 draws worse than the published cross-validation error and 249 worse than the published external error, with none matching both. Those descriptors still reach a correlation of 0.63 under a modern learner on nine times the data, and the properties they encoded, charge, polar surface and hydrogen bond donors, are what a free search over 1,301 descriptors selects today. A genetic algorithm searches descriptor sets of variable length, scored by the out-of-bag error of a random forest inside each extraction's training group, and the descriptors recurring across 250 extractions form one fixed consensus set, which is evaluated the same way. The deployed absorption model, 25 consensus descriptors and a random forest on 780 compounds, gives 20.47 %HIA units rms at R 0.76 and places a compound on the correct side of the 30 percent threshold 92.60 percent of the time. The same procedure reaches permeability, reported as log10 Papp: Caco-2 small molecules, 3,771 compounds at 0.35 rms and R 0.84; Caco-2 cyclic peptides, 1,281 peptides at 0.49 rms and R 0.84; and cyclic peptides across an artificial membrane, 7,298 peptides at 0.82 rms and R 0.67. The two Caco-2 models predict the same quantity on different chemistry and are not interchangeable. Every prediction carries the distance from the query to its nearest training compound. Predict a structure at https://predicthia.ai. The original model: Wessel, M. D.; Jurs, P. C.; Tolan, J. W.; Muskal, S. M. J Chem Inf Comput Sci 1998, 38, 726-735 (https://doi.org/10.1021/ci980029a). The current work: Muskal, S. M.; Wessel, M. D. Human Intestinal Absorption Revisited: The 1998 Descriptors Survive, and the Method Extends to Permeability. bioRxiv 2026 (https://doi.org/10.64898/2026.10.03.756481). - [ToxPred](https://eidogen-sertanty.com/toxpred.php): toxicity by consortium. A query molecule is compared to a public reference set of measured compounds, and those similar enough vote on its toxicity; when none is close enough the method declines rather than guessing. 92.1% of answers land within one log unit of the measured rat oral LD50, on the 1,336 of 2,873 held-out compounds it answered for, and the same method extends unchanged to 28 further endpoints. A second signal counts the distinct protein families a compound's structural matches span across 27,910 indexed ligands: matches reaching no family inhibit CYP3A4 2.5 percent of the time, matches reaching sixteen families or more inhibit it 51.0 percent of the time. Score a molecule at https://toxpred.ai. Code: https://github.com/smuskal/toxpred. The method was first described in Muskal, S. M.; Jha, S. K.; Kishore, M. P.; Tyagi, P. J Chem Inf Comput Sci 2003, 43, 1673-1678 (https://doi.org/10.1021/ci034080c). The current work rebuilds it on the public record: Muskal, S. M.; Jha, S. K.; Kishore, P.; Tyagi, P. Toxicity by Consortium Revisited: Structural Neighbors and Cross-Family Reach. bioRxiv 2026 (https://doi.org/10.64898/2026.10.04.756566). - [ChIP](https://eidogen-sertanty.com/chip2026.php): de novo design that searches reaction space rather than molecule space, so a designed molecule comes with a synthetic route. - [Reverse Screen](https://reversescreen.ai/): given a molecule, the protein targets it might interact with. Every co-crystal ligand in the Protein Data Bank, 27,797 of them across 28,579 target sites, is indexed by the three-dimensional pharmacophore fingerprint PharmCast predicts from flat structure. A query is fingerprinted in 4 milliseconds and compared against the whole index in 40, and AutoDock Vina then docks only into the sites retrieval returns, against 40.5 hours to dock every characterized site. Across 3,000 held-out molecules the five nearest indexed ligands give 21.1 candidate proteins and contain the molecule's own known target 48.8 percent of the time. This is the question that arises once a compound exists, and the only question available for a molecule a generative method proposes. 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 overview: https://www.youtube.com/watch?v=mFRBfK6iv_w ## Structural evidence - [Contact Atlas](https://eidogen-sertanty.com/contactatlas.php): every small molecule co-crystallized with a protein in the Protein Data Bank, searchable by the structure you draw. Four modes: substructure, an exhaustive graph match over the whole index that is never sampled and never truncated and also takes a SMARTS pattern; exact structure, matched on the InChIKey connectivity layer so a stereoisomer still finds it; pharmacophore, using the same predicted three-point PharmCast fingerprint Reverse Screen ranks with, which is the mode that finds a scaffold hop between two molecules sharing no common substructure; and by name, where component code, generic name, trade name, research code and ChEMBL id all resolve to the same ligand. Every hit opens on its contact map: the ligand drawn flat as deposited, ringed by the residues it touches, with hydrogen bonds, salt bridges, pi stacking, cation to pi, metal coordination and bridging waters labelled by residue and measured from the deposited coordinates, and the potency ChEMBL records for that compound against that target. Nothing is docked and nothing is predicted. The index is rebuilt weekly and the search states the archive snapshot it answers from. The search itself runs at https://reversescreen.ai/atlas.html, free and without an account. - [Kinase contact maps](https://eidogen-sertanty.com/ligsite.php): what each Kinase Knowledgebase compound is doing in its pocket, measured from the deposited coordinates, for every compound with a matching Protein Data Bank structure. - [Bound is not necessarily potent](https://eidogen-sertanty.com/bound-is-not-necessarily-potent.php): fourteen protein sites with every co-crystal ligand superposed and its measured potency, which spans up to six orders of magnitude inside one pocket. - [Reverse peptide mimetics](https://eidogen-sertanty.com/reversePeptideMimetic.php): short peptides that present the same pharmacophore as ligands observed bound in co-crystal structures. - [Loops that mimic drugs](https://eidogen-sertanty.com/loopMimeticDrugs.php): the peptides above that occur as complete loops in real protein structures, shown in the conformation the loop actually adopts. ## Proactive wellness - [Food Health](https://foodhealthscan.com/): mobile app for iPhone and Android. A meal photo or a typed note becomes calories, protein, carbs, fat and fiber in seconds; the app syncs activity, sleep and workouts from Apple Health, captures readings from health devices, keeps entries offline until a connection returns, and shows nutrition reports and scatter plots over time. Data stays on the device by default. App Store: https://apps.apple.com/app/id6759629101. Google Play: https://play.google.com/store/apps/details?id=com.foodhealthai.android - [Food Health TxD](https://foodhealthscan.com/txd/): spot the meals that spike you. The same carb estimation overlaid on continuous glucose readings on one timeline, with a menu scanner for the lowest-carb choice, built for type 1, type 2 and prediabetes. Reads glucose from Apple Health and writes carbs back. For education and tracking only, not a medical device. App Store: https://apps.apple.com/us/app/food-health-txd/id6766341399. Google Play: https://play.google.com/store/apps/details?id=com.foodhealth.txd - The thinking behind both, by Steven M. Muskal: Your Kitchen Is a Pharmacy. Here's the Proof (https://medium.com/@smuskal/your-kitchen-is-a-pharmacy-heres-the-proof-9631b555cf2e); The Metabolic Crisis Is Not a News Story (https://medium.com/@smuskal/the-metabolic-crisis-is-not-a-news-story-3ea7ffada493); Food Is Medicine. But First It Has to Become Data (https://medium.com/@smuskal/food-is-medicine-but-first-it-has-to-become-data-663c2ee26a98). ## Company - [Products and services](https://eidogen-sertanty.com/products.php) - [About](https://eidogen-sertanty.com/about.php) - [Publications and presentations](https://eidogen-sertanty.com/publications.php) - [Contact](https://eidogen-sertanty.com/contact-information.php) ## The people behind the work - [Steven M. Muskal, Ph.D.](https://www.stevenmuskal.com/): founder and Chief Executive Officer, principal author of the foundation models, Reverse Screen and PharmCast. His site carries a project card for each, with the videos and papers: https://www.stevenmuskal.com/projects.html. His publications, including the bioRxiv papers: https://www.stevenmuskal.com/publications.html. His own llms.txt: https://www.stevenmuskal.com/llms.txt - PharmCast is joint work with Malcolm J. McGregor. Reverse Screen is joint work with George Nicola. ## Boundaries Knowledgebase contents are available under license, and the customer support and download areas are gated by that license. Pages here describe what the data covers and how to evaluate it. Model predictions are predictions, and a pharmacophoric similarity indicates agreement of shape and features rather than measured biological activity.