Publications & Presentations

A compound's toxicity can be gauged from the toxicities of the compounds it resembles. We showed that in 2003 for rat oral LD50, on a reference set that is no longer public. Rebuilt on 11,989 curated public values, the consortium method answers for 67.0 percent of the held-out compounds carrying an exact value and 92.1 percent of those answers land within one log unit. Run unchanged over 28 further endpoints, 149,224 measurements on 34,162 compounds, it correctly identifies a median 72.9 percent of the compounds that showed a toxicity and 96.5 percent of those that did not. A second signal, cross-family reach, counts the protein families a compound's matches belong to among 27,910 ligands solved in protein structures: a compound reaching no family inhibits CYP3A4 2.5 percent of the time, one reaching sixteen or more 51.2 percent, and the same pattern holds in 2,335 OpenADMET compounds measured since. Scoring needs only a two-dimensional structure, with no docking and no conformer generation at query time.

In 1998 we predicted percent human intestinal absorption for 86 compounds from six descriptors chosen by a genetic algorithm and a small neural network, on one division of the compounds. Repeating that division 250 times shows it was a fortunate one: 14 extractions match the published cross-validation error and only one matches the published external error, so every experiment here is repeated 250 times and reported as a distribution. The same six descriptors, carried unchanged onto 780 public absorption measurements, still hold up, and the properties they measure, charge, polar surface and hydrogen bond donors, are what a free search over 1,301 modern descriptors selects today. The procedure then extends, with nothing retuned, to Caco-2 permeability and to cyclic peptide permeability, with each deployed model reporting an applicability flag from its distance to the nearest training compound.

Two separately fitted comparator models over one roster spanning 34 protein families, from kinases and GPCRs to proteases, ion channels and the epigenetic readers, writers and erasers. The compound preference model holds one target fixed and ranks two compounds; the target preference model holds one compound fixed and ranks two targets, from different families or from one family, with each kind measured on its own held-out comparisons. Every answer is an ordering with a prediction strength from 0.5 to 1.0, and accuracy climbs with strength: 0.71 over 65,725 held-out compound preference comparisons, and 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. Trained on ChEMBL 37 alone, so both sets of weights are freely downloadable, and either model can be extended with a user's own measurements on their own machine, including targets added from an amino acid sequence.

Given one molecule, which proteins does it bind? Docking it into every characterized binding site answers directly and takes 40.5 hours on twenty cores. Instead every co-crystal ligand in the Protein Data Bank is indexed by the three-dimensional pharmacophore fingerprint it presents, predicted from flat structure by PharmCast, together with the proteins it was solved against. A query is fingerprinted in 4 milliseconds and searched against all 27,797 indexed ligands, covering 28,579 target sites, in 40 milliseconds, and only the proteins its nearest neighbors were crystallized with are docked. Across 3,000 held-out molecules, pooling the five most similar indexed ligands gives 21.1 candidate proteins and contains the known target 48.8 percent of the time at 108 seconds of docking; pooling twenty-five gives 104.5 proteins and 60.8 percent. Run cold, orforglipron returned the GLP-1 receptor ranked first and daraxonrasib returned its KRAS and cyclophilin A tri-complex fourth.

In the reference pipeline, generating 100 conformers requires 2.82 s of the 2.86 s needed to fingerprint one catalog compound; the bit calculation requires 0.039 s. We therefore removed the conformational stage. PharmCast is a feedforward neural network that predicts all 10,549 bits of a PharmPrint ensemble fingerprint directly from a SMILES string, comparing two molecules in 0.584 ms where the conventional pipeline takes 5.71 s. Version 10 was trained on 5,887,229 molecules from a screening collection, activity-backed ChEMBL compounds and peptide loops excised from crystal structures. Evaluated on 155,648 purchasable screening collection compounds excluded from every training set, 139,700 held-out ChEMBL compounds and 13,500 reserved peptide loops, median fingerprint error, Pearson r and pairwise ranking accuracy are 0.008, 0.980 and 0.936 for screening collection chemistry; 0.016, 0.984 and 0.952 for loop peptides; and 0.027, 0.936 and 0.889 for activity-backed ChEMBL compounds, against the reference calculation's own reproducibility of 0.006 and 0.995.

The consortium method: a compound's toxicity is predicted from the measured compounds it structurally resembles. Rebuilt in 2026 on the public record as ToxPred, it lands within one log unit of the measured rat oral LD50 92.1 percent of the time on the 1,336 of 1,994 held-out compounds carrying an exact measured value that it answered for, against 94 percent reported here. The rebuild is described in Toxicity by Consortium Revisited, bioRxiv 2026, doi 10.64898/2026.10.04.756566.