Multimodal transformer for drug discovery that predicts human pharmacokinetics by transferring signal from abundant preclinical data.
No providers recorded yet. Browse all providers
A drug discovery program generates laboratory data on thousands of molecules, but only a handful of those molecules ever reach human volunteers, so the properties that decide clinical success, such as how long a drug persists in the body, are measured for almost nothing. A model trained only on human pharmacokinetic (PK) data has a few hundred compounds to learn from across the whole public literature. Enchant attacks this "data wall" by training one transformer on many kinds of discovery-stage data at once, so that a scarce clinical endpoint can borrow signal from abundant preclinical measurements, often made on unrelated molecules.
Enchant was developed by Iambic Therapeutics, a clinical-stage biotech in San Diego, and announced on 29 October 2024 in a company whitepaper and press release. Iambic began the project about a year before launch, and the model completed training in early 2024. There is no peer-reviewed paper or preprint. Iambic later released Enchant v2, a successor trained at more than ten times the scale, and this entry covers the original model only.
Enchant is an internal platform model. Iambic uses it to screen out molecules with poor predicted clinical properties before they are nominated as candidates, and it has released no code, weights, training data or public API.
Enchant is a transformer trained on dozens of data sources and modalities, which trade press reporting of the launch lists as including PubMed, bioRxiv, the Protein Data Bank and Iambic's own laboratory data. Iambic has not disclosed the parameter count, the tokenization scheme, the size of the training corpus or the training objective. The whitepaper's footnote states that the work primarily addresses small molecules.
The main evaluation uses the Obach human PK dataset. Iambic ordered the preclinical data the way it is usually gathered in a discovery program, from in vitro assays to mouse PK to higher-species PK, and trained on successively more of it alongside 1%, 5% or all of the Obach clinical data. With 1% and 5% of the human data, Enchant became strongly predictive of half-life on held-out molecules through transfer from preclinical data, and similar trends held for the other PK parameters Obach reports. Trained on all of the preclinical data plus the full Obach set, Enchant reached a Spearman correlation of 0.74 for human half-life, against 0.58 for the best prior published model, a CFA model from Nan Jiang and colleagues. A later benchmarking exercise, published in Iambic's May 2025 Enchant v2 post rather than in the v1 whitepaper, put Enchant at a median per-kinase Pearson R of 0.61 on the DGBC Kinase200 benchmark, against 0.55 for a fine-tuned TxGemma 9B. All of these results are self-reported by the developer and have not been independently reproduced.
Enchant is aimed at medicinal chemists and program leads who must decide which compounds to advance long before any human data exists. Iambic applies it to predict PK properties such as half-life, along with properties such as solubility, when selecting preclinical drug candidates. The goal is to catch liabilities that would otherwise surface only in clinical trials, saving time and cost and reducing the burden on trial participants. Because the model is proprietary, outside groups can reach it only through partnerships with Iambic.
Enchant was one of the first industrial attempts to treat drug discovery data as a single multimodal pretraining corpus rather than a collection of separate property-prediction tasks, each with its own small model. Its central claim, that more preclinical data makes clinical predictions better, speaks directly to the translational bottleneck behind late-stage clinical failures. Outside researchers can weigh that claim only through Iambic's own publications, the launch whitepaper and the later posts that re-benchmark the model. The evidence covers mainly one public PK dataset, and trade press noted at launch that the examples may have been selected to show the model at its best. Enchant established the design that Enchant v2 scaled up, and Iambic has continued the family with further releases.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.