Proprietary multimodal transformer for small-molecule drug discovery, predicting ADME, PK, potency and clinical endpoints from sparse assay data.
No providers recorded yet. Browse all providers
A new drug discovery program starts with almost no data about the properties that decide its fate. A team may have measured cellular potency for a dozen compounds and human pharmacokinetics for none, and a model trained only on those few points has nothing to learn from. Enchant v2 addresses this by pretraining one transformer on a large, heterogeneous corpus of drug-discovery data, so that a new endpoint borrows signal from everything the model already knows about related chemistry, assays and biology.
Enchant v2 was developed by Iambic Therapeutics, a clinical-stage biotech in San Diego, and announced on 5 May 2025 in a technical post, with training completed in the first quarter of 2025. It succeeds Enchant v1, which finished training in early 2024, and was trained with more than ten times the compute. Iambic presented the work as the poster "Scaling Multi-Modal and Multi-Task Transformers for Small Molecule Drug Discovery" at the NeurIPS 2025 AI4Science workshop. No peer-reviewed paper or preprint describes the model.
Enchant v2 is an internal platform model. Iambic uses it across its own programs and partnerships and has released no code, weights or API.
Iambic unified and curated large public databases, including PubChem, ChEMBL and the Protein Data Bank, with its own internal data spanning chemical, biological, preclinical and clinical contexts. An isoFLOP study, sweeping model size against training tokens at fixed compute, set the size of Enchant v2. Neither the parameter count nor the size of the training corpus has been disclosed. In the scaling study, fine-tuned Enchant v2 reaches Pearson R 0.78 and a mean absolute error of 0.43, against 0.71 and 0.46 for Enchant v1.
Iambic compared Enchant v2 with Enchant v1 and with TxGemma 9B, all fine-tuned on identical train/test splits. Enchant v2 led on every endpoint of the Biogen ADME benchmark: human and rat liver microsomal clearance, human and rat plasma protein binding, MDR1-MDCK efflux ratio and solubility. On the DGBC Kinase200 set it reached a median per-kinase Pearson R of 0.69, against 0.61 for Enchant v1 and 0.55 for TxGemma. The same ordering held on internal Iambic assays for permeability, CDK4 inhibition, MCF-7 cellular pharmacodynamics and solubility. In one example, the base model had seen MCF-7 data for only 12 molecules and still predicted held-out compounds zero-shot at Pearson R 0.45. Trained on the Obach human pharmacokinetics dataset, the Enchant family reached a Spearman R of 0.74 for human half-life. These results are self-reported by the developer and have not been independently reproduced.
Iambic applies Enchant v2 to multi-parameter optimization of drug candidates and to ranking discovery programs by their predicted endpoints, and it describes using the model to inform clinical trial design. In one program, compounds' in vitro clearance did not predict their clearance in rats. Enchant v2, trained jointly on the plentiful in vitro data and a small set of in vivo pharmacokinetic measurements, separated high- from low-clearance compounds with an AUROC of 0.74. Extrapolating from rat microsomes gave an AUROC of 0.51, no better than chance. In hit finding, 10 of the model's 12 top-ranked compounds were confirmed active in the assay.
Enchant v2 is an industrial test of whether transformer scaling laws carry over to molecular property prediction, a field long dominated by small models trained one assay at a time. Its zero-shot results bear on a practical bottleneck: getting predictive signal for a new assay before enough compounds have been made to train a dedicated model. Outside researchers can only weigh that evidence from Iambic's own reports. The model is proprietary, the evidence is a company post and a workshop poster, and the internal benchmarks rest on data that is not public. Enchant v3 succeeds it as the next generation of the family.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.