ChemLab

A chemist optimizing a lead series wants the same four answers about every analogue: will it dissolve, will it cross a membrane, will the liver clear it in an hour, and will it light up a toxicity flag. The conventional in-silico answer is one model per assay, each fit to whichever compounds were run through it. But those readouts are not independent — the lipophilicity that sinks aqueous solubility is the same property that drives microsomal turnover — so a model-per-endpoint approach discards exactly the correlations a medicinal chemist reasons with. ChemLab takes a SMILES string and predicts 33 ADMET endpoints together in a single multi-task model, letting densely measured assays lend statistical strength to sparse ones.

ChemLab was built by Eli Lilly and Company and is distributed through Lilly TuneLab, the company's collaborative AI/ML drug discovery platform within Lilly Catalyze360. TuneLab launched in September 2025 with 18 predictive models trained on Lilly's internal drug disposition, safety, and preclinical datasets, valued at over $1 billion in research investment, and offered to biotech members in exchange for assay data of their own. Those first-generation models were single-task, an architecture Lilly's machine learning team describes as poorly suited to federated learning across partners whose assay panels differ. ChemLab and its antibody-developability counterpart AbLab were released in June 2026 as multi-task replacements, trained on the same Lilly data but rebuilt to absorb heterogeneous assay contributions; the original single-task models remain available alongside them.

This is a closed commercial model. There is no paper, preprint, code repository, public checkpoint, or model card, and no independent benchmark. Access requires TuneLab membership, conditioned on contributing preclinical assay data.

#Key Features

• One model across 33 ADMET endpoints: A single multi-task network covers solubility, permeability, microsomal stability, and toxicity signals simultaneously, learning shared chemistry rather than fitting each assay in isolation.
• Conformal uncertainty quantification: Predictions carry calibrated uncertainty from conformal methods, separating a confident call from an extrapolation past the chemistry the model has seen.
• Designed for heterogeneous contributions: The multi-task architecture accepts partner assay panels that only partially overlap Lilly's, making federated training across many small biotech datasets tractable.
• Fine-tuning and API inference: Members can fine-tune the model on their own chemistry and call it programmatically, so predictions sit inside an existing design-make-test loop rather than a separate portal.
• Available inside the electronic lab notebook: Benchling tenant admins install ChemLab from the Model Hub registry and run it against molecules already registered there, with predictions written back alongside the experimental record.

#Technical Details

ChemLab consumes SMILES and emits 33 small-molecule ADMET endpoints spanning absorption, distribution, metabolism, excretion, and toxicity, including aqueous solubility, permeability, and microsomal stability. Its training corpus is Lilly's proprietary drug disposition, safety, and preclinical data — more than 500,000 preclinical datapoints collected over more than 20 years, including in vivo and in vitro pharmacokinetic and toxicology measurements. Uncertainty quantification uses conformal prediction. Federated training runs over partner-contributed data in an environment operated by Rhino Federated Computing, with Lilly reporting more than 90 contributed small-molecule ADMET datasets in that pipeline; Tamarind Bio hosts the inference platform. Parameter count, molecular featurization, architecture family, and per-endpoint accuracy are undisclosed, and Lilly has published no benchmark results or technical report, so performance rests entirely on internal validation.

#Applications

The intended use is triage during hit-to-lead and lead optimization: score a virtual or newly synthesized series against the full ADMET panel, then send only the top-ranked molecules for testing, saving the weeks that solubility, permeability, and metabolic stability assays take. The audience is deliberately small and mid-size biotechs that have reached preclinical development but lack the historical assay data to train comparable models themselves. Delivery through Benchling puts predictions next to the registry entries they describe, so a chemistry team can prioritize compounds without exporting structures to a separate tool. Membership is priced in data rather than cash: a small-molecule company joining TuneLab contributes at least 1,000 experimental readouts — one molecule measured for one property — trained at the member's own node so that only weight updates, never structures, reach the aggregation server.

#Impact

ChemLab is the clearest test yet of whether a large pharmaceutical company's preclinical data advantage can be shared without being surrendered. TuneLab passed 100 member companies within its first year, and ChemLab is the model those members' contributions now train — a design where each participant's data improves a model every other participant uses. The arrangement is not symmetric: members gain predictions built on data they could never assemble, while Lilly gains a broadening view of chemical space and a platform positioned as a default step in early discovery. The caveats are that nothing about the model has been externally validated, and its accuracy on chemistry unlike Lilly's is unmeasured in public.