AbLab

An antibody can bind its target at picomolar affinity and still be undevelopable. It self-associates at the concentrations a subcutaneous formulation demands, or the solution turns too viscous to push through a needle, or it unfolds a few degrees below where the manufacturing process needs it to hold. These properties are read out by physical assays — AC-SINS for self-association, size-exclusion chromatography for aggregation, nanoDSF for thermal stability — every one of which needs expressed and purified protein. The answers therefore arrive after the expression campaign, often after months of engineering are already committed to a lead.

AbLab moves that triage upstream to the sequence. Built by Eli Lilly and Company, it takes an antibody's heavy- and light-chain sequences and returns predictions for 18 developability endpoints — thermal stability, aggregation, viscosity, solubility and others — in a single pass. The endpoints are predicted jointly rather than by a separate model per assay, which shares statistical strength across assays measuring correlated physical behavior and, critically, lets the model absorb a contribution covering only one endpoint. That is what distinguishes AbLab from the antibody models Lilly first released: those were single-task, and were replaced precisely because they suited learning from partners' partial assay panels poorly.

The measurements that anchor developability are the scarce ingredient. Public antibody language models such as AbLang and AntiBERTa2 learn from repertoire sequences, which carry no biophysical labels; the labels live in pharmaceutical companies' internal assay archives. AbLab is trained on Lilly's, and reaches outside users through Lilly TuneLab — a federated learning platform launched in September 2025 where biotechs gain access to Lilly's discovery models in exchange for their own assay data. AbLab arrived with TuneLab 2.0 in June 2026, and became installable from Benchling's Model Hub registry in August 2026.

#Key Features

• Sequence-only input: heavy and light chain sequences in, developability out — no structure, no expressed material, no purification.
• Eighteen endpoints from one model: thermal stability, aggregation, viscosity, solubility and further biophysical readouts are predicted simultaneously, not by an assay-specific model each.
• Designed for heterogeneous contributions: the multi-task architecture accepts partners whose data covers only part of the endpoint panel, which makes federated training across many biotechs workable.
• Calibrated uncertainty: conformal methods attach upper and lower confidence bounds to every prediction, so a weakly supported call can be routed to the bench rather than acted on.
• Adaptable and callable: the model can be fine-tuned to a project's own sequence space, and is reachable through an API or from Benchling's Notebook and Registry.

#Technical Details

Lilly describes AbLab as a multi-task foundation model trained on its proprietary antibody data and architecturally rebuilt from the first-generation TuneLab models for the federated setting. Lilly puts the platform's aggregate corpus at more than 500,000 preclinical datapoints collected over more than 20 years, spanning in vivo and in vitro pharmacokinetic and toxicology measurements across small molecules and antibody-based therapeutics, and values that research investment at over $1 billion. Training is on monoclonal antibody data, with additional formats in progress and VHH models announced as forthcoming. Beyond the multi-task framing, the architecture, parameter count and benchmark scores are undisclosed, and no paper, code or weights accompany the model. Tamarind Bio operates the inference infrastructure; Rhino Federated Computing runs the environment in which partner contributions update the shared model.

#Applications

The intended use is candidate triage during antibody lead optimization. A team with a panel of binders can rank them by predicted liability before committing expression, purification and assay capacity, sending only the top-ranked sequences for wet-lab confirmation. Because developability gates the move from discovery into lead optimization, cell line development and formulation, catching a liability at the sequence stage removes a full make-and-measure cycle. The clearest beneficiaries are small and mid-size biotechs, which rarely hold a developability archive deep enough to train a comparable model. Membership is priced in data rather than cash: an antibody company joining TuneLab contributes at least one measured endpoint across a minimum of 125 sequences, trained at the member's own node so that only weight updates, never sequences, reach the aggregation server.

#Impact

AbLab's significance is as much structural as technical: it is one of the first production antibody models a large pharmaceutical company has opened to outside biotechs, trained on exactly the proprietary assay data historically unavailable for modeling. TuneLab passed 100 member companies within its first year, and the Benchling integration puts AbLab within reach of more than 1,300 biotechnology companies. The caveats are substantial and worth stating plainly. There is no publication, no public benchmark and no released weights or code, so reported accuracy rests on Lilly's own statements and cannot be independently reproduced. Access is conditional on TuneLab membership, making the model a commercial arrangement rather than a released artifact.