Products
Find the right model or tool for your drug discovery work.
AI/ML models and tools available now on TuneLab
AbLab v1, v2
Predicts antibody developability across 22 endpoints, including stability, aggregation, and solubility, from sequence alone.
Antibody-Antigen Binding Affinity (ΔΔG)
TuneLab’s Antibody–Antigen Binding Affinity model predicts how a specific mutation changes binding affinity between an antibody and its antigen.
ChemLab v1, v2 Federated
Predicts ADMET and physicochemical properties to help prioritize compounds for synthesis.
Immunogenicity Model (MHC-II Peptide Presentation)
Predicts MHC class II presentation likelihood and 9-mer binding cores for antibody peptides across 12 HLA-DRB1 alleles.
in vivo PK (developed by insitro)
Predicts preclinical in vivo PK, including clearance and Vss, across multi species, developed by insitro for Lilly.
NanoLab (VHH Developability)
Predicts five biophysical developability endpoints for VHH nanobody sequences to rank leads before wet lab testing.
Small Molecule Computational Workflows
Configurable workflows that generate, filter, score, predict properties and rank small molecules to automate drug discovery campaigns.
ChemLab is Lilly TuneLab's multi-task small molecule property prediction model, covering physicochemical and ADMET endpoints including solubility, lipophilicity, permeability, protein binding, metabolic stability, transporter activity and safety-related liabilities.
You will see ChemLab v1 and ChemLab v2 Federated. ChemLab version 1 was released in June and predicts 33 ADMET endpoints from molecular SMILES using a Chemprop message-passing neural network (MPNN) with task embeddings. The newly released ChemLab v2 Federated is trained through federated learning across Lilly's internal assay collections and contributed TuneLab member data, giving it a wider applicability domain than the previous releases. It is built for scientists prioritizing which compounds to make.
Predictions across a large panel of physicochemical and ADMET endpoints from a single molecular structure input.
A substantially broader applicability domain than earlier versions. Roughly a third of the training signal now comes from member contributions spanning many organizations and chemotypes, improving reliability on chemistry outside Lilly's historical space.
Explainability indicates which parts of a molecule drive a property prediction. The module can be operated at either the fragment or atom level. For fragment mode, the molecule is fragmented, then attributions are assigned at the fragment level. For atom mode, attributions are assigned at the atom level.
Improved uncertainty quantification, so low-confidence predictions on out-of-domain chemistry can be identified and handled differently from confident ones.
Adaptive multi-task sharing. Each assay is preserved as its own prediction task with its own head, learning what information is useful across assays while preserving the distinctions that make each endpoint unique. This strengthens predictions through relevant shared evidence for endpoints.
Fine-tuning is available for ChemLab. Fine-tuning lets a member adapt the model to their own project. It builds on the shared learnings from the federated datasets, so a member can get useful predictions from a small dataset of their own. Fine-tuned models are private to the member.