Question 1

How is TuneLab designed to protect proprietary data when training the AI/ML models that use federated learning?

Accepted Answer

TuneLab federated training is hosted by a 3rd party.  During a training run, a global AI model is distributed from a central orchestration server to each participating company's node. Each node then trains this model on its proprietary data, with the raw data remaining in that node. Only model updates, such as mathematical weights and parameters, not the underlying molecular structures, are transmitted back to the central server. The server aggregates these updates from all participating nodes to improve the global model, which is then redistributed for the next training round. This iterative process continues until the model converges to optimal performance. To confirm data quality during this process, Tunelab reviews certain aggregated statistics and analytics derived from the data, together with information about assay conditions and protocols used to generate it.  This review is designed not to disclose, and does not require access to, any underlying molecular structures, sequences, or individual data points.

Question 2

What is Lilly TuneLab?

Accepted Answer

TuneLab is a first-of-its-kind, collaborative AI-powered drug-discovery and development platform that gives biotech researchers access to powerful AI models trained on decades of Lilly proprietary research data and member data contributions. These are the same models used by Lilly's own scientists every day.

Question 3

What kinds of models and use-cases does Lilly TuneLab support?

Accepted Answer

TuneLab currently supports preclinical drug-discovery programs, from first synthesis through to in vivo testing, with core models for small molecules and antibodies designed to work across different data sources and in federated environments.

Question 4

What training data does a biotech contribute in exchange for access?

Accepted Answer

To become a member of Lilly TuneLab, we are asking small molecule companies to contribute a minimum of 1,000 experimental readouts (an experimental readout is defined as one molecule observed in a wet lab experiment to measure a property). For antibody companies, we are asking a minimum of at least 1 measured property/endpoint across 125 sequences. A number of member companies contribute larger amounts of data than these minimums, with the aim of improving TuneLab’s AI/ML models further, and potentially increasing their performance in that specific member’s projects.

Question 5

Who is Lilly TuneLab designed for?

Accepted Answer

Biotechs with at least one preclinical small molecule or antibody program. Every molecule that advances too far before its liabilities costs time, capital, and can delay the development of new medicines for patients. Lilly TuneLab supports predictive characterization of antibody developability as well as small molecule ADMET profiles prior to in vitro or in vivo validation, with the aim that only the most viable candidates progress, and minimizing dead-ends in drug development.