Data-driven federated learning in drug discovery with knowledge distillation

Vandenbruaene Nathalie — Tue, 18 Mar 2025 12:00:52 +0000

Yogesh Sabnis, Global CADD and Annie Delaunois, Non clinical Safety Evaluation

Vandenbruaene … Tue 18/03/2025 - 13:00

We’re proud to announce the publication of joint research on FLuID (Federated Learning Using Information Distillation) in Nature Machine Intelligence. This innovative approach has the potential to reshape how industries like pharmaceuticals collaborate while safeguarding sensitive data.

A key challenge for AI in scientific research is accessing high-quality data for impactful models. Valuable knowledge often remains locked in confidential corporate data silos, despite industries being more open to sharing non-competitive insights. Federated learning allows knowledge sharing while preserving data privacy but has limitations.

In the publication we introduce FLuID (federated learning using information distillation) tailored to drug discovery to maintain data privacy. Validated through public data and real-world collaboration among eight pharmaceutical companies, FLuID addresses domain shift challenges and enhances knowledge sharing. This leads to improved models for biological activity predictions, paving the way for a new generation of models with better performance and broader applicability.

Here’s how it works. Instead of sharing raw data, companies train private, local models and use them to annotate a shared public dataset. These annotations are then combined, creating a powerful blend of insights that organizations can leverage collaboratively. The process ensures complete privacy while producing models that outperform those built from individual datasets.

FLuID has already demonstrated its impact. By collaborating, multiple pharmaceutical companies have improved their ability to predict how chemical compounds interact with the human body, helping drive innovation and support drug safety predictions.

With its privacy-first design and scalable framework, we hope FLuID opens the door to ethical, large-scale collaboration across industries, paving the way for smarter, faster discoveries in fields where data security has traditionally been a barrier.

I invite you to explore this achievement and learn how FLuID is setting new standards for innovation in science and beyond by reading the full publication here.

innovation

55 Likes

Nature Machine Intelligence

Data-driven federated learning in drug discovery with knowledge distillation

Leave a Comment