EU-backed healthcare project rolls out platform for federated learning in drug discovery

The three-year project MELLODDY, which had received €18-million from the European Union and partnering EFPIA companies under the Innovative Medicines Initiative (IMI2), has now built a secure platform for privacy-preserving and federated learning.

This AI framework avoids the need for competitive data and models to ever leave the owner’s custody while still allowing collaborative machine learning that can train and evaluate drug discovery predictive models. The results show that its operation at scale yields improvements across all pharmaceutical partners in the predictive performance of collaboratively trained models over single partner models.

In its three years, the platform trained models on billions of industrial experimental data points documenting the behaviour of more than 20 million chemical small molecules in over 40,000 biological assays.

Several pharma companies contributed to the development of this project by providing training data for the global model and evaluating if the global model performed better than the one built solely on their data. The consortium members included 10 pharmaceutical companies including Amgen; Astellas; AstraZeneca; Bayer; Boehringer Ingelheim; GSK; Institut de Recherches Servier; Janssen Pharmaceutica NV; Merck KGaA; and Novartis, five technology partners including Iktos; Kubermatic; NVIDIA; Owkin; and Substra Foundation and two academic partners Budapest University of Technology and Economics and KU Leuven.

According to Mathieu Galtier, project coordinator and chief data and platform officer at Owkin, it is a massive win for drug discovery and ultimately patients.

“For years, it was assumed that competing pharmaceutical companies could never work together at scale to discover the next generation of drugs. But our three-year project showed that it is not only possible – it is also more effective than going at it alone. Crucially, we demonstrated that federated learning protects commercially sensitive data, meaning that no one loses out when collaborating.”

Hugo Ceulemans, project lead and senior scientific director of drug discovery data sciences at Janssen Pharmaceutica NV, added: “Throughout the pharmaceutical industry, increasingly powerful machine learning approaches are leveraging ever more data and insights to better focus and accelerate the real-world experiments and studies that bring life-saving drugs to patients.”

Going forward, the technology developed as part of the project will be extended to new fields in healthcare to facilitate privacy-preserving collaboration in AI. While the consortium focused on the domain of small molecule drug discovery, its approach has the potential to benefit other areas in the development pipeline, including areas of interest such as biologics, histology, and genomics.