In economics, the prisoner’s dilemma describes a paradox where all players in the game would benefit from collaboration. Yet, they decide not to do so and settle for a suboptimal outcome due to a lack of trust in each other - ultimately leading to a worse outcome for everyone.
This was exactly the challenge pharmaceutical companies faced when developing machine learning (ML) models for drug discovery. Machine Learning (ML) and Artificial Intelligence (AI) models require large amounts of data to be trained on - and their accuracy increases as more high-quality datasets are fed into the model. If pharma companies could combine their datasets, they could build more powerful predictive models, leading to faster scientific breakthroughs in drug development.
However, they faced two major obstacles: data privacy and competition concerns. Not only were companies wary of sharing their proprietary data with competitors, but there were also potential regulatory challenges involved. As a result, they developed their models in isolation.
The solution: MELLODDY
The MELODDY project (Machine Learning Ledger Orchestration for Drug Discovery) was created to tackle these challenges. As a European-funded project, MELLODDY brought together 10 leading pharma companies - including Bayer, GSK, and Novartis - alongside key technical partners like NVIDIA, BME-HIT, Owkin, KU Leuven, the Substra Foundation, and Kubermatic.
The goal? Train ML models collaboratively - without ever sharing raw data.
How it worked
MELLODDY used federated learning, a technique that allows models to learn from distributed data while keeping it private. Instead of pooling raw data in a central location, each partner trained the model locally, and only encrypted training updates (gradients) were shared with the global model. This ensured that no confidential data was ever exposed.
Kubermatic played a key role in providing the scalable Kubernetes infrastructure for each pharma partner, based on the Kubermatic Kubernetes Platform (KKP). The platform was deployed on Amazon Web Services (AWS) multi-account architecture, and it allowed companies to run Kubernetes clusters in private subnets.
The Impact
MELLODDY successfully created the first industry-scale federated learning platform for drug discovery.
This breakthrough project resulted in a global model that outperformed any individual company’s ML models. It proved that pharma companies can work together without compromising their data.
In the end, everyone won - MELLODDY broke free from the prisoner’s dilemma, turning competition into “coopetition”. And the biggest winner? Society, which will benefit from faster, more efficient drug discovery in the years to come.
To learn more about the MELLODDY project, explore the project report or this article.
Get in touch with us! Book a demo or learn more about KKP.
