MELLODDY Using Federated Learning To Improve Drug Development
By Deborah Borfitz
July 2, 2019 | A 17-partner consortium in Europe is seeking to confirm the utility of a machine learning platform for better predicting promising compounds for drug development. The three-year initiative, which launched in June, represents the first large-scale deployment of blockchain technology to extract insights from multiple preclinical datasets without having to first pool the data, according to Hugo Ceulemans, industry leader of the Machine Learning Ledger Orchestration for Drug Discovery (MELLODDY) project.
Ceulemans is scientific director, Discovery Data Sciences at Janssen Pharmaceutica NV (a Johnson & Johnson company), one of 10 European pharmaceutical companies that compose the MELLODDY consortium and are collectively funding more than half of the project’s €18.4 million (over $20 million) budget. The remainder of the funding is being funneled through the Innovative Medicines Initiative (IMI) by the European Union.
The pharma partners will for the first time be collaborating in their core competitive space to improve the overall efficiency of bringing a drug to market, which on average takes about 13 years and 1.9 billion euros (Journal of Health Economics). Artificial intelligence was already on their radar as a potential solution, says Ceulemans, so they were eager to participate when they learned the requisite information exchange could happen “without imperiling the privacy and sensitivity” of proprietary databases.
The pharma partners will for the first time be collaborating in their core competitive space to bring much-needed efficiencies to a drug development process that notoriously consumes an average 13 years and many billions of euros per medicine, Ceulemans notes. Artificial intelligence was already on their radar as a potential solution, so they were eager to participate when they learned the requisite information exchange could happen “without imperiling the privacy and sensitivity” of proprietary databases.
The predictive machine learning platform will enable each company to focus on molecules most worth pursuing, leveraging the annotated chemical library of each company while owners keep control of their data and their intellectual property remains protected, he says.
For participating pharma companies, one advantage of consortium membership is that they are at liberty to use whatever prediction results emerge during the three-year project period, Ceulemans says. One large AI company, four subject matter experts and a pair of universities are also part of the platform-creating initiative.
Mapping of the MELLODDY platform to the hardware layer is supported by NVIDIA, manufacturer of graphics processing units popular in the gaming industry, which is “very keen on prominently positioning its products in novel deep learning applications,” says Ceulemans. IT startup Loodse will deploy platform software components across the distributed infrastructures of pharma companies.
Owkin and the Substra Foundation will bring in the distributed ledger technology (aka “blockchain light”) that will enable partners to keep control and visibility over their own private data, Ceulemans continues. Communications between a dispatcher and the ledger require the approval of all other partners. Like a bank statement, the ledger holds a log of all activities that can be requested after a federated run.
Machine learning algorithms are being supplied by Katholieke Universiteit Leuven, a Belgian research university; Budapesti Muszaki és Gazdaságtudományi Egyetem, a Hungarian technical university; and Iktos, a French biotech startup specializing in machine learning for drug discovery. Privacy and security features are embedded in the algorithms, allowing for data-sharing, he says.
If the MELLODDY platform succeeds in demonstrating its predictive prowess, Ceulemans says, the technology will become commercially available and likely find additional applications to include clinical data sharing—at which point the personal data protections of GDPR would kick in.