In June 2019, a new consortium of pharmaceutical, technology, and academic partners launched the MELLODDY project. The objective? Leveraging the world’s largest collection of small molecule data with known biochemical or cellular activity to enable more accurate predictive models and increase efficiencies in drug discovery. Apart from the huge technological achievement that it represents, the consortium at its heart is a human adventure - 118 people are involved and dedicated to making this 3-year project successful.
In this article, we’ll take a look at the MELLODDY consortium from the inside, with Hideyoshi Fuji, a Computational Chemist/Chemoinformatician leading AI-driven drug discovery at Astellas in Japan.
Joining the MELLODDY consortium
“Back in 2016, when Hugo Ceulemans, Scientific Director Discovery Data Sciences at Janssen, presented the MELLODDY project to Astellas, my first impression was that this was an incredibly challenging task to achieve.
We are in an AI and big data era, and I truly believe that this will be a game-changer in accelerating the drug development process for pharmaceutical companies. Today, internal data, particularly molecular data, is limited and AI is data-hungry. MELLODDY solves both challenges as it increases data volume while preserving privacy and enabling machine learning at scale in a coopetitive effort to promote drug development"
“To find a solution (and more data), we needed to start collaborating with other pharmaceutical companies, by sharing the insights from our internal data with each other. But at that time, there was no technology to do that in a privacy-preserved way. The MELLODDY research project proposed a new privacy-first machine learning technique to unlock maximum potential of pharma industry data, so we decided to join.”
The development cycle of AI methodology has increased rapidly, so pharmaceutical companies didn't want to lose out on a new development in AI. Joining this consortium allows us to use cutting edge technology and be at the forefront of innovation in machine learning. With increasing competition from GAFA (Google, Apple, Facebook, Amazon) also trying to develop healthcare solutions, we wanted to make sure we are one step ahead in the AI/ML field and joining this consortium was a great place for us to start.
The main challenges faced by Astellas
Astellas has been working on various open innovation efforts to turn innovative science into value for patients. But we were being very cautious about data sharing with other partners since drug discovery data, such as chemical structure, biological activity, and ADME/Tox properties, are directly related to intellectual property. In other words, they are competitive assets for pharmaceutical industries. Initially, we were skeptical that the federated learning approach could preserve our drug discovery data.
Frequent discussions on federated learning with Hugo Ceulemans (Janssen Pharmaceutica NV) helped us to understand how the technology works and to evaluate the benefits and risks. In this new federated learning approach, the underlying data contributions are not shared, and remain under control of the respective data owners. In the end, we realized all the pharma partners had the same feelings, that is, the highest priority in this MELLODDY platform is security and privacy preservation and our goal is to leverage big data to better efficiencies in drug discovery. The same ambition within pharma partners pushed us to join this consortium.
At Astellas, we are now thinking about further applications of federated learning. I have identified several applications that could be useful in solving problems in sharing clinical image data or health care data like genomic data. The potential of federated learning in healthcare is unlimited!
In September 2020, the MELLODDY project met its first year objective, with the deployment of the world’s first secure platform for multi-task federated learning in drug discovery at scale involving 10 pharmaceutical companies. In April 2021, it initiated its second cycle of privacy-preserving federated Machine Learning at scale on the drug discovery data of the 10 pharma partners. The data-sharing challenges within pharma research were successfully overcome, even with increased data complexities.
Over the next two years, the MELLODDY project will focus on improving the performance of the multi-partner predictive models by exposing it to an increasing amount of data and exploring new scientific and technical options.
________________________________________________________________________
Contact
MELLODDY Communication: Tinne Boeckx – Janssen (TBoeckx1@its.jnj.com) ; Ashley Nicollet (ashley.nicollet@owkin.com)