MELLODDY: cross pharma federated learning at unprecedented scale unlocks benefits in QSAR without compromising proprietary information

Abstract

Federated multi-partner machine learning can be an appealing and efficient method to increase the effective training data volume and thereby the predictivity of models, particularly when the generation of training data is resource intensive. In the landmark MELLODDY project, each of ten pharmaceutical companies realized aggregated improvements on its own classification and/or regression models through federated learning. To this end, they leveraged a novel implementation extending multi-task learning across partners, on a platform audited for privacy and security. The experiments involved an unprecedented cross-pharma dataset of 2.6+ billion confidential experimental activity data points, documenting 21+ million physical small molecules and 40+ thousand assays in on-target and secondary pharmacodynamics and pharmacokinetics. Appropriate complementary metrics were developed to evaluate predictive performance in the federated setting. In addition to predictive performance increases in labeled space, the results point towards an extended applicability domain in federated learning. Increases in collective training data volume, including by means of auxiliary data resulting from single concentration high-throughput and imaging assays, continued to boost predictive performances, albeit with saturating return. Markedly higher improvements were observed for pharmacokinetics and safety panel assay-based task subsets.

Industry-Scale Orchestrated Federated Learning for Drug Discovery

“…This would allow partners to contribute more equally to the aggregated gradients. In the work of Heyndrickx et al (2022) it was however found that default weighting was more beneficial for the federated model.…”

“…MELLODDY Federated Learning (FL) Simulator During the MELLODDY project, many options were considered to improve the performance of the model (Heyndrickx et al 2022). In order to quickly assess each option, the partners conducted single partner studies.…”

“…To this end, each partner can select the group of already trained models from the pool of single-and multipartner models. The best model for each task is then selected based on the separate dataset and the performance is measured based on the held-out test set (Heyndrickx et al 2022).…”

“…In this paper, we describe a real world application for drug discovery in the context of a European Innovative Medicines Initiative (IMI) project called MELLODDY (grant n°831472), which gathered 10 pharmaceutical companies, academic research labs, large industrial companies and startups. The platform developed for MELLODDY project enabled to produce scientific results described in the work of Heyndrickx et al (2022). The data used in MEL-LODDY were chemical and assay data resulting from research conducted by ten pharmaceutical companies over the span of decades.…”

Federated Learning in Computational Toxicology: An Industrial Perspective on the Effiris Hackathon

“…Federated learning has already been applied to drug discovery, as assessed by the recent outcomes of the MELLODDY project, which involved ten pharma companies, sharing federated information for a total of more than 21 million small molecules. , The MELLODDY approach uses multitask learning and underlying deep neural network architectures. In brief, local models were trained at each company, and only the gradients were exchanged, thus avoiding the disclosure of information on the underlying data and on the modeled end points.…”