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

In silico approaches have acquired atoweringrole in pharmaceutical research and development, allowing laboratoriesall around the world to design, create, and optimize novel molecularentities with unprecedented efficiency. From a toxicological perspective,computational methods have guided the choices of medicinal chemiststoward compounds displaying improved safety profiles. Even if therecent advances in the field are significant, many challenges remainactive in the on-target and off-target prediction fields. Machinelearning methods have shown their ability to identify molecules withsafety concerns. However, they strongly depend on the abundance anddiversity of data used for their training. Sharing such informationamong pharmaceutical companies remains extremely limited due to confidentialityreasons, but in this scenario, a recent concept named "federatedlearning" can help overcome such concerns. Within this framework,it is possible for companies to contribute to the training of commonmachine learning algorithms, using, but not sharing, their proprietarydata. Very recently, Lhasa Limited organized a hackathon involvingseveral industrial partners in order to assess the performance oftheir federated learning platform, called "Effiris".In this paper, we share our experience as Roche in participating insuch an event, evaluating the performance of the federated algorithmsand comparing them with those coming from our in-house-only machinelearning models. Our aim is to highlight the advantages of federatedlearning and its intrinsic limitations and also suggest some pointsfor potential improvements in the method.