Développement de modèles d'apprentissage fédéré pour améliorer l'évaluation des variants génétiques dans un cadre clinique multi-sites

Federated learning (FL) is a machine learning (ML) technique that enables multiple data holders to collaboratively train a model, without raw data sharing. This approach is particularly relevant in the field of genomics, where data is often distributed across institutions, and regulatory constraints, such as General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA), restrict data centralization. In addition to improving privacy and data security, FL allows the training of more robust ML models, by leveraging access to a larger and more diverse dataset. FL was first proposed by Google researchers in 2017 as an approach for training ML models on a federation of mobile devices coordinated by a central server. In this setup, the mobile devices contained data that was either sensitive or large in size with respect to the ML model. In their implementation, the server defined a global ML model and communicated the parameters to a subset of clients. The clients then optimized the received model by performing Stochastic Gradient Descent on local data, and sent back the local updates to the server. The server derived a new global model by aggregating the local updates through weighted averaging. This process was repeated for a predefined number of rounds or until the model converged. FL has also been adapted to cross-silo settings, where clients (typically 2-50) are organizations, such as hospitals and research institutions. The objective of this thesis is to study the effectiveness of cross-silo FL for the clinical assessment of human genetic variants. To that extent, we leveraged the public-available database ClinVar for simulating realistic multi-institutional collaborations in the assessment of coding Single Nucleotide Variants (SNVs), non-coding SNVs, and copy number variants (CNVs). Concretely, we evaluated the performance of a diverse set of supervised ML models, trained in a FL manner across multiple institutions, in classifying genetic variants into pathogenic or non-pathogenic, and compared it to the centralized and individual- institution (local) model counterparts. Our results showed that FL generally achieved competitive or superior performance than the centralized model, and systematically outperformed the local models, highlighting the advantages of collaboration. In the experiments we benchmarked several FL aggregation strategies, including FedProx, FedAdagrad, FedAdam, and FedYogi, which refer to the methods used by the server to combine local updates into a new global model. Our results showed that FedProx generally provided the best performance. Furthermore, we analyzed the performance degradation of both FL and its centralized model counterpart when one institution decided not to participate in the collaborative training. We found that FL was more resilient than centralized approaches in most cases, demonstrating that FL can generalize adequately to unseen datasets with smaller training sets. To the best of our knowledge, this thesis presents the first simulated FL study for the pathogenicity classification of genetic variants. With our findings, we expect to incentive the adoption of FL for establishing secure multi-institutional collaborations in human variant interpretation.