Substra

Substra enables federated learning to train machine-learning models across distributed datasets while preserving data privacy for applications such as predicting neoadjuvant chemotherapy (NACT) response in triple-negative breast cancer (TNBC) using whole-slide images and clinical data.

Key Features:

• Federated learning (FL): Implements federated learning workflows for collaborative training of machine-learning models across distributed datasets.
• Privacy-preserving data locality: Keeps patient data secured within institutional firewalls so raw data do not leave local sites during model training.
• Single-machine simulation: Supports virtual partitioning of datasets to simulate federated settings and debug experiments on a single machine.
• Multimodal data integration: Enables training of models using whole-slide images and clinical information.
• Interpretability: Produces interpretable models that identify histological patterns associated with treatment response.
• Proven deployment: Has been deployed in multicentric projects such as MELLODDY and a federated study of early-stage TNBC to predict NACT response.
• Performance improvement via FL: Collaborative federated training improves predictive performance relative to locally trained models and can reach performance comparable to expert annotations.

Scientific Applications:

• NACT response prediction in TNBC: Training ML models to predict neoadjuvant chemotherapy response in triple-negative breast cancer from whole-slide images and clinical data.
• Multicentric federated studies: Enabling multicentric studies that pool model training across hospitals while keeping data behind institutional firewalls.
• Biomarker discovery: Supporting biomarker discovery by enabling large-scale analysis of histological patterns across previously inaccessible datasets.
• Collaborative biomedical research: Supporting collaborative pharmaceutical and biomedical research workflows as demonstrated in the MELLODDY project.

Methodology:

Federated learning of local ML models on whole-slide images and clinical data across distributed datasets; virtual partitioning to simulate FL on a single machine; and model interpretability to identify histological patterns associated with treatment response.