Federated Deep Learning Enables Cancer Subtyping by Proteomics.

Area of Science:

• Biomedicine
• Artificial Intelligence
• Proteomics
• Machine Learning

Background:

• Artificial intelligence (AI) in biomedicine is hindered by stringent data privacy regulations.
• Clinically annotated tissue proteomic data presents unique privacy challenges.
• Developing large-scale AI models requires access to diverse, multi-institutional datasets.

Purpose of the Study:

• To develop a privacy-preserving federated deep learning approach for analyzing tissue proteomic data.
• To address the challenge of data privacy in AI applications within biomedicine.
• To enable international collaboration for proteomic data analysis and biomarker discovery.

Main Methods:

• Developed ProCanFDL, a federated deep learning framework.
• Trained local models on simulated and private firewall-protected proteomic datasets (pan-cancer and 29 cohorts).
• Aggregated local model updates to create a global model, validated on hold-out and external datasets.

Main Results:

• ProCanFDL achieved a 43% performance gain in 14 cancer subtyping tasks compared to local models.
• The federated approach matched the performance of a centralized model.
• Demonstrated generalizability by retraining on diverse external proteomic datasets (DIA-MS and TMT proteomics).

Conclusions:

• ProCanFDL offers a viable solution for privacy-compliant machine learning with proteomic data.
• Enables collaborative initiatives for biomarker and treatment target discovery.
• Facilitates the development of large-scale, privacy-preserving AI models, including foundation models, across global institutions.