Federated learning for multi-omics: A performance evaluation in Parkinson's disease.

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in healthcare

Background:

• Machine learning (ML) application in omics is limited by data access.
• Federated learning (FL) offers a solution for collaborative data curation across institutions.
• Parkinson's disease (PD) research can benefit from advanced ML techniques applied to multi-omics data.

Purpose of the Study:

• To compare the performance of FL-trained models against classically trained ML models for multi-omics PD prediction.
• To evaluate the impact of sample dispersion within a federation on FL model performance.
• To highlight challenges and opportunities of applying FL in multi-omics studies.

Main Methods:

• Simulated performance evaluation of multiple ML models trained using FL.
• Comparison with ML models trained using traditional centralized approaches.
• Implementation of several open-source FL frameworks.
• Analysis of multi-omics data for Parkinson's disease prediction.

Main Results:

• FL model performance closely mirrors that of centrally trained ML models.
• The top-performing FL model achieved an AUC-PR of 0.876 ± 0.009, only slightly below its centralized counterpart.
• Sample dispersion within the federation significantly influences FL model performance.

Conclusions:

• Federated learning is a viable approach for training high-performance ML models in multi-omics studies, even with distributed datasets.
• FL can mitigate data access limitations in sensitive research areas like Parkinson's disease.
• Further research into optimizing FL strategies based on data dispersion is warranted.