NoisyFlow

NoisyFlow integrates optimal transport theory with neural networks and differential privacy mechanisms to align data distributions across institutions while preserving individual privacy for secure biomedical data sharing.

Key Features:

• Differentially Private Neural Networks: The framework incorporates differential privacy mechanisms to provide privacy guarantees for individual data points during the data alignment process.
• Optimal Transport-Based Distribution Alignment: NoisyFlow uses optimal transport theory combined with neural networks to align data distributions across institutions without requiring direct data sharing.
• Covariate and Batch Effect Reduction: The framework reduces distribution shifts caused by covariate and batch effects when integrating data from multiple sources.
• High-Dimensional Data Handling: NoisyFlow handles high-dimensional single-cell genomic data and histopathology images while maintaining utility in downstream tasks.

Scientific Applications:

• Secure Biomedical Data Sharing: NoisyFlow enables collaboration across institutions on sensitive genomic and clinical datasets without direct data exchange.
• Single-Cell Genomic Data Integration: The framework aligns high-dimensional single-cell genomic data across diverse patient populations for improved analysis.
• Histopathology Image Analysis: NoisyFlow supports privacy-preserving sharing and analysis of histopathology images across institutions.
• Disease Classification: The framework maintains high utility for downstream disease classification tasks after privacy-preserving data alignment.

Methodology:

NoisyFlow integrates optimal transport theory with neural networks and differential privacy mechanisms to align data distributions across institutions.