NoisyFlow: Differentially Private Optimal Transport Using Neural Networks for Secure Biomedical Data Sharing

Abstract

MotivationAdvancing data sharing in biomedical research, particularly for sensitive genomic and clinical datasets, is crucial for improving model performance across diverse patient populations. However, stringent privacy concerns hinder collaboration and limit insights derived from multi-institutional datasets. Current approaches to privacy-preserving data sharing fail to address gaps between data distributions. ResultsWe introduce NoisyFlow, a differentially private neural network-based optimal transport framework designed to enable secure and unbiased biomedical data sharing. By integrating optimal transport theory with neural networks and differential privacy mechanisms, our framework aligns data distributions across institutions while preserving individual privacy. NoisyFlow eliminates the need for direct data sharing and reduces distribution shifts caused by covariate and batch effects. Empirical evaluations demonstrate the frameworks effectiveness in handling high-dimensional single-cell genomic data and histopathology images, achieving superior privacy guarantees while maintaining high utility in downstream tasks such as disease classification. Availability and implementationThe implementation of NoisyFlow is available at https://github.com/liyy2/NoisyFlow. Contactmark@gersteinlab.org. Supplementary informationSupplementary data are available online.