privateST: a feasible framework for privacy-preserving spatial transcriptomics prediction from histopathology images

Area of Science:

• Computational Biology
• Genomics
• Medical Imaging

Background:

• Spatial transcriptomics prediction from histology images offers cost-effective biological insights.
• Privacy concerns hinder cross-institutional data sharing, limiting the scope of such predictions.
• Existing methods face computational and privacy challenges.

Purpose of the Study:

• To develop a privacy-preserving framework for spatial transcriptomics prediction using homomorphic encryption.
• To demonstrate the feasibility of secure, cross-institutional analysis of histology data for gene expression prediction.
• To optimize deep learning architectures for homomorphic encryption's computational constraints.

Main Methods:

• Developed privateST, a framework optimized for homomorphic encryption.
• Downsampled histology images using bilinear interpolation to manage computational load.
• Incorporated auxiliary gene predictions to enhance target gene prediction accuracy.
• Modified neural network architecture (ResNet-18) by replacing Max-Pooling with Average-Pooling and ReLU with polynomial approximation.
• Implemented multiplexed packing for efficient encrypted data processing.

Main Results:

• Achieved prediction accuracy comparable to non-encrypted ResNet-18 models despite reduced image resolution.
• Demonstrated the feasibility of secure spatial transcriptomics prediction using homomorphic encryption.
• Successfully adapted deep learning models for homomorphic encryption by minimizing non-linear operations.

Conclusions:

• The privateST framework enables secure and accurate spatial transcriptomics prediction from histology images.
• Homomorphic encryption can be effectively applied to complex deep learning tasks in genomics.
• This approach overcomes privacy barriers, facilitating collaborative research and broader data utilization.