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

Predicting spatial transcriptomics from histology images offers cost-effective insights but faces privacy barriers preventing cross-institutional data sharing. To address this, we present privateST, a homomorphic-encryption-optimized framework that substantiates the feasibility of secure spatial transcriptomics prediction. To accommodate the computational constraints of homomorphic encryption, we downsampled the input images using bilinear interpolation. To improve prediction accuracy for the 100 target genes, we incorporated predictions for an auxiliary set of 150 highly expressed genes. This approach enables the model to learn from a more diverse genomic context, thereby refining the feature representations for the high-priority target genes. To adapt the architecture for homomorphic encryption, we replaced Max-Pooling and ReLU with Average-Pooling and polynomial approximation, respectively. These modifications eliminate non-linear comparison operations, significantly reducing the multiplicative depth. We also implemented multiplexed packing to optimize the efficiency of encrypted data processing. Remarkably, our results demonstrate that despite the reduced input resolution, this proposed approach achieves accuracy comparable to the original ResNet-18 configurations without downsampling.