Collaborative Research: Privacy-Preserving Statistical Learning for High-Dimensional Spatial Omics

Spatial genomics technologies can now measure gene expression while maintaining their tissue context, showing not only which cells are present, but also where they are located and how they interact with their surrounding environment. These spatial maps are especially important for diseases such as cancer, where the tumor microenvironment may influence disease progression, therapeutic resistance, and treatment response. However, a single hospital or research center often does not have enough patient samples to identify statistically reliable biological patterns, while privacy rules, data-use agreements, and institutional policies often prevent direct sharing of patient-level molecular data. This project will develop new statistical theories and computational methods that allow multiple hospitals and research laboratories to extract statistical insights from each other’s spatial genomics data without directly sharing sensitive patient information. The project will help researchers create a decentralized collaborative network to detect molecular spatial patterns from high dimensional raw measurements, measure how reliable those discoveries are based on statistical evidence, and protect patient privacy during collaborative analysis. By enabling secure and rigorous analysis of distributed spatial genomics data, the project will advance statistical science, support privacy-aware prediction medicine, broaden access to advanced spatial omics analytics, and provide interdisciplinary training opportunities for undergraduate and graduate students at the intersection of statistics, computation, machine learning, and genomics. This project develops a statistically principled and computationally scalable framework for privacy-preserving decentralized learning from high-dimensional spatial omics data. Recent advances in spatial transcriptomics and multiplexed imaging now contain thousands of genes measured across hundreds of thousands to millions of spatial locations, with complex spatial dependence and substantial heterogeneity across tissues, patients, platforms, and institutions. As data dimensions grow, cross-institutional collaboration becomes essential for capturing population variability and improving statistical power. However, privacy concerns and institutional data-sharing restrictions pose significant challenges for joint statistical analysis. To address these challenges, this research aims to establish the statistical foundation for distributed learning of spatial omics. The project pursues three synergistic aims. Aim 1 develops decentralized spatial factor analysis algorithms that identify spatially coherent latent features from distributed heterogeneous spatial omics data. The approach integrates priors from partial differential equations with factor analysis to capture spatial dependencies and enable large-scale collaborative inference without sharing raw data. Aim 2 constructs uncertainty quantification frameworks based on Riemannian stochastic gradient Langevin dynamics. This aim will implement manifold optimization procedures that propagate uncertainty across institutions while respecting communication and privacy constraints. Aim 3 builds scalable nonlinear spatial operators that extract nonlinear spatial embeddings through decentralized learning under a privacy-preserving framework tailored to spatially correlated data. The expected outcomes include new theory for decentralized high-dimensional statistical inference, new methods for uncertainty-aware collaborative learning, new privacy protection mechanisms for correlated spatial domains, and open-source software tools for secure multi-institution analysis for spatial omics data. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.