Advances in Statistical and AI Methods for Digital Pathology and Spatial Molecular Imaging

Virtual staining is a cross-domain image synthesis approach that applies deep learning to generate digitally "stained" pathology images—either from unstained tissue or by translating one staining modality into another (e.g., mIF to H&E). This approach reduces reliance on labor-intensive laboratory protocols and enables scalable, cost-effective digital alternatives. Leveraging co-registered mIF and H&E whole-slide images, we align cell type labels derived from mIF (via the CellGate pipeline) with H&E slides to create a large-scale dataset of labeled H&E images at single-cell resolution. These data provide a powerful resource for training virtual staining models that enable automated, cell-level interpretation of H&E slides without requiring multiplexed imaging. Virtual staining thus expands access to high-quality spatial tumor microenvironment analysis, advancing both research and clinical applications. 

AI-powered pathology foundation models provide general-purpose representations of histopathological images by encoding image tiles into numerical embeddings. However, these embeddings are not directly interpretable in biological or clinical terms and must be translated into biologically meaningful features, such as cell-type composition or gene expression, to enable downstream clinical applications. To bridge this gap, we developed STpath, a framework that integrates histopathology images embedding derived from existing pathology foundation models with matched spatially resolved transcriptomics data. STpath consists of cancer-specific XGBoost models trained to infer cell-type compositions and gene expression from histopathology image tiles. We evaluated STpath in colorectal and breast cancer datasets and showed that it provides accurate estimates of the composition of major cell types and the expression of a subset of genes, with further performance gains achieved by combining embeddings from multiple foundation models. Finally, we demonstrated that STpath inferred features can be used in down-stream association studies to assess their associations with clinical outcomes.
 

The recruitment of various types of cells into the tissue microenvironment and how these cells engage with other cells in the tissues play critical biological roles. However, it is difficult to study these processes on a genome-wide scale using low-throughput experiments. To fill this void, we introduce SPACER, an interpretable and generative multi-instance learning framework, which digests the transcriptomic and spatial modalities of spatially resolved transcriptomics (SRT) data to elucidate the mechanisms of cellular engagement, in a manner that is isomorphic to the tissue spatial architecture. We deployed SPACER to 17 high and 20 low definition whole transcriptomic SRT datasets. SPACER identified tumor cell- and stromal/immune cell-specific features that determined the recruitment of these cells into tumors. SPACER could also map transcriptomic features that were associated with T cell infiltration into cardiomyocytes during myocarditis. SPACER out-performed all benchmark methods in terms of prediction accuracy and recovery of biological signals. Overall, SPACER establishes a new spatially resolved paradigm for studying cellular localization in situ.

 

Integrating pathology images with spatial transcriptomics provides comprehensive insights into the tissue microenvironment. However, a significant challenge remains in translating these multi-modal data into clinical outcome predictions at the cohort level. While existing spatial transcriptomics (ST) tools excel at within-sample analysis and cross-modal gene expression imputation, there remains a lack of a unified framework for direct, cohort-level prognostication. Here,  we introduce PathoSpatial, an interpretable multimodal framework that integrates co-registered whole slide images (WSIs) and ST to learn spatially informed prognostic representations. PathoSpatial uses task-guided prototype learning within a hierarchical, multi-level expert architecture, decoupling unsupervised modality-specific prototype discovery from supervised cross-modal prototype aggregation. This design improves discriminative capability while preserving biological interpretability. Evaluated on a large breast cancer ST cohort, PathoSpatial delivers the best overall performance on survival outcome prediction, consistently surpassing or matching leading uni- and multimodal methods. Post-hoc prototype interpretation and molecular risk decomposition provide quantitative, biologically grounded explanations and highlight clinically actionable determinants of prognosis, advancing scalable multimodal integration for next-generation computational pathology. 

Spatial transcriptomics generates high-dimensional, spatially indexed gene expression data with complex dependence structures and measurement noise, making low-dimensional spatial representation learning a critical step for characterizing tissue organization and spatial heterogeneity. While recent advances have largely relied on auto-encoder-based 'black-box' approaches to learn spatial embeddings, these methods often lack interpretability, biological traceability, and principled uncertainty quantification. From a statistical perspective, Bayesian spatial factor models provide a transparent and generative alternative, but their practical use has been limited by identifiability issues, unstable posterior inference, and scalability challenges. This talk presents an explainable spatial representation learning framework based on Bayesian spatial factor modeling that addresses these limitations. A projected MCMC sampling strategy is introduced, leveraging conditional conjugacy and projection to constrain latent factor sampling to a scaled Stiefel manifold, substantially improving posterior stability, mixing efficiency, and robustness to initialization. Applications to large-scale human kidney spatial transcriptomics data demonstrate that the proposed approach yields spatially smooth, interpretable embeddings, biologically meaningful gene-level spatial effects, and uncertainty-aware inference for spatial molecular analysis.