Research

My work is organized around four connected themes, rather than a single project. Across them, I develop Bayesian graphical models, global-local shrinkage methods, spatial statistics, and multimodal learning for cancer genomics and the tumor microenvironment. See Publications for full citations and Data & Projects for code and model figures.

1. Bayesian Graphical Models & Sparse Network Estimation

Foundational methods work on Bayesian Gaussian graphical models, where a sparse precision matrix encodes conditional-dependence edges between variables. I study scalable covariance selection and neighborhood selection under horseshoe and spike-and-slab priors, joint estimation of related networks through shared shrinkage, and regression with a network-valued predictor.

• Bhadury, S., Mitra, R., Gaskins, J. Fast Bayesian High-Dimensional Gaussian Graphical Model Estimation
• Bhadury, S. et al. Bayesian Joint Estimation of Multiple Graphs through Shared Shrinkage

2. Spatial Statistics for Multiplexed Tissue Imaging

I develop spatially varying graphical models for the tumor immune microenvironment, recovering shared and region-specific cell–cell interaction networks from 2D and 3D multiplexed immunofluorescence data. The models connect spatial heterogeneity, cellular neighborhoods, and tissue architecture with Gaussian-process structure and spatially adaptive shrinkage.

• Bhadury, S. et al. ISPAT: Informed Spatially Aware Patterns for Multiplexed Immunofluorescence data, Nature Scientific Reports. Code
• Bhadury, S., Rao, A. ISPAT 3D: Spatially Varying Conditional Volumetric Network Estimation for 3D Tumor Imaging. Code
• Bhadury, S., Gaskins, J., Rao, A. Spatially Varying Graphical Models for Cell-Cell Interaction Networks in Multiplexed Tissue Imaging Code

3. Multi-Modal Integration & Imaging Genomics

This work fuses histopathology, spatial transcriptomics, radiomics, and molecular features into coherent predictive or inferential models. The scientific targets include spatially variable gene programs, gene regulatory and ligand–receptor interactions, image–transcriptome associations, multimodal survival analysis, and biomarker discovery for precision oncology.

• Tsang, A., Krishnan, S., Kulkarni, R., Bhadury, S., Rao, A., et al. Encoding functional edges in graphs to model spatially varying relationships in the tumor microenvironment, npj Artificial Intelligence
• Bhadury, S. et al. Graph-Based Multi-Modal Learning for Integrating Histopathological and Transcriptomic Features in Lung Cancer Code
• Bhadury, S. et al. RISM: Radiomic feature Integration with Spike-and-slab for Multimodal survival analysis. Code
• PathMI — mutual information estimators (OLS, LASSO, Bayesian horseshoe) linking histopathology embeddings to molecular subtype. Code

4. Causal Inference & Functional Data Analysis

I use causal mediation and sensitivity analysis to study spatial cell-type co-occurrence, with explicit attention to the assumptions separating association from causal interpretation. In parallel, functional-depth-based classifiers and interpretable tree-derived rules turn curve-valued histopathology features into robust biomarkers for tumor subtyping.

• Bhadury, S. et al. Functional Depth biomarkers Distinguish Lung Cancer Subtypes. Code
• Mediation and sensitivity analyses of spatial cell-type co-occurrence patterns. Code
• Statistical shape analysis (PCA-based) of MRI-derived tumor contours in lower-grade glioma. Code