SimSpace: a comprehensive in-silico spatial omics data simulation and modeling framework

Abstract

Spatial omics technologies provide rich measurements of molecular states in their native tissue context, but the development and benchmarking of computational methods remain constrained by limited access to datasets with known ground truth. We present SimSpace, a flexible framework for simulating spatial omics data for method development, stress-testing, and controlled in silico experiments. SimSpace uses a hierarchical generative model that couples tissue-scale spatial organization, local cell-type interactions, and modular omics-profile generation. Its Markov random field-based spatial model enables controllable simulation of niche structure, cell-type co-localization, phenotype-dependent density, and spatially mediated molecular interactions. SimSpace supports both reference-free simulations under user-defined generative assumptions and reference-guided simulations that calibrate spatial parameters to real datasets while generating new spatial realizations. Across Xenium, MERFISH, and CODEX examples, SimSpace reproduced multiple evaluated spatial statistics, preserved key domain-level structure in reference-guided settings, and supported benchmarking of cell-type deconvolution and spatially variable gene detection methods. We further show that inferred interaction parameters can be perturbed to generate interpretable counterfactual spatial proteomics scenarios. SimSpace is implemented as an open-source Python package with reproducible workflows for rigorous benchmarking and methodological development in spatial omics.