D3aist: data-driven detection of atypical interactions in spatial transcriptomics.

Tissues exhibit complex spatial organization that governs cellular function and interactions, yet traditional transcriptomic methods fail to capture this context. Spatial transcriptomics preserves cellular locations while profiling gene expression, enabling high-resolution mapping of tissue architecture. Molecule-resolved spatial transcriptomics data, in which individual transcript detections are associated with spatial coordinates and gene identities, can be naturally modeled as multitype point patterns within the framework of spatial statistics. Second-order summary statistics, such as the Ripley cross K-, cross L-, and pair correlation functions, are commonly used to characterize spatial interactions; however, systematic comparison across multiple types remains challenging. We propose a data-driven method, termed D3AIST, based on empirical envelopes constructed via a leave-one-pair-out strategy over the set of observed functions. Each gene-pair type is evaluated by comparing its second-order summary functions against an empirical envelope constructed from the corresponding functions of all other pairs, excluding the one under evaluation. This allows the detection of atypical spatial interactions without relying on classical null models or asymptotic results. D3AIST provides a global comparative scheme across all type pairs, facilitating the identification of singular interaction patterns within a fully empirical framework. We applied D3AIST to transcript-level Xenium data from two publicly available studies. In a primary colorectal cancer case study based on GSE280318/GSE280314, D3AIST identified atypical spatial interaction networks across expert-defined tumor microenvironment regions. An independent external evaluation using the GSE267680 Xenium dataset further illustrated the applicability of the framework in a distinct pancreatic intraepithelial neoplasia context. Controlled simulations and computational benchmarking were additionally used to evaluate empirical error behavior and scalability. Overall, D3AIST provides a descriptive and hypothesis-generating framework for prioritizing atypical transcript-level spatial associations in molecule-resolved spatial transcriptomics data.