Extrachromosomal DNA as a Causal Instrument for Spatial Multi-Omics

Spatial transcriptomics, multiplex imaging, and computational pathology now map tissue organization at cellular resolution, but the analyses applied to these data remain correlational. Clustering and co-occurrence statistics describe which features appear together; they cannot say which feature drives the others. We propose a framework for causal inference in spatial multi-omics built on a specific feature of extrachromosomal DNA (ecDNA). ecDNA carries no centromere; it does not attach to the mitotic spindle and partitions randomly to daughter cells at division. Two neighboring cells in the same microenvironment can therefore inherit very different oncogene copy numbers for reasons unrelated to local signaling. Together, these cell-intrinsic randomization properties of ecDNA provide the framework for ecDNA copy number serving as an instrumental variable (IV) separating the effects of oncogene dosage from the downstream cellular effects. In this study, we formalize this within a structural causal model, implement two-stage least squares estimation with sibling-comparison and falsification diagnostics, and provide sensitivity analyses for residual confounding. To benchmark these methods against ground truth, we built CAUSANTA, an agent-based simulator that generates spatial tissue with a known causal graph and stochastic ecDNA inheritance. The framework recovers known oncogene-to-phenotype effects from simulated data, and we outline its application to real tumor sections, including multi-instrument settings where independent ecDNA species carrying different oncogenes enable factorial designs within a single tumor.