New Score Quantifies Directional Influence in Spatial Graphs

Understanding directional influence between different components in complex systems, particularly in spatial biological contexts like cell-cell interactions, is a critical challenge in graph-based modeling. Traditional methods, such as those based on attention or correlation, often identify associations but lack a robust framework for assessing true directional effects under controlled experimental conditions. This new research proposes a novel approach to address this limitation. The core of the method is the Counterfactual Directionality Score (CDS), which leverages structured counterfactual interventions within graph-based models. It begins by training a Neighbor Influence Model (NIM) to predict the state of a node based on its local neighborhood. Subsequently, constrained interventions are applied to modify the neighborhood composition while carefully preserving essential spatial and structural properties of the graph. The CDS then quantifies the change in the predicted node state resulting from these targeted perturbations, offering a principled measure of local intervention sensitivity. To ensure the reliability of uncertainty estimates, the researchers developed a core-level bootstrap procedure that accounts for dependencies inherent in spatial samples. Validation on synthetic spatial graphs with known directional structures confirms that CDS accurately recovers directional influence, maintains calibration under null conditions, and exhibits robustness against confounding signals. Early applications to spatial transcriptomics data have also revealed biologically plausible and consistent interactions across different tissue samples.