Directional deconvolution of spatial conductance proxies resolves prognostic signal cancellation in oral squamous cell carcinoma

BackgroundTumor cells are increasingly understood as physically connected collectives whose intercellular communication is gated by gap junctions and modulated by microen-vironmental ion fluxes. While spatial transcriptomics provides the geometric substrate for building transcriptomic proxies of bioelectric organization, no robust pipeline currently translates spot-level connectivity features into independent clinical prognostic markers. MethodsWe analyzed 12 oral squamous cell carcinoma (OSCC) Visium sections (GSE208253). A K-nearest-neighbor (K=6) spatial graph was built on full-resolution coordinates and edge-weighted by a conductance-like transcriptomic proxy in which gap-junction proxy expression was scaled by an exponential acid-gating penalty. Geometric edge artefacts were controlled with concave-hull edge distance and partial rank correlation under permutation testing. A 25-gene BCI-Signature was extracted by intra-sample top/bottom conductance differential expression and cross-sample consensus voting ([≥] 6/12). The signature was spatially back-projected, directionally decomposed from prior biology, and then projected to TCGA-HNSC (n = 519) and GSE65858 (n = 270) for survival analysis. Cohort-level effects were combined by inverse-variance fixed-effect meta-analysis. ResultsDiagnostic controls falsified the initial isolation-driven hypothesis: across all 12 sections, the partial rank correlation between the isolation index and depolarization-footprint expression was negative after edge-distance adjustment. Feature ablation identified the conductance sum as the best transcriptomic proxy of physical network state, and section-level sensitivity analyses preserved the positive conductance-stress direction after long-edge removal and graph-parameter perturbation. Spatial back-projection showed that aggressive and differentiation programs are positively correlated within every section (median{rho} = 0.43) and co-enrich in high-conductance regions. This predicted bulk-level signal cancellation: the unweighted 25-gene mean was non-prognostic in TCGA-HNSC (HR=1.17, p=0.35), whereas the locked directional composite BCI_net was independently associated with worse OS (HR=1.38, 95% CI 1.06-1.79, p=0.015 after adjustment for age, stage, HPV status and gender). The effect persisted after separate adjustment for composition, EMT and proliferation proxies, but attenuated in a saturated all-proxy benchmark model. The biologically matched HPV-negative oral-cavity subset of GSE65858 (n = 77) preserved the direction with a larger effect size (HR=2.45, 95% CI 0.96-6.27, p=0.062). Inverse-variance fixed-effect pooling of the two cohorts yielded a significant pooled effect (HR=1.48, 95% CI 1.07-2.05, p=0.019). ConclusionsSpatial graph features can be transferred to bulk transcriptomic cohorts only after the structural and aggressive programs that co-localize within the same physical network are explicitly deconvolved. The equal-weight directional metric BCI_net is a biology-driven candidate prognostic readout that remains preliminary pending broader independent validation.