Abstract We report a failed empirical prediction and the methodological insight it produced. Using the Coherence-Decoherence Remodelling (CDR) framework, we predicted that spatial autocorrelation of gene expression — measured as Local Coherence Length (LCL) via variogram analysis — would be shorter in tumour tissue than in matched normal tissue, reflecting loss of regulatory coordination in cancer. We tested this prediction on 10x Visium spatial transcriptomics data from a treatment-naive NSCLC patient (E-MTAB-13530, De Zuani et al. 2024). The prediction was completely reversed: tumour tissue showed dramatically longer spatial autocorrelation than normal tissue across all four measured signalling channels (EMT, immune, proliferation, hypoxia), confirmed independently by Global Moran's I. Investigation revealed that the reversal reflects a fundamental scale mismatch. Standard-resolution spatial transcriptomics (Visium, 55 μm spot diameter) captures mesoscale pathophysiological organisation (hypoxic gradients, EMT fronts, immune exclusion zones) that emerges in tumours. It cannot resolve microscale developmental coordination (cell-cell junctions, basal-apical polarity, positional identity) that is maintained in healthy tissue. Healthy lung tissue is developmentally coordinated but spatially uniform at Visium resolution — no gradients to detect. Tumour tissue is developmentally decoherent but spatially structured at Visium resolution — abundant gradients from resource limitation, clonal expansion, and immune evasion. This scale mismatch is not specific to the CDR framework. Any study interpreting spatial autocorrelation in tumour atlases as a measure of tissue coordination risks the same conflation. We propose that the field distinguish two orthogonal properties: developmental coherence (microscale, decreases in cancer, requires single-cell resolution) and pathophysiological organisation (mesoscale, increases in cancer, captured by standard spatial transcriptomics). This distinction has implications for how spatial transcriptomics data from tumour atlases — including the Human Tumor Atlas Network, TRACERx, and disease-specific spatial studies — should be interpreted. Keywords: spatial transcriptomics, spatial autocorrelation, tumour microenvironment, developmental coordination, scale mismatch, variogram analysis, Local Coherence Length, Visium, cancer biology, methodology
Smith et al. (Thu,) studied this question.