Context, not contradiction: a pan-cancer framework for conforming and nonconforming co-expression programmes
Reports of a gene's role in cancer frequently conflict: the same transcript is described as an adverse prognostic marker in one tumour type and a favourable one in another, and such discordance is usually attributed to cohort differences or to irreproducibility. An alternative explanation is that the transcriptional company a gene keeps genuinely differs between lineages. That claim is testable, but testing it requires separating real lineage-dependence from three artefacts that pervade pan-cancer co-expression analysis: public summaries report only coefficients above a fixed threshold, a gene in a recurrently gained region acquires correlates through dosage alone, and conservation across cancer types is assumed rather than measured. We present a framework that addresses all three for any index gene. Correlations are recomputed threshold-free from each cohort's expression matrix; transcripts are partitioned into cis and trans sets relative to the index gene after retired symbols are resolved; a recurrence rule freezes the gene set before any downstream analysis; cohorts are scored against expression-decile-matched permutation nulls; and copy number is residualised out using the cis set as a built-in positive control, so the adjustment is shown to work before the trans result is read. Applied to PTK2, encoding focal adhesion kinase, across 17 TCGA cohorts, expression was associated with poor outcome in several cancers and better outcome in others, independently of tumour purity. The framework shows this is not noise: cohorts separate into conforming and nonconforming groups enriched for different biology. The 44-transcript trans programme is unchanged by copy-number adjustment while the cis control collapses by 38%, and co-expression is recoverable within individual tumours in 133 of 243 single-cell cases.