Tissue Memory Score measures transcriptional retention in solidtumors and predicts survival in clear cell renal cell carcinoma

Abstract

Solid tumors retain transcriptional programs inherited from their tissue of origin, but the degree of this retention has not been quantified as a continuous metric, nor has its prognostic relevance been evaluated systematically. We developed the Tissue Memory Score (TMS) by projecting tumor bulk RNA-seq onto principal component axes constructed from 2,694 Genotype-Tissue Expression V8 samples across eight healthy tissues, and applied it to 3,900 primary tumors from eight Cancer Genome Atlas cohorts. Nearest-centroid classification recovered tissue identity in six of eight tumor types (80.1%-100.0% accuracy; chance 12.5%; all p < 10-). The two failures, gastric and endometrial carcinoma, reflected limitations of the reference atlas composition and not loss of tissue memory. Among cohorts passing pre-specified retention and dedifferentiation gates, TMS was independently associated with overall survival in clear cell renal cell carcinoma (ccRCC; multivariable HR = 0.59 per SD; 95% CI 0.46-0.77; p = 6.6 x 10-) and accounted for 16.3% of explained survival variance by Shapley decomposition. The association held under purity residualization, rank transformation, stage stratification, and progression-free interval as an alternative endpoint. In hepatocellular carcinoma, TMS showed a protective univariable association in the same direction, which did not survive multivariable adjustment and is reported as exploratory. External application to CPTAC-ccRCC yielded a pooled hazard ratio of 0.71 in the same direction (I{superscript 2} = 0%). Tissue-of-origin transcriptional memory is quantifiable in solid tumors and independently prognostic in ccRCC, but did not extend to the remaining seven cohorts. Broader applicability depends on whether the reference atlas represents the cellular compartment from which a given carcinoma arises.