Cross-Tissue Concordance of Blood-Trained Protein EpiScores in Donor-Matched GTEx Tissues

Abstract

Background: Protein EpiScores are DNA methylation (DNAm)-derived predictors of circulating protein levels that may provide accessible biomarkers of disease-related molecular variation. However, because these models were developed in blood, it remains unclear whether they preserve interindividual variation when applied to other tissues. We evaluated the cross-tissue concordance of blood-trained protein EpiScores in matched donors from the Genotype-Tissue Expression (GTEx) project. Methods: We applied 109 published protein EpiScores to DNAm data from whole blood and six matched non-blood tissues using MethylDetectR. The analysis included 54 unique donors contributing 158 blood-tissue pairs: transverse colon (n=45), lung (n=37), prostate (n=33), kidney cortex (n=24), breast mammary tissue (n=10), and ovary (n=9). Fifteen EpiScores had unusable variation, leaving 94 estimable scores per tissue and 564 protein-tissue comparisons. Positive concordance required Pearson and Spearman correlation coefficients to each be at least 0.50; strong concordance required both to be at least 0.70. Benjamini-Hochberg false-discovery-rate (FDR) correction was applied separately across all comparisons for each method. Leave-one-out analyses assessed sensitivity to individual donors, and partial Pearson correlations evaluated sensitivity to available covariates. Protein secretion categories, functional groups, and exploratory demographic and donor-cohort strata were also examined. Results: Ninety-seven of 564 comparisons (17.2%) met the positive-concordance criterion, and 73 of these (75.3%) met the leave-one-out stability criteria. Lung showed the broadest correspondence, with median absolute Pearson and Spearman correlations of 0.594 and 0.555, respectively; 56 of 94 lung EpiScores (59.6%) were concordant, 26 were strongly concordant, and 49 were stable. Other tissues showed more selective concordance, including 10 scores in transverse colon, three in prostate, two in kidney cortex, 15 in breast, and 11 in ovary. After covariate adjustment, median absolute partial r in lung was 0.522, and 68/94 scores had global q-values below 0.05. Among initially concordant pairs, 48/56 in lung retained positive partial r of at least 0.50 and q-values below 0.05, compared with 3/10 in colon, 1/2 in kidney cortex, and 0/3 in prostate. Breast and ovary were too small for adjusted analyses. Lung concordance extended across multiple target-protein secretion categories and functional groups. Exploratory donor-cohort differences did not survive FDR correction. Conclusions: Blood-trained protein EpiScores showed broad correspondence in lung that persisted after adjustment for available covariates. Initially concordant findings in the other tissues eligible for adjustment were mostly covariate-sensitive. These findings identify score-tissue combinations for independent replication and molecular validation and support evaluating cross-tissue correspondence separately for each score and tissue. Concordance between methylation-derived scores does not establish that blood EpiScores measure local tissue protein abundance.