Directional Gene-Level Concordance and Methodological Constraints in Blood Transcriptomic and DNA Methylation Studies of Parkinson’s Disease

Assessing reproducibility across different molecular profiling studies is a persistent methodological challenge (Zhang et al., 2009; Sweeney et al., 2017; Ioannidis, 2005). Differences in platform technology, cohort composition, analytical pipelines, and feature definitions often make it difficult to interpret cross-study comparisons based solely on gene-identity overlap.

In this study, we conducted a retrospective computational analysis of seven publicly available analytical datasets (including alternative analytical pipelines applied to the same cohort) derived from five biologically independent peripheral blood transcriptomic and DNA methylation cohorts, comprising 3,487 samples (1,824 Parkinsons disease cases and 1,663 controls). Reproducibility was evaluated using gene-identity overlap, enrichment-based comparisons, and a permutation-based framework to assess directional consistency of effect estimates across datasets. We also tested the robustness of results by varying false discovery rate thresholds and applying alternative probe-to-gene collapsing strategies. All analyses were performed using reproducible workflows implemented in R and Python with fixed random seeds.

Across independent cohorts, gene-identity overlap was generally limited, with enrichment ratios close to one, especially when datasets were generated using different platforms. In several datasets, limited numbers of statistically significant features further constrained overlap-based comparisons. In contrast, directional consistency showed greater stability. High levels of directional consistency were observed across independent cohort comparisons when restricted to overlapping statistically significant features and remained stable across statistical thresholds (90.0% at FDR < 0.05 and 82.8% at FDR < 0.10). When evaluated across the full shared gene universe without conditioning on statistical significance, directional consistency was substantially lower (∼30 to 32%) but remained significantly above permutation-based null expectations. Permutation testing confirmed that the observed directional consistency exceeded what would be expected by chance. A combined analysis including methodological replicates (n ≥ 3 datasets) showed 98.3% directional consistency; however, this estimate includes non-independent analytical pipelines applied to the same cohort and reflects analytical stability rather than independent biological replication. Rather than introducing a new statistical method, this study examines how commonly used reproducibility metrics behave under crossstudy heterogeneity and identifies their practical limitations and appropriate use boundaries.