Context-dependent correlations mislead transcriptomic network inference in bulk and single-cell data.

Area of Science:

• Genomics
• Bioinformatics
• Systems Biology

Background:

• Correlation analysis is foundational for co-expression module discovery and miRNA-target inference.
• These methods often assume a single, meaningful Pearson coefficient across heterogeneous samples.
• Simpson's paradox poses a theoretical challenge to this assumption due to potential between-group mean shifts.

Purpose of the Study:

• Quantify the frequency of correlation reversal due to Simpson's paradox in real transcriptomic data.
• Assess the impact of sample heterogeneity on pooled correlation estimates.
• Evaluate the necessity of reporting correlations within their specific biological contexts.

Main Methods:

• Analyzed 8,890 TCGA tumors across 31 cancer cohorts and 23,170,038 miRNA-mRNA pairs.
• Utilized GTEx and 10x PBMC scRNA-seq data to examine correlations across different biological modalities.
• Applied statistical measures including heterogeneity (I^2) and False Discovery Rate (FDR) to assess correlation consistency.

Main Results:

• 94.8% of miRNA-mRNA pairs exhibited both positive and negative within-cohort correlations.
• 13.3% of high-variance pairs showed a reversal in pooled correlation sign compared to the within-cohort majority.
• Significant heterogeneity (median I^2 = 0.86) and correlation inconsistency across cohorts were prevalent (99.5% rejected equality).
• Context refinement, such as using molecular subtypes, reduced correlation reversal rates.

Conclusions:

• Pooled correlation coefficients can misleadingly invert direction relative to within-context associations.
• Context-specific reporting of correlations, including distribution, heterogeneity, and mean shifts, is crucial.
• A provided R interface facilitates the computation of these essential correlation summaries.