Background

Genome-wide association studies (GWAS) have expanded the horizons of modern medicine by discovering numerous genetic variants associated with diseases. However, a major challenge remains: most disease variants are located in non-coding regions that do not directly encode proteins, making it difficult to reveal specific biological mechanisms.

To fill this gap, the academic community introduced expression quantitative trait loci (eQTL) analysis. This is an approach that statistically identifies the impact of genetic variation on tissue-specific gene expression levels. Furthermore, transcriptome-wide association study (TWAS) and colocalization analysis, which combine eQTL reference panels with GWAS summary statistics, have become key tools for identifying causal genes driving disease onset.

The problem lies in the foundation that guarantees the reliability of the analysis. While the academic community has applied strict standards to complex post-hoc statistical modeling and multiple testing corrections, it has been relatively lax regarding the upstream preprocessing stages of ribonucleic acid sequencing (RNA sequencing, RNA-seq), which serve as the starting point for analysis. It has overlooked the fact that downstream research results can be destabilized depending on which transcriptome annotation database is used to convert raw sequence reads into gene expression levels, and whether alignment-based or pseudo-alignment-based quantification is chosen.

Key Findings

An international research team demonstrated in a paper published on October 26 in the journal Nature Genetics that the choice of upstream RNA-seq processing steps can significantly impact genomic analysis. The team systematically compared combinations of transcriptome annotation systems and expression quantification pipelines using identical genomic and transcriptomic raw data.

Clear discrepancies emerge starting from the baseline setting of transcriptome annotation. When applying different annotation versions with varying GENCODE, Ensembl, or exon definition ranges, the assigned gene-specific expression levels for the same sequencing reads changed significantly. In particular, quantification distortions were prominent in genomic regions where multiple transcript isoforms coexist or gene boundaries overlap.

These preprocessing biases are directly propagated to the eQTL mapping stage. The composition of statistically significant eQTL variant-gene pairs fluctuated significantly due to differences in quantification algorithms alone. The concordance rate of eQTL discovery between traditional alignment-based quantification tools and algorithms applying transcript deconvolution fell below expectations.

Downstream analyses, such as TWAS and colocalization, are also directly affected. It was observed that the ranking of candidate genes identified as key causal mediators of disease risk reversed simply by changing the upstream preprocessing method. Even the posterior probability of colocalization, which evaluates whether GWAS signals and tissue-specific expression signals originate from the same variant, reached opposite conclusions depending on the preprocessing choice. The initial choice of data processing essentially dictated the final results of disease gene identification.

Implications and Outlook

This study issues a stark warning to the functional genomics research ecosystem. It demonstrates that even when advanced statistical models are applied, subtle biases introduced during the raw data preprocessing stage can distort the overall conclusions of the research.

The researchers urged the establishment of specific methodological reporting standards to ensure data reproducibility in future genomic studies. They pointed out that researchers must move away from the practice of simply listing analysis pipeline names and instead specify in detail the build version of the transcriptome annotation used and the quantification parameters. A sensitivity verification procedure, which compares multiple preprocessing combinations rather than relying on a single pipeline, was also identified as an essential element.

In the long term, the academic community anticipates that the widespread adoption of long-read sequencing technology will serve as a turning point in mitigating annotation incompleteness. This is because it can fundamentally reduce errors in isoform protein quantification by directly decoding the entire transcriptome without fragmentation. Attention is now focused on whether discussions regarding the standardization of algorithms and reference databases will intensify.