The research, led by Yadusayan Appulingam and colleagues at King’s College London in collaboration with the NYGC ALS Consortium, focused on two distinct sources of confounding that plague transcriptomic datasets. The first arises from unmeasured technical and biological heterogeneity within the expression data itself — things like batch effects, differences in tissue handling, or subtle variations in cell-type composition that are never directly recorded. The second source is more subtle: expression differences driven by population stratification, where genetic ancestry differences between cases and controls can create spurious gene-expression signals that have nothing to do with the disease itself.

Statisticians have long developed tools to tackle each of these problems independently. Surrogate variables, or SVs, are statistical constructs derived directly from the expression data that capture and remove unmeasured technical and biological noise. Genotype-based principal components, or PCs, on the other hand, are computed from genetic data and capture population structure — the way individuals cluster by ancestry. Both approaches are widely used, but until now, no study had directly evaluated what happens when they are combined within a differential expression framework. The researchers hypothesised that using both correction layers simultaneously would produce more biologically valid and reproducible results than either approach alone.

To test this, the team turned to two independent post-mortem RNA-sequencing datasets of ALS cases and controls with paired genotype data. The first, known as KCLBB, comprised 96 ALS cases and 52 controls from the London Neurodegenerative Diseases Brain Bank. The second, from the ALS Consortium, included 272 cases and 35 controls. The researchers constructed four nested differential expression models: one corrected only with genotype PCs, one corrected only with expression-based SVs, one corrected with both SVs and PCs, and one with neither correction. They then evaluated all four models across three metrics: cross-dataset effect size concordance, cross-dataset replicability quantified by the Jaccard Similarity Index, and biological recall against a curated reference set of 66 known ALS genes.

The results were striking. The combined SV-plus-PC framework outperformed the simpler models across most metrics. Replicability — the proportion of significant genes that appear in both datasets — improved nearly ten-fold compared to the non-corrected model, rising from a Jaccard index of 2.28 percent to 19.5 percent. The combined framework also exhibited a statistically significant 2.2 percent gain over the SV-only model. Biological recall, the ability to recover known ALS genes from the differential expression results, doubled compared to SV correction alone. Effect size magnitude consistency was preserved across datasets, though a reduction in Spearman’s rank correlation indicates that PC correction refines the magnitude of effects rather than reshuffling gene rankings.

These findings remained generally robust to the number of genotype PCs included and to sequencing platform adjustment. However, more variable performance was observed under ancestry restriction — when the analysis was limited to a single ancestry group — and upon extension to an independent tissue. This suggests that the combined approach is most valuable in ancestrally diverse cohorts, where population stratification is a greater concern, and that its benefits may be tissue-specific.

The researchers concluded that SVs and genotype PCs address non-redundant sources of confounding. In other words, they are not interchangeable — each captures a distinct type of distortion that the other cannot fully remove. The team recommends their combined use as standard practice in differential expression analysis wherever paired genotype data are available, with the greatest benefit expected in ancestrally diverse cohorts. Notably, PCs capturing population structure can also be derived directly from RNA-sequencing data, potentially extending this framework’s applicability to studies lacking paired genotype data.

Although this analysis was restricted to ALS datasets, the researchers expect these findings to generalise to other traits. The implications are broad: any study using differential expression analysis to identify disease-associated genes — from cancer genomics to psychiatric disorders — could benefit from this dual-correction approach. As transcriptomic datasets grow larger and more ancestrally diverse, the need for robust confounding correction will only intensify. This study provides a clear, evidence-based recommendation for how the field should move forward, and it may well change the way thousands of future gene-expression studies are conducted.

Subject of Research: Computational Biology and Bioinformatics

Article Title: Exploring transcriptomic and genomic latent variable correction approaches in differential expression analysis

Article References: Exploring transcriptomic and genomic latent variable correction approaches in differential expression analysis. (n.d.). https://doi.org/10.1038/s41598-026-69619-8

Image Credits: AI Generated

DOI: 10.1038/s41598-026-69619-8

Keywords: differential expression analysis, surrogate variables, principal components, transcriptomics, ALS, population structure, batch effects, Exploring, transcriptomic, genomic, latent, variable