Batch Effect Correction in a Confounded Scenario: a Case Study on Gene Expression of Chornobyl Tree Frogs

Abstract

When large omics datasets present unwanted latent variability, a critical analysis step is to control these so-called batch effects properly. However, most batch effect-correction algorithms (BECAs) face limitations when the source of unwanted variation and the variable of interest are confounded. In this paper, we use RNA-seq data to study the effects of radiation contamination on tree frogs (Hyla orientalis) collected in the Chornobyl Exclusion Zone. We identify the site of collection of the frogs as a confounding factor in the transcriptomics analysis. We present our strategy to correct this confounding effect using the following BECAs: ComBat-seq, linear residualization, and Surrogate Variable Analysis. We show that the severe confounding between the site and radiocontamination level makes the correction step challenging. Instead, we investigate the site-to-site variability and successfully deconvolute the batch variable from the radiation level by adjusting for the population genetic structure. Our strategy allowed us to reveal the effects of low-dose radiation on the gene expression of Chornobyl tree frogs and appropriately preprocess the RNA-seq dataset for future multimodal integrative analyses.

A Appendix: Close-Up on the Impact of Sparsity in Gene Selection

To assess how forcing sparsity in PCA influenced the selected gene list, we also ran a similar approach using standard PCA. We performed PCA on 1000 bootstrap samples of the non-corrected (raw) variance-stabilized matrix. In each model, genes were ranked by the absolute value of their weight and the top 400 genes were selected from components 1 and 2. Genes stably selected across bootstrap iterations in components 1, 2, or both were submitted to gene functional annotation, as mentioned previously.

Table 4 shows that the genes selected using sparse PCA were more stable across bootstrap samples than with standard PCA. This led to identifying a larger number of deregulated pathways in the uncorrected dataset with sparse PCA than with PCA. In Fig. 5, we notice that the alteration of biological processes related to energy metabolism (GO terms “oxidative phosphorylation” or “energy derivation by oxidation of organic compounds”) was recovered with sPCA and not with PCA. The identification of pathways typically linked with low-dose radiation, despite the presence of batch effects, suggests that sparsity in the PCA weight vectors mitigated the influence of noise.

Table 4. Feature selection approaches and Gene Ontology terms enrichment

Fig. 5.

Enriched Gene Ontology-terms network after feature selection by PCA or sparse PCA on the uncorrected count matrices