Using regulatory variants to detect gene-gene interactions identifies networks of genes linked to cell immortalisation

Abstract

The extent to which the impact of regulatory genetic variants may depend on other factors, such as the expression levels of upstream transcription factors, remains poorly understood. Here we report a framework in which regulatory variants are first aggregated into sets, and using these as estimates of the total cis-genetic effects on a gene we model their non-additive interactions with the expression of other genes in the genome. Using 1220 lymphoblastoid cell lines across platforms and independent datasets we identify 74 genes where the impact of their regulatory variant-set is linked to the expression levels of networks of distal genes. We show that these networks are predominantly associated with tumourigenesis pathways, through which immortalised cells are able to rapidly proliferate. We consequently present an approach to define gene interaction networks underlying important cellular pathways such as cell immortalisation.

Fig. 1. Employing a gene’s cis-eQTL complement to identify genetic interactions.

a SNPs within 1 Mb of a gene’s transcription start and termination sites were used to train an expression level prediction model using PredictDB. Predictions from these models correspond to the additive effect of all cis-regulatory variants for a given gene. b The observed expression of gene^A can then be modelled as an interaction (X) between its cis-eQTL (predicted expression) and the observed expression of other genes in the genome (gene^B). The error term represents, for example, uncaptured environmental or trans effects linked to variation in the expression of gene^A, not capture by these other terms.

Fig. 2. Model prediction performance and validation.

a R² of predicted versus observed gene expression levels in the Lothian Birth Cohort (LBC1936) against the rank of the R² value. Only the 5033 genes that passed the kurtosis threshold are shown. The horizontal line indicates an R² = 0.1, for which the 1205 genes exceeding this threshold were retained for further analysis. b Validation of the LBC1936 gene prediction models using the 1000 Genomes data. Genes with an R² > 0.1 across both datasets are plotted, with genes with an R² > 0.5 in both datasets labelled. Genes annotated twice have more than one probe on the array used in the LBC1936 dataset.

Fig. 3. Interactions can lead to genotype-dependent and regulatory variant-set-dependent variance in a gene’s expression levels.

Simulated data illustrating that a gene A’s expression is only highest when the cis-eQTL genotype is 1/1 and the expression level of a distal gene (gene^B) is low. b Replotting the data in a on different axes illustrates how this interaction is associated with genotype-dependent variance in the expression of gene^A, with greater variance in expression between individuals when the eQTL genotype is 1/1. c The interaction leads to genotype-dependent differences in the slope of the relationship between gene^A and gene^B’s expression levels. d–f The same principle extended to sets of cis-regulatory variants. P-values derived from Spearman’s rho test.

Fig. 4. Detecting regulatory variant-set veQTL.

An example regulatory variant-set-based veQTL identified at the SLFN5 locus is shown. a SLFN5 shows a strong correlation between its predicted and observed expression levels in the Lothian Birth Cohort suggesting this gene is under the strong control of nearby regulatory variants. b Plotting the residual after regressing out this cis-eQTL effect against predicted expression levels suggests that this gene also shows set-dependent variance in expression. c Correlating the gene’s predicted expression level to the square of the residuals from b confirms the presence of a veQTL at this locus. d–f the same as a–c but resulting from analysis of the 1000 Genomes data. P-values derived from Spearman’s rho test.

Fig. 5. Significant regulatory variant-set-dependent interaction between SLFN5 and DANCR.

Example of a significant set-dependent interaction, showing patterns consistent with the simulated data presented in Fig. 3. The expression of SLFN5 is linked to the combination of cis-regulatory variant-set and expression of DANCR. This is shown to be the case for the Lothian Birth Cohort (a–d) and is reproduced in the 1000 Genomes data (e–h), with both datasets exhibiting statistical significance (FDR < 0.05) and sign concordance for the interaction term coefficient. P and FDR values derived from ANOVA F-test.

Fig. 6. Binding of key EBV-associated transcription factors at the promoters of genes in the interaction networks.

The median binding levels of four EBV nuclear antigens (EBNA2, EBNA3A, EBNA3C and EBNALP) and five NF-κB subunits (cREL, p50, p52, RelA and RelB) at the promoters of interacting genes (i.e., interacting gene^Bs. Blue line) as well as genes expressed and tested for an interaction but for which one was not observed (red line). The grey shaded area around the red line represents the 95% confidence interval calculated by sampling 100 times the same number of genes from this background set as were in the foreground set of interacting genes. The processed binding data used in this plot was obtained from Jiang et al..

Fig. 7. The proportion of variance explained by additive and interaction effects between cis-regulatory variants and the expression of distal genes.

A given interaction explains on average a an additional 0.68% of the variation in the gene’s expression in the Lothian Birth Cohort, and b an additional 2.0% in the GEUVADIS dataset. c The total set of non-redundant interacting genes identified by stepwise regression explained 13.2% of the variation in gene expression in the Lothian Birth Cohort and 17.6% in the GEUVADIS dataset, with the interaction effects alone explaining 2 and 5% of the variation on average respectively (points are coloured by gene name).