Testing for genetic interactions in complex disease with distance correlation

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: Hi, I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "Testing for genetic interactions in complex disease with distance correlation".

Marcus: Here's the gist of the paper. The gist Testing for genetic interactions in complex disease with distance correlation proposes distance correlation as a novel tool for detecting epistasis from case-control data of single-nucleotide polymorphisms (SNPs) to understand the genomic basis of common diseases like schizophrenia.

Ines: The role of heredity in psychiatry has been studied for almost a century, with Pearson (1931) not having “the least hesitation” in asserting its relevanceToday it is known that a majority of psychiatric disorders are multifactorial, complex traitsMore precisely, the genome can explain up to 80 % of the susceptibility to suffer some of these diseases, like schizophrenia (Sullivan et al., 2018)The genetic susceptibility to a psychiatric disorder lies on a large number of variants along the genomeAlthough the specialised literature usually focuses simply on additive models (International Schizophrenia Consortium, 2009), biological knowledge suggests that gene-gene interactions (or epistasis) could be oneWe are interested in datasets of case-control GWASs (i.e., a collection of genotypes of “healthy” individuals and “patients”) for schizophreniaThe statistical challenge hinges on using this data to detect pairs of genetic variants that significantly increase or decrease the susceptibility to develop schizophrenia, which further research can confirm with biological criteria Distance Correlation in Metric Spaces Distance correlation is an association measure that characterises general statistical independence between random variables, not only the linear oneLogically, distance correlation is defined as the quotient of variance and the product of standard deviations (as long as none of the latter vanish)It is an improved version of Pearson’s correlation squared because it has values in

0,1: and, more importantly, it is zero if and only if X and Y are (statistically) independentWhenever X and Y have strong negative type, dcov(X, Y) def.

Marcus: = dcov(θ) = 0 ⇒ X, Y independent Distance Correlation-Based Test for Epistasis The scope of this work will be address those questions, using an association measure named distance correlationThe aim is testing the independence of their joint sample (Zi,Zj) or, equivalently,(H0ij: dcov(Zi, Zj) = 0 H1ij: dcov(Zi, Zj) 6= 0 with the philosophy of the large-scale multiple tests by Cai (2017)In order to approximate the null distribution of the test statistic dcov([Zi, Zj), one can take advantage of finite marginal spaces, in which only a few of the coefficients of the quadratic form that gives the asymptotic null distribution of distance covariance will be non-nullThe most logical approach is, once again, resorting to permutation tests Simulations and Application The simulations show that the calibration of significance is acceptable or even good for low levels of nominal αIn all cases, R = 1000 replicates were carried outThe power is very satisfactory and that, as expected, it increases as one gets further away from the null hypothesisExperiment II indicates that we are detecting some genuine signal, at the genetically-regulated gene expression levelThe results of Experiment II indicate that we are detecting some genuine signal, at the genetically-regulated gene expression level Distance correlation has been shown to characterise independence in certain metric spaces, establishing how this theory is valid for the case of 3-point marginal spacesSimulations show that the calibration of significance is adequate and that power is considerably high against various alternativesWe also show that we outperform the “gold standard” BOOST (Wan et al., 2010) in the scenarios we have studiedThe schizophrenia database has been interrogated with our methodology, obtaining biologically sound results at the level of genetically-regulated gene expressionSome very recent studies show evidence of epistasis between regulatory regions of the human genome (Lin et al., 2022; Patel et al., 2022), which supports our findingsThe main limitation of our method (as it is of any other for this task) is that it is very difficult to make any solid discoveries when working with real dataFuture work on our research line may include investigating other distances, e.g.

Title and authors: Ines: We’re starting with the title and the authors of "Testing for genetic interactions in complex disease with distance correlation." This paper proposes distance correlation as a novel tool for detecting epistasis from case-control data of single nucleotide polymorphisms.

Marcus: The authors are Fernando Castro-Prado, Javier Costas, Dominic Edelmann, Wenceslao Gonzalez-Manteiga, and David R. Penas. They come from institutions like the Department of Statistics at the University of Santiago de Compostela and the German Cancer Research Center in Heidelberg.

Yuki: For population genetics purposes, having authors from both psychiatric genetics labs and computational biology centers suggests a good mix between clinical relevance and robust statistical methods for handling complex genetic architecture.

Ines: The implication here is that they are suggesting this distance correlation measure can be used to find pairs of genetic variants that significantly increase or decrease susceptibility to schizophrenia, which is what’s the main goal of testing epistasis.

Marcus: It’s about using these case-control GWAS datasets, which contain genotypes for healthy versus patients, to find interactions that explain why some people develop the disorder while others don't.

Yuki: This is directly addressing the question about how we can move beyond just additive models to understand the genomic basis of these common diseases.

The paper's summary: Ines: The paper summarizes that they are interested in datasets of case-control GWAS for schizophrenia, which involves single nucleotide polymorphisms, or SNPs, which are variants at one specific point on the genome.

Marcus: They specifically mention considering only autosomal variants because each individual can carry zero, one, or two copies of the minor allele. This helps define the space they are working in.

Yuki: The paper is motivated by a limitation in existing tools for epistasis detection; they’re interested in which association measures characterize the independence of ternary variables and how to extend techniques like LCTs to be applicable to SNP data under less stringent conditions.

Ines: So, the main summary is that they are using distance correlation as an association measure because it characterizes general statistical independence between random variables, not just linear dependence.

Marcus: They then introduce a novel testing procedure for independence in ternary data and organize the rest of the paper around summarizing the state of the art in characterizing independence in metric spaces.

Yuki: It’s essentially proposing distance correlation as a way to test for epistasis within this specific context, which is pretty novel if you think about how complex traits are usually studied.

The paper's improvements: Ines: They suggest that the main improvement they bring to the table is using distance correlation because it provides a measure of independence that is zero if and only if two variables are statistically independent, which is a strong property.

Marcus: They also discuss how to approximate the null distribution of this test statistic by leveraging finite marginal spaces where only a few coefficients of the quadratic form that gives the asymptotic null distribution will be non-null.

Yuki: The simulation studies reported in Section four show that they are detecting some genuine signals at the genetically-regulated gene expression level, which is interesting because it suggests their method can see signals beyond just missense SNPs <ref:2012.05285#pg1>.

Ines: They also report that the calibration of significance is acceptable or even good for low levels of nominal alpha, and they showed that the power is very satisfactory and increases as you get further away from the null hypothesis.

Marcus: Furthermore, they explicitly state that their method outperforms existing tools like BOOST in the scenarios they studied, which indicates improved detection capability for genetic interactions.

Conclusion: Ines: So to wrap up, "Testing for genetic interactions in complex disease with distance correlation" shows that distance correlation can characterize independence in certain metric spaces and that their simulation results show adequate calibration of significance and high power against various alternatives.

Marcus: They also confirm that the schizophrenia database was interrogated with this methodology, yielding biologically sound results at the level of genetically-regulated gene expression.

Yuki: For us in population genetics, what this means is that recent studies showing evidence of epistasis between regulatory regions of the human genome support these findings. It reinforces the idea that interactions matter even when we look at broader genomic features.

Ines: The main limitation they point out is that making any solid discoveries with real data remains very difficult for this task, which is a standard caution in this kind of research.

Marcus: They also noted that the method itself has limitations, but the overall performance against existing methods seems quite strong in the scenarios they tested.

Yuki: So, while it’s promising for detecting those interactions at the expression level and showing good statistical calibration, we still have to wait for more solid discoveries in real data before we can really say this is a definitive tool.

Ines: That’s it for this paper on "Testing for genetic interactions in complex disease with distance correlation." We’ll be looking at what comes next on our show.