A genome wide association study to identify germline copy number variants associated with cancer cachexia: a preliminary analysis

Procurement of samples and isolation of DNA

In this preliminary study, we utilized a subset of samples from our previously published candidate SNP study in cachexia^{10, 11} in which weight loss history and germline DNA were sampled (n = 1276) from patients at risk for cachexia with United Kingdom, Canada, Norway, Switzerland, and Greece as participating countries. We limited our investigation to two primary cancers (lung and gastro-oesophageal) for their relative higher prevalence of cachexia in patients.² We classified individuals based on the cachexia consensus definition of weight loss to stratify patients as cases and controls. This being a preliminary study, we selected individuals with extremes of weight loss of >10% over a period of 6 months (hereafter termed cases) for comparison with weight stable patients (<2% weight loss in 6 months, hereafter termed controls). A total of 179 patients met these criteria (74 cachectic cases and 105 controls) and were included in the study for genotyping. Complete demographics, clinical information, inclusion, and exclusion criteria of the study participants from multiple cancer types are provided elsewhere.^{10, 11} All patients provided written and informed consent for the study. DNA was extracted from buffy coat and was stored at −80°C until further use. The Health Research Ethics Board of Alberta (HREBA)- Cancer Committee approved the study protocol (# 17-0517).

Genotyping of samples using single nucleotide polymorphism 6.0 arrays and quality control

Affymetrix SNP 6.0 arrays contains probes for both SNPs and CNVs totaling 1.8 million genetic markers. As a part of quality control (QC), we used Contrast QC (CQC) to calculate the allelic (allele A and allele B) intensities into three genotype clusters (AA, AB, and BB) wherein the contrast distribution is calculated between the homozygote genotype peak and their shared valley with the heterozygote peak. Contrast QC (CQC) is a cluster-based algorithm and is a good predictor of sample genotyping performance. We met the default CQC as recommended by Affymetrix of ≥1.7 and the CQC for majority of our samples were consistently above 2.0, as described in our previous publications.^{12, 18, 19} All the quality control metrics utilized in the study is given in the user manual for reference (https://assets.thermofisher.com/TFS-Assets/LSG/manuals/gtc_4_2_user_manual.pdf). SNP genotyping calls from the array data were used to identify the genetic ancestry of the cohort we used. Principal component analysis (PCA) was performed using EIGENSTRAT to identify population structure using HapMap 270 genotype data as reference (European, Han Chinese or African as sub populations). The study participants were mapped to identify their genetic ancestry and samples that were not mapping to European ancestry were removed as outliers. PCA is a statistical method to address data dimensionality reduction, identify batch effects in genotype calls and adjust association tests using covariates, to increase interpretability and limit information loss. Sampled cases and controls may represent different ancestral populations (population stratification) and have different allele frequencies at a given (disease) locus, and if uncorrected, contributes to false positive associations. GWAS relies on identifying common variants for disease and phenotype associations on the assumption of unrelated and random sampling of populations unlike the genetic linkage studies, which rely on sampling from pedigrees. We therefore have also considered the possibility of hidden cryptic relationships (such as first-degree relative) among the study samples sharing genetic similarities. Such relationships can lead to an increase in the false positive associations. Identity by descent (IBD) analysis, widely used in GWAS, was performed to identify cryptic relatedness.²⁰

Identification of copy number variants

CNV analysis was performed using Partek Genomics Suite v6.6 [PGS, (Partek® Genomics Suite software, Version 6.6 beta, Copyright © 2009 Partek Inc., St. Louis, MO, USA)]. CNV intensity files from hybridized arrays were imported and the default parameters such as GC wave corrections were used. We used all sample normalization approach to create a reference baseline and to calculate the copy number estimates for each sample. Briefly, average hybridization intensities from all samples and all probes were treated as a representative of a diploid genome as described previously.^{12, 18, 21} To identify the copy number status and define the CNV boundaries, genomic segmentation was carried out using the default parameters in PGS: minimum consecutive genomic markers >10, P-value threshold = 0.001, signal to noise ratio = 0.3. A segment of CNV was called copy gain when the intensity value was above 2.3 and called copy loss if the intensity value was below 1.7, a range optimized for genomic segmentation algorithm in a diploid region. A copy number was assigned a diploid status when the intensity values fell in the range of 1.7 and 2.3. CNVs with P < 0.05 was considered significant (this P-value is distinct from the P-value in the genomic segmentation). The PGS output for CNV is given as contiguous segments with frequency and P-value calculated for each CNV call. The nature of the segmentation algorithm is that the CNV calls are made for stretches of DNA and the number of probes present along the linear length of the chromosome. As a result, some of the CNVs vary in length; multiple CNVs within the same genomic region were joined to form a contiguous segment and were referred to as a single copy number variable region (CNVR) as described in our previous studies.¹⁸ As CNVs harbour both coding and non-coding regions of the genome, we restricted our present analysis to map CNVs or CNVRs that showed 100% overlap with the protein coding genes for a detailed analysis. The non-coding genes were also catalogued, as these may confer regulatory roles and await further studies to identify their significance. Genome build hg19 was used as the reference genome for mapping CNVs. CNVs with a total frequency of 5% or more were defined as common CNVs.^{12, 18, 21}

In addition, we also mapped the significant CNV/CNVRs to 1000 Genomes Project phase 3 data and Database of Genomic Variants (DGV, http://dgv.tcag.ca/dgv/app/home) to interrogate if the identified CNV/CNVRs are recorded as common variants at a population level (distinct from de novo variants) in well annotated independent structural variation databases. Data deposited into such databases utilize independent CNV calling algorithms and independent genotyping platforms. Overlap of variants identified in this study to those from the curated databases may be viewed as an independent validation of CNVs identified in the current study. Those CNVs/CNVRs not mapping to these databases were considered as potentially novel variants that may require independent validation.

Biological interpretation of copy number variant embedded protein coding genes

CNVs that had a 100% overlap with protein coding genes at pre-defined cut-offs were considered for in silico pathway analysis (see results for defining cut-offs). To contextualize the protein coding genes to cachexia pathophysiology, in silico pathway analysis provides an overall understanding of the potential pathways that are involved in cachexia. These are curated databases where the genes associated with specific functions are binned together against which the gene from our current study is queried. It should be noted that each database has its strengths and limitations in terms of its functionality. We recognize the caveat that these in silico databases are updated constantly based on a few diseases, which is being extrapolated to most conditions.

Pathway analysis was performed using Metascape.²² Pathways with P < 0.05 were considered significant. To identify nodal molecules and its downstream targets, we performed network analysis using STRING database using CNVs, which had a 100% overlap with protein coding genes (https://string-db.org/). Genes that had more than five connections to its neighbours were used to build the interaction network. To generate a readable gene–gene interaction network, data obtained from STRING database were exported to Cytoscape v3.8 to generate the gene network.

To further understand the biological relevance of the identified CNV/CNVRs, we mapped the 803 unique protein coding genes obtained from 896 CNVs/CNVRs and compared them with the publicly available human skeletal muscle gene expression datasets. For this, we utilized differentially expressed genes reported in cancer cachexia studies from two datasets: GSE133979 and GSE18832.^{23, 24} Both these gene profiling studies utilized non-cancer controls and patients with cancer presented with weight loss. We only considered differentially expressed genes that had a fold change of >1.5 and P < 0.05.

Demographics of study participants and quality control of single nucleotide polymorphism 6.0 arrays

PCA identified five outliers (three cachectic cases and two weight stable controls), which were removed from further analysis (see principal component analysis, Figure 2) leaving 71 cases and 103 controls for further analysis. Identity by descent analysis revealed no cryptic relatedness in the samples.

Table 1 summarizes the patient demographics. There was no significant difference for age and sex between weight losing (cases) and weight stable cancer (controls) individuals. The weight losing group showed low BMI compared with weight stable cancer group and the differences were statistically significant. A mean of 15.8% weight loss was observed among cases; control participants did not undergo weight loss. Tumour type was also significantly different between the groups.

Characterization of copy number variants

After performing quality control, and calculating the copy number estimates, genomic segmentation for CNV with predefined cut-off led to the identification of 236 502 CNVs. Since CNVs may encompass both protein coding and non-protein coding genes, we interrogated for all the genes within CNVs (Figure 3). Approximately, 56% of the genes within CNVs were protein coding genes and the remaining were noncoding RNAs, which included miRNAs, lncRNAs, and piRNAs, among others. Of these, only miRNAs are well characterized for their gene regulatory functions and the roles of other non-coding RNAs are still evolving.^{12, 18} The length distribution of CNVs in protein coding region is given in Figure 3. (See Table S1 for a complete list of statistically significant CNVs associated across all 22 autosomes.) The majority of the CNVs fell between 1–5 kb length. CNVs in X, Y and mitochondrial (MT) chromosomes were filtered resulting in 227 074 CNVs, which were subjected for further filtering. At a statistical significance cut-off of P < 0.05, 10 923 CNVs were identified (any bp length); 10 801 CNVs were presented with >50 bp length. From this subset of CNVs, 5414 CNVs were retained when filtered for CNVs that showed 100% overlap with protein coding genes. Of these, 1583 CNVs had more than 5% frequency in the study population.

Further, contiguous genomic regions were merged from the 1583 CNVs into copy number variable regions (CNVRs).^{12, 18} CNVs that did not have contiguous regions were not merged and were called as CNVs. In all, 896 non-redundant CNV/CNVRs were identified, which embedded 803 unique protein coding genes. Copy gain and loss may occur in the same subject at a given locus in both cases and controls contributing to a redundancy of CNV counts. The cumulative copy number gain regions from 896 CNV/CNVRs, were 4208 and 2492 in cachectic cases and weight stable controls, respectively. The observed copy number loss regions were 4153 and 2774 in cachectic cases and weight stable controls, respectively. The observed differences in copy gain/loss in cases and controls were statistically significant (P = 0.0006, chi-square test); 743/896 CNV/CNVRs (~83%) overlapped with 1000 Genomes phase 3 data or DGV and the remaining CNVs may potentially be novel variants requiring further independent validations.

Copy number variant embedded genes are associated with pathways related to muscle wasting and metabolism

The top 30 CNV/CNVRs with the highest frequency identified in this study are presented in Table 2. To date, several of the identified CNV embedded genes have not been directly implicated in cachexia. A complete list of significant CNVs/CNVRs identified in this study are presented in Table S1 and Figure 4 and further data summaries pertaining to pathways and biological networks are represented in Figure 4. Pathways associated with protein synthesis and atrophy such as PI3K-Akt signalling, Jak–STAT signalling and FOXO signalling were identified (Figure 4). Representative upstream regulators include RBFOX1, JAK1, JAK2, JAK3, and LIFR (Figure 4). Many of these genes have been associated with muscle wasting in patients with cancer.

Copy number variant embedded protein coding genes are also differentially expressed in publicly available skeletal muscle datasets

We identified 52 genes that are common between CNV/CNVRs embedded genes (this study) and skeletal muscle gene expression datasets (GSE133979 and GSE18832).^{23, 24} Genes such as SPON1, CPT1B, SLC37A2, ARID5B, RBFOX1, ABLIM1, and GALNT15 (Tables 3 and S2) have not been reported to play a role in cachexia pathophysiology at the germline level. CNVs harbour genes spanning several loci and as such a subset of genes that are also expressed in skeletal muscle gene expression data sets are emphasized. For example, RASSF1 gene that was embedded in chr3:50344972-50454597 was differentially expressed. But CYB561D2 and CACNA2D2 that were present in the same CNV region were not differentially expressed in skeletal muscle. The complete list of CNVs/CNVRs embedded protein coding genes that are differentially expressed in presented in Table S2.