Analysis of KIR gene variants in The Cancer Genome Atlas and UK Biobank using KIRCLE

Abstract

Background: Natural killer (NK) cells represent a critical component of the innate immune system's response against cancer and viral infections, among other diseases. To distinguish healthy host cells from infected or tumor cells, killer immunoglobulin receptors (KIR) on NK cells bind and recognize Human Leukocyte Antigen (HLA) complexes on their target cells. However, NK cells exhibit great diversity in their mechanism of activation, and the outcomes of their activation are not yet understood fully. Just like the HLAs they bind, KIR receptors exhibit high allelic diversity in the human population. Here we provide a method to identify KIR allele variants from whole exome sequencing data and uncover novel associations between these variants and various molecular and clinical correlates.

Results: In order to better understand KIRs, we have developed KIRCLE, a novel method for genotyping individual KIR genes from whole exome sequencing data, and used it to analyze approximately sixty-thousand patient samples in The Cancer Genome Atlas (TCGA) and UK Biobank. We were able to assess population frequencies for different KIR alleles and demonstrate that, similar to HLA alleles, individuals' KIR alleles correlate strongly with their ethnicities. In addition, we observed associations between different KIR alleles and HLA alleles, including HLA-B*53 with KIR3DL2*013 (Fisher's exact FDR = 7.64e-51). Finally, we showcased statistically significant associations between KIR alleles and various clinical correlates, including peptic ulcer disease (Fisher's exact FDR = 0.0429) and age of onset of atopy (Mann-Whitney U FDR = 0.0751).

Conclusions: We show that KIRCLE is able to infer KIR variants accurately and consistently, and we demonstrate its utility using data from approximately sixty-thousand individuals from TCGA and UK Biobank to discover novel molecular and clinical correlations with KIR germline variants. Peptic ulcer disease and atopy are just two diseases in which NK cells may play a role beyond their "classical" realm of anti-tumor and anti-viral responses. This tool may be used both as a benchmark for future KIR-variant-inference algorithms, and to better understand the immunogenomics of and disease processes involving KIRs.

Keywords: Immunogenomics; Killer immunoglobulin receptors; Natural killer cells.

Fig. 1

Description of the KIRCLE methodology. a Flowchart describing the 4 steps of the KIRCLE algorithm as it processes a single KIR gene (KIR2DL1 as an example here). Inputs are green, computations are blue, and outputs are gold. KIRCLE hyperparameters are listed in parentheses where they are implemented. b Depiction of step 4 of KIRCLE (thresholding). Allele probabilities generated by expectation-maximization may lead to a homozygous solution, a heterozygous solution, or no solution at all, depending on the user-selected value of the threshold hyperparameter t. In the depicted example, running KIRCLE with t = 0.5 (blue) would generate a homozygous solution (2 copies of KIR2DL2*003), whereas using t = 0.2 (green) would generate a heterozygous solution (1 copy each of KIR2DL2*002 and KIR2DL2*003). Using t = 0.8 (red) or t = 0.05 (purple) would have yielded no solution. c Depiction of one step of expectation-maximization. The initial allele-read matrix $M^{t_0}$ is collapsed into an expectation vector $E^{t_0}$ that is used to compute the next iteration of the matrix $M^{t_1}$. This process is repeated until the convergence criterion is satisfied, at which point the final expectation vector represents an estimate of KIR allele probabilities

Fig. 2

KIRCLE accuracy and consistency validation. a Contour plot demonstrating the effect of varying the bootstrap-proportion (p) and threshold (t) hyperparameters on KIR2DL1 allele inference, as measured by empirical calculation of the inferred genotypes’ entropy. b KIRCLE’s performance on KIR2DL4 allele inference was similarly characterized. c Fraction of each KIR gene’s KIRCLE-inferred allele genotypes that were called identically between 531 samples and their biological replicates in TCGA. Error bars represent the normal approximation confidence intervals. d TCGA sample coverages (binned) versus TCGA sample allele probability entropies for all 15 KIR genes. The allele probability entropies of a set of 20 “pseudo-BAMs” (green) are presented as negative controls

Fig. 3

KIR allele distributions in UK Biobank and TCGA. a Non-linear correlations between KIR allele frequencies among Caucasian individuals in TCGA versus those in UK Biobank, stratified by KIR gene. Error bars represent the standard deviation of correlation coefficients after re-sampling with replacement 100 times. b Comparison of allele frequency ranks between Caucasian individuals in TCGA and UK Biobank. c KIR2DL4 allele frequencies among Caucasian individuals in TCGA (left), UK Biobank (center), and US NMDP (right). d t-SNE plot of individuals in TCGA colored by participants’ ethnicities. Caucasian individuals were down-sampled by a factor of 8 for ease of visualization. e t-SNE plot of UK Biobank individuals colored by ethnicity. Caucasian individuals were down-sampled by a factor of 16 for ease of visualization

Fig. 4.

KIR allele associations with HLA alleles. a Bar plot depicting and listing KIR alleles that are significantly associated with HLA alleles at the FDR < 1e−10 level. b Volcano plot of KIR allele correlations with HLA alleles in UK Biobank. Associations are color-coded by the activity (inhibitory or activating) of the KIR allele. c Presence of at least 1 copy of HLA-A74 positively correlates with presence of at least 1 copy of KIR3DL3*005 (left) and presence of at least 1 copy of HLA-A36 negatively correlates with presence of at least 1 copy of KIR2DL3*002. d KIR t-SNE of UK Biobank individuals with those possessing HLA-B*42 highlighted in red. e Volcano plot of KIR allele correlations with HLA alleles among Caucasians in UK Biobank

Fig. 5

KIR allele associations with clinical correlates. a QQ-plot of KIR allele correlations with ICD10 diagnosis codes in UK Biobank among Caucasian individuals. b Odds of developing peptic ulcer disease is increased among those with the KIR3DL3*080 phenotype. c Bar plot showing decreased mean age of hay fever, rhinitis, or eczema in those with at least one copy of KIR3DL2*107. d Bar plot showing increased mean age of hay fever, rhinitis, or eczema in those homozygous for KIR3DL2*062. e Bar plot showing decreased mean age of atopy in those with at least one copy of KIR3DL2*107. f Bar plot showing increased mean age of atopy in those homozygous for KIR3DL2*062. All error bars in bar plots depict standard error of mean