HLA-DPB1*13:01 associates with enhanced, and KIR2DS4*001 with diminished protection from developing severe COVID-19

Abstract

Extreme polymorphism of human leukocyte antigens (HLA) and killer-cell immunoglobulin-like receptors (KIR) differentiates immune responses across individuals. Additional to T cell receptor interactions, subsets of HLA class I act as ligands for inhibitory and activating KIR, allowing natural killer (NK) cells to detect and kill infected cells. We investigated the impact of HLA and KIR polymorphism on the severity of COVID-19. High resolution HLA class I and II and KIR genotypes were determined from 403 non-hospitalized and 1,575 hospitalized SARS-CoV-2 infected patients from Italy collected in 2020. We observed that possession of the activating KIR2DS4*001 allotype is associated with severe disease, requiring hospitalization (OR=1.48, 95% CI 1.20-1.85, p_c=0.017), and this effect is greater in individuals homozygous for KIR2DS4*001 (OR=3.74, 95% CI 1.75-9.29, p_c=0.003). We also observed the HLA class II allotype, HLA-DPB1*13:01 protects SARS-CoV-2 infected patients from severe disease (OR=0.49, 95% CI 0.33-0.74, p_c=0.019). These association analyses were replicated using logistic regression with sex and age as covariates. Autoantibodies against IFN-α associated with COVID-19 severity were detected in 26% of 156 hospitalized patients tested. HLA-C*08:02 was more frequent in patients with IFN-α autoantibodies than those without, and KIR3DL1*01502 was only present in patients lacking IFN-α antibodies. These findings suggest that KIR and HLA polymorphism is integral in determining the clinical outcome following SARS-CoV-2 infection, by influencing the course both of innate and adaptive immunity.

1. Introduction

The symptoms of SARS-CoV-2 infection vary from none (asymptomatic) to multiple forms of coronavirus disease (COVID-19). Consequences include mild through severe disease, which can lead to respiratory failure and death^1-3. Diversity of COVID-19 outcome is caused by differential immune responses to SARS-CoV-2 infection, driven in part by polymorphism of immune-response genes^4,5. Class I and II human leukocyte antigens (HLA) guide immune responses by presenting peptide fragments to CD8⁺ T and CD4⁺ T cells, enabling infected cells to be killed and antibodies to be produced. Additionally, some HLA allotypes can interact specifically with killer cell immunoglobulin-like receptors (KIR) expressed by natural killer (NK) cells and some T cells⁶. HLA and KIR are the most polymorphic genes in the human genome. Their extreme variation affects resistance to infectious disease and susceptibility to autoimmunity, among other conditions^7-9. Combinatorial diversity of HLA and KIR has established effects on NK cell responses to respiratory infections, as demonstrated by H1N1 influenza and 2003 SARS-CoV^10,11. In this context, HLA and KIR polymorphism likely impacts immune responses to SARS-CoV-2 and subsequent disease outcome^12-14.

NK cells are lymphocytes that can recognize virus-infected cells and thereby influence the course of both innate and adaptive immunity^15-17. NK cell functions are modulated by the balance of inhibitory and activating signals transduced from multiple cell surface receptors, of which inhibitory KIR are the most polymorphic and have dominant effect^6,18. NK cells are educated during their development through ligation of inhibitory KIR, allowing them to respond to any HLA class I loss that may occur upon viral infection¹⁹. In addition to controlling acute viral infections, NK cells can also contribute to tissue pathogenesis. In severe COVID-19 patients, the NK cell gene expression profile is enriched towards pro-inflammatory cytokine production^20-22. In addition, peripheral NK cells are reduced in number during COVID-19 and impaired in patients having severe disease^22,23. Pro-inflammatory cytokines, including interferon-α (IFN-α), can also modulate NK cell function during SARS-CoV-2 infection²⁴. Genetically driven abnormal levels of IFN-α²⁵, as well the presence of autoantibodies against IFN-α have been associated with COVID-19 severity^26-29.

KIR and HLA are encoded by genes on separate chromosomes, and their combinatorial diversity thus depends on random segregation¹⁸. The KIR genomic region is characterized by gene presence/absence, copy number variation (CNV), and allele and haplotypic variability, any of which can impact NK cell functions⁹. KIR bind HLA ligands via their extracellular immunoglobulin-like domains (KIR2D have two, KIR3D have three) and those having long (L) cytoplasmatic tails transduce inhibitory signals. Transduction of activating signals by KIR having short (S) cytoplasmic tails is aided by accessory proteins³⁰. Interaction of KIR with their specific HLA ligands modulates NK cell functions, for example, KIR3DL2 is an inhibitory receptor that binds HLA-A*03 and A*11, and KIR2DS4 is an activating receptor that binds HLA-A*11 and a subset of HLA-C allotypes^9,18. Importantly, the interactions modulate a subset of NK cells having significant cytolytic activity and cytokine secretion^15,19.

Focused analyses have identified that the distinct HLA variants associating with COVID-19 severity differ across populations⁴. Although the first genome-wide association studies (GWAS) of COVID-19 did not identify HLA variability in disease susceptibility^31-33, a recent study in the UK showed, through imputation of HLA genotypes, that HLA-DRB1*04:01 protects against developing severe disease³⁴. On the other hand, KIR have remained largely intractable to GWAS studies. Nevertheless, presence and absence polymorphism has been reported associated with COVID-19 severity, and differential KIR expression levels by NK cells have been associated with specific clinical outcomes^35-37. Due to limitations of HLA genotyping methods or KIR presence/absence analysis, limited sample size, different population ancestries and clinical outcomes evaluated, the role of combined HLA and KIR polymorphism in determining COVID-19 severity remains unclear. Population-based studies in Italy^38-42 and one association study with disease severity in Sardinia³⁵ have reported that HLA diversity could impact on disease severity and susceptibility. The aim of this study is to analyze HLA and KIR diversity in hospitalized and non-hospitalized COVID-19 from a large Italian cohort and its potential impact on NK cell function following SARS-CoV-2 infection. Additionally, we evaluated the impact of presence of IFN-α antibodies in hospitalized patients.

2. Methods

2.1. COVID-19 cohort

We studied 1,978 SARS-Cov-2 infected subjects of differing disease severity belonging to the Italian GEN-COVID Multicenter study (https://sites.google.com/dbm.unisi.it/gen-covid). This is a network of more than 40 Italian Hospitals which, between March 2020 and February 2021, collected plasma, serum, and blood samples as well as clinical data for COVID-19 research. Patients infected by SARS-CoV-2 were diagnosed using RT-PCR from nasopharyngeal swab⁴³. Peripheral blood samples were collected in BD Vacutainer^® PPT^™ Plasma Preparation Tube K2E (EDTA). Plasma was isolated immediately using centrifugation and stored at −80°C in aliquots for subsequent analyses. DNA was extracted from EDTA peripheral blood samples using MagCore HF16 kit (Diatech Lab-Line, Jesi, Ancona, Italy) according to the manufacturer’s instructions. Written informed consent was obtained from all patients. The GEN-COVID study was performed in accordance with the relevant international, EU, national, and institutional guidelines and approved in advance by the University Hospital (Azienda Ospedaliero-Universitaria Senese) Ethical Review Board, Siena, Italy (Prot n. 16917, dated March 16, 2020). For association analysis, the cohort was divided into 403 patients having pauci-symptoms who were not hospitalized, and 1,575 patients who were hospitalized due to severe COVID-19. For the latter, a sub-category of critical patients (required CPAP/BiPAP and high-flow oxygen therapy, or intubation, or the outcome was death) was analyzed where indicated. The demographic data are shown in Table 1.

Table 1.

Clinical and demographic frequencies of cohort and association with COVID-19 severity (logistic regression).

	Non-hospitalized N=403	Hospitalized N=1575
	N (%)	N (%)	OR	CI (95%)	p
Clinical categories
1. Non-hospitalized	403 (100%)	-	-
2. Severe (a)	-	900 (57.1%)
(b)	-	675 (42.9%)
Sex
Female	242 (60.0%)	564 (35.8%)	Ref
Male	161 (40.0%)	1011 (64.2%)	2.67	2.14 – 3.35	<0.001
	N=403 F (%)	N=1573 F (%)
Age categories
<60 years	335 (83.1%)	611 (38.8%)	Ref
>= 60 years	68 (16.9%)	962 (61.2%)	7.89	5.99 – 10.52	<0.001
age 18-29	35 (8.7%)	28 (1.8%)	-
age 30-39	61 (15.1%)	77 (4.9%)
age 40-49	125 (31.0%)	148 (9.4%)
age 50-59	114 (28.3%)	358 (22.8%)
age 60-69	43 (10.9%)	372 (23.6%)
age 70-79	20 (5.0%)	331 (21.0%)
age 80-89	4 (1.0%)	202 (12.8%)
age >=90	0 (0.0%)	57 (3.6%)

N: number of patients, OR: odds ratio, CI: confidence interval, Ref.: reference group for association study.

The clinical categories were as follows: 1. non-hospitalized: asymptomatic or pauci-symptomatic, 2. Severe COVID-19 hospitalized patients a) who did not require oxygen therapy, or oxygen support only and b) patients who required CPAP/BiPAP and high-flow oxygen therapy, or intubation, or the outcome was death.

2.2. Sequencing HLA and KIR genes

HLA-class I and II and KIR genes were targeted for DNA sequencing using a well-established biotinylated DNA probe-based capture method⁴⁴, with modifications as follows. Genomic DNA (500ng from each sample) was fragmented enzymatically using the NEBNext^® Ultra II FS module (E7810, New England Biolabs, Boston, MA) followed by NEBNext^® Ultra II Ligation module (E7959, New England Biolabs, Boston, MA) with IDT xGen Stubby Adapters (Integrated DNA Technologies, Coralville, IA). Next, we performed post ligation cleanup and 800 bp DNA fragments were isolated by dual size selection with Agencourt AMPure XP beads (Beckman Coulter Life Sciences, Indianapolis, IN). Individual samples were labelled uniquely using 5 μl of 10 μM custom dual-index primers (Integrated DNA Technologies, Coralville, IA). Post amplification cleanup was based on the Kapa Hyper Prep protocol (Kapa Biosystems, Wilmington, MA). Equivalent quantities of each DNA library were pooled for enrichment. The KIR and HLA enrichment step⁴⁴ was modified as previously described⁴⁵. Paired ends of 250bp each were sequenced using a NovaSeq^™ 6000 instrument and SP Reagent Kit (Illumina Inc, San Diego, CA). The read depth was greater than 150X for all HLA and KIR genes. HLA alleles were determined from the sequence data using the consensus calls obtained from three algorithms: NGSengine^® 2.10.0 (GenDX, Utrecht, the Netherlands), HLA Explore^™ (Omixon Biocomputing Ltd. Budapest, Hungary) and HLA*LA⁴⁶. In > 99% of cases the three algorithms agreed at two fields of resolution. KIR copy number variation (CNVs) and alleles were obtained using the PING (Pushing Immunogenetics to the Next Generation) pipeline⁴⁷, which operates in R⁴⁸. We obtained 2 fields of resolution for all KIR and HLA alleles.

2.3. Association of HLA and KIR polymorphism with COVID-19 hospitalization

We first inferred the genetic ancestry of individual HLA-A-B genotypes using a model-based clustering algorithm, STRUCTURE v2.3.4⁴⁹ as described previously⁵⁰. We used HLA-A~B data from the 1000 Genomes^51,52 and the present cohort, with a burn-in length of 10,000, 20,000 MCMC resamples, and the number of assumed populations (K) was set to 3. HLA-C was not included in the analysis, since STRUCTURE expects independent genetic markers, and HLA-B and -C are in strong LD and could bias the inference. Two-sample non-paired Wilcoxon tests were applied to compare the distribution of ancestry proportions between the hospitalized and non-hospitalized groups. We performed HLA and KIR association by comparing allele frequencies of hospitalized with non-hospitalized patients, using the BIGDAWG package for R⁵³. For association studies, Fisher’s exact test, odds ratio (OR), 95% confidence interval (95% CI) and p-values were calculated for any HLA and KIR alleles and haplotypes having frequency > 2% across the cohort. We analyzed HLA class I genes: HLA-A, -B, -C, -E, -F, -G, and HLA class II genes: HLA-DRB1, -DQA1, -DQB1, -DPA1, -DPB1 for 403 non-hospitalized and 1,575 hospitalized patients. For KIR, we analyzed KIR2DL1-5, KIR2DS1-5, KIR3DL1S1, KIR3DL2 for 402 non-hospitalized and 1,569 hospitalized patients. KIR2DS4 alleles were combined as those expressing a functional receptor (KIR2DS4-full) or not (KIR2DS4-null)^54,55. HLA and KIR haplotypes were estimated using ‘haplostats’, included in the BIGDAWG package⁵³. The Bonferroni method was used to adjust for multiple tests, where p < 0.05 after correction (p_c) was considered significant. The p-values were corrected for 96 alleles for HLA and 57 alleles for KIR.

Any HLA or KIR genes having one or more alleles shown as significant on association analysis were then tested for sensitivity to confounders of sex and age > 60 years old using logistic regression analysis using the ‘glm’ function in R⁴⁸. As this is a sensitivity test, we did not correct for multiple testing⁵⁶. For any alleles identified through association analysis, we assessed if there was any effect due to ethnicity or recruitment location by including these factors as covariates using logistic regression.

2.4. Interactions of KIR with HLA class I

Where indicated, we analyzed specific KIR allotypes in combination with their respective HLA-ligands. HLA-A*03 and -A*11 are ligands for KIR3DL2⁵⁷. Any HLA-A or B allotypes carrying the Bw4 motif are ligands for KIR3DL1, with exceptions being A*25 and B*13⁵⁸. C1+HLA-C and some C2+HLA-C are ligands for KIR2DL2/3, and C2+HLA-C are ligands for KIR2DS1 and KIR2DL1^59,60. HLA ligands for KIR2DS2 are A*11, C*01:02, C*16⁶¹. HLA ligands for KIR2DS4 are C*01:02, C*02:02, C*04:01, C*05:01, C*14:02, C*16:01, and A*11:01⁵⁴. The ligand for KIR2DS3 is unknown, and the only KIR2DS5 allotype present in the cohort (2DS5*002) does not bind to HLA class I⁶². In analyzing the lineages of KIR3DL1, allotypes of the 015-lineage or 005 lineage were combined as described⁶³. KIR3DL1*004 and others sharing residue-86L were considered null alleles due to lack of cell-surface expression⁶⁴.

2.5. Detection of interferon-α specific antibodies in COVID-19 patients

Plasma from 156 hospitalized patients was subject to IFN-α specific autoantibody detection. The clinical and demographic data and results are presented in Table S1 and Table S2. Differences between subjects with or without IFN-α antibodies were compared using Pearson’s χ² using R, with p < 0.05 as the threshold for significance. The titer of antibodies against one or more of the 12 IFN-α subunits was measured using ELISA. Human recombinant proteins of IFN-α (Vinci-Biochem, Firenze, Italy) were coated onto flat-bottomed plates and incubated overnight at 4°C, with the following concentrations: IFNA1 5 μg/mL, IFNA2 5 μg/mL, IFNA4 10 μg/mL, IFNA5 8 μg/mL, IFNA6 20 μg/mL, IFNA7 10 μg/mL, IFNA8 8 μg/mL, IFNA10 5 μg/mL, IFNA14 8 μg/mL, IFNA16 20 μg/mL, IFNA17 20 μg/mL, IFNA21 10 μg/mL, all in 0.1 mol/L Na₂CO₃. Plasma was diluted at 1:500 in PBS containing 1% bovine serum albumin (BSA) and 0.05% Tween-20. Anti-human HRP (GE Healthcare, Milan, Italy) diluted 1:3000 was then added followed by tetramethylbenzidine (Tebu Bio, Magenta, Italy) incubation. Rabbit polyclonal antibody against human IFN-α (Vinci-Biochem) was included as appropriate positive control. The Delta values of human plasma were calculated by subtracting the optical density of coated wells from the uncoated wells to account for background signal. We compared HLA and KIR allele frequencies between patients with and without IFN-α specific antibodies, analyzing alleles that were > 2% in frequency (N=150).

2.6. Estimation of HLA binding affinity for SARS-CoV-2 Spike derived peptides.

We performed in silico prediction of the affinity for HLA class II allotypes of peptide fragments derived from SARS-CoV-2 Spike protein. We used netMHCIIpan v4.0, setting a < 0.5% rank for strong binders⁶⁵. The Spike protein of SARS-CoV-2 alpha (B.1.1.7) virus strain was used (UniProtKB P0DTC2)⁶⁶. Sliding windows for peptide fragments were set at 15 amino acids in length⁶⁵. Viable HLA class II allotypes^67,68 were subdivided into two groups: ‘broad’ or ‘limited’ presenters, with the threshold being the median number of distinct peptides predicted to bind across all HLA-DP, DQ or DR allotypes. Multivariate transformation and visualization was performed using the function 'prcomp' and package ‘factoextra’ (https://cloud.r-project.org/package=factoextra/); the two first principal components (PC) were visualized⁴⁸.

3. Results

3.1. HLA-DPB1*13:01 is associated with protection from severe COVID-19

To investigate the impact of HLA polymorphism on SARS-CoV-2 infection course, we analyzed a large cohort from Italy, collected in early 2020 and comprised of 403 individuals with mild disease (non-hospitalized) and 1,575 patients hospitalized with severe COVID-19. We analyzed 11 HLA genes, comprising highly polymorphic HLA class I (HLA-A, -B, -C), less polymorphic HLA class I (-E, -F, and -G) and HLA class II (HLA-DPA1, -DPB1, -DQA1, -DQB1, -DRB1). In this cohort, we observed 96 HLA alleles having greater than 2% frequency (Table S3). HLA-B is the most variable HLA class I gene, having 16 alleles, followed by HLA-C (13 alleles) and HLA-A (12 alleles). HLA-G has five alleles in the cohort and HLA-E and HLA-F each have two alleles. For HLA class II, HLA-DQB1 is the most variable gene with 12 alleles, followed by HLA-DRB1 (11 alleles), HLA-DQA1 and HLA-DPB1 (10 alleles), and HLA-DPA1 (3 alleles). No differences in genetic ancestry proportions were observed between the severe COVID-19 and non-hospitalized patients (Figure S1).

We compared the HLA allele frequencies of hospitalized with non-hospitalized COVID-19 patients. Regarding the less polymorphic HLA class I genes (HLA-E, -F, and -G), no significant frequency differences were found in relation to disease severity (Table S3). In analyzing highly polymorphic HLA class I, HLA-B*08:01 and -C*07:01 were more frequent in the non-hospitalized patients, whereas -C*07:02 was more frequent in hospitalized patients (Table S3). These differences were not significant after p-value correction (p_c > 0.05, Table 2). Of note, the frequencies of the HLA class I ligands for KIR did not differ between hospitalized and non-hospitalized patients (Table S3). Regarding HLA class II, HLA-DQB1*03:03 and HLA-DPB1*13:01 were more frequent in non-hospitalized than hospitalized patients (Table S3). However, only DPB1*13:01 remained statistically significant after Bonferroni correction (OR = 0.49, 95% CI 0.33-0.73, p = 0.0002, p_c = 0.019, Table 2). The results were stable when including sex, age, and ethnicity as covariates (OR = 0.50, 95% CI 0.32 - 0.80, p = 0.003, Table S4). In addition, HLA-DPB1*13:01 remained associated when adjusted by the hospital of recruitment (OR = 0.35, 95% CI 0.14-0.92, p = 0.032, Table S4). Lastly, we evaluated if HLA-DPB1*13:01 could protect from the most severe outcomes, including patients that need mechanical breathing assistance (CPAP/BiPAP), intubation, or when the outcome was death. Indeed, HLA-DPB1*13:01 was more frequent in non-hospitalized patients than in these critical cases (OR = 0.46, 95% CI 0.26 - 0.83, p = 0.010, Table S5).

Table 2.

HLA alleles and haplotypes showing frequency differences, comparing non-hospitalized with severe COVID-19 patients, in an Italian cohort.

	Non-hospitalized 2N=806	Severe 2N=3150
Alleles	Obs. (%)	Obs. (%)	OR	CI (95%)	p	pc
HLA-B*08:01	52 (6.5%)	142 (4.5%)	0.68	0.49-0.97	0.023	1.000
HLA-C*07:01	139 (17.2%)	433 (13.7%)	0.76	0.62-0.95	0.012	1.000
HLA-C*07:02	48 (6.0%)	253 (8.0%)	1.38	1.00-1.94	0.047	1.000
HLA-DPB1*13:01	41 (5.1%)	83 (2.6%)	0.49	0.33-0.74	0.0002	0.019
HLA-DQB1*03:03	36 (4.5%)	90 (2.9%)	0.63	0.42-0.96	0.020	1.000
HLA class II haplotypes
DPA1*01:03~DPB1*13:01	3 (0.4%)	5 (0.2%)	0.43	0.08-2.75	0.229	1.00
DPA1*02:01~DPB1*13:01	39 (4.8%)	75 (2.4%)	0.48	0.32-0.73	0.0002	0.002
DPA1*02:02~DPB1*13:01	0 (0%)	3 (0.1%)	-

OR: odds ratio, CI: confidence interval, -: not calculated, p_c: Bonferroni correction for 96 HLA alleles, 10 DPA1~DPB1 or 10 DQA1~DQB1 haplotypes analyzed. In bold, p_c < 0.05

HLA class II molecules are expressed as heterodimers, with specific alleles encoding the respective α and β subunits occurring in genomic proximity and having high linkage disequilibrium (LD). HLA-DPB1*13:01 can be observed in high LD with one of three -DPA1 alleles (-DPA1*02:01, -DPA1*02:02, or -DPA1*01:03)⁶⁹. The HLA-DPB1*13:01~HLA-DPA1*02:01 combination was more frequent in non-hospitalized than hospitalized patients (OR = 0.48, 95% CI 0.32-0.73, p = 0.0002, p_c = 0.002, Table 2). Concurrence of HLA-DPB1*13:01 with HLA-DPA1*01:03 was also more frequent in non-hospitalized than hospitalized patients, although in low frequency (0.4% vs 0.2%). The HLA-DPB1*13:01~DPA1*02:02 haplotype was not analyzed as it was present in only three patients (Table 2). As neither of the DPA1 alleles associated with disease state, these findings suggest DPB1*13:01 is driving the observed protection from severe COVID-19.

3.2. HLA-DPβ1*13:01 has distinct peptide and NK cell receptor binding properties

Antibodies against SARS-CoV-2 proteins help control infection and disease severity^26,29. As the differential peptide binding characteristics of HLA class II allotypes guide production of antibody specificity, we estimated the binding affinities of peptides derived from the Spike protein for each potential HLA-class II allotype (Table S6). HLA allotypes were divided according to the number of distinct Spike peptides they could present (broad, limited, or none). By this measure, the HLA-DPα1*02:01~DPβ1*13:01 heterodimer was classified as a broad binder (Table S6). The other two HLA-DPβ1*13:01 heterodimers observed in our cohort (DPα1*01:03 or DPα1*02:02) were also predicted to be broad Spike peptide binders (Table S6). Suggesting differential peptide binding contributes to differential disease risk, the Spike-derived peptide binding repertoire of DPβ1*13 clusters distinctly from that of other DP molecules (Figure 1). This finding suggests HLA-DPβ1*13:01-Spike peptide complexes encourage production of a distinct antibody specificity profile. Some HLA-DP allotypes can also bind NKp44, which is an NK cell activating receptor⁷⁰. Among these molecules HLA-DPα1*02:01~DPβ1*13:01 and DPα1*02:01~DPβ1*01:01 are the strongest NKp44 binders⁷⁰. As we saw no association with disease for HLA-DPB1*01:01 (Table S3), and interaction with Nkp44 is peptide-specific⁷⁰, this distinction could also be explained by the differential peptide repertoires of DPβ1*13:01 and DPβ1*01:01. Although it is important to note that the in silico analysis assessed peptide binding without regard to their immunogenic properties, these findings imply that DPβ1*13 can influence both antibody production and NK cell driven killing in response to SARS-CoV-2 infection. An alternative proposed mechanism is that DPβ1*13:01 may be able to resist downregulation by SARS-CoV-2 immunomodulatory ORF7a⁷¹.

Figure 1. HLA-DPβ1*13:01 is predicted to bind a distinct repertoire of SARS-CoV-2 Spike protein derived peptide fragments.

Shown is a principal component analysis (PCA) plot of the peptide fragments predicted to bind the HLA-DPβ1 allotypes indicated. DPβ1*13:01 is shown in red and DPβ1*01:01 in blue.

3.3. KIR2DS4*001 increases the risk of developing severe COVID-19

We analyzed 12 KIR genes KIR2DL1, KIR2DL2/3, KIR2DL4, KIR2DL5, KIR2DS1, KIR2DS2, KIR2DS3/5, KIR2DS4, KIR3DL1/S1 and KIR3DL2 to high resolution through sequencing. In our cohort, 57 alleles were observed having >2% allele frequency (Table S3). KIR3DL2 was the most variable gene observed, with 11 alleles, followed by KIR3DL1/S1 (8 alleles), KIR2DL4 (7 alleles), KIR2DS4 and KIR2DL1 (6 alleles), KIR2DL2/3, KIR2DL5A/B and KIR2DS3/5 (5 alleles), and KIR2DS1 and KIR2DS2 (2 alleles). Of the alleles identified, KIR2DL4*00102, KIR2DS3*00103, KIR2DS3*00201, KIR2DS4*00101, KIR3DL2*00201, and KIR3DL1*00201 were associated with increased risk of hospitalization, and KIR2DL4*00802, KIR2DS4*00601, KIR3DL1*00401 with decreased risk (Table 3). However only KIR2DS4*00101 remained associated with COVID-19 severity after correction for multiple testing (OR = 1.48, 95% CI 1.20-1.85, p_c = 0.017, Table 3). We therefore analyzed KIR2DS4 using logistic regression with age >60 and sex as covariates, which showed that carriers of KIR2DS4*00101 were significantly more frequent in hospitalized than non-hospitalized individuals (OR = 1.48, 95% CI = 1.14-1.93, p = 0.003, Table 4). The results were stable when including sex, age, and ethnicity as covariates (OR = 1.4, 95% CI 1.07 - 1.83, p = 0.014, Table S4). In addition, KIR2DS4*001 remained nominally significant when adjusted by the hospital of recruitment (OR = 1.62, 95% CI 1.00-2.66, p = 0.052, Table S4). Again, we evaluated if KIR2DS4*001 could associate with protection from the most severe outcomes, including patients that need mechanical breathing assistance (CPAP/BiPAP), intubation, or when the outcome was death. Indeed, KIR2DS4*001 was more frequent in non-hospitalized patients than these critical cases (OR = 1.51, 95% CI 1.10-2.08, p = 0.010, Table S5).

Table 3.

KIR alleles and haplotypes associated with COVID-19 severity, comparing non-hospitalized to hospitalized patients.

	Non-hospitalized 2N=804	Hospitalized 2N=3138
a) Alleles	F (%)	F (%)	OR	CI (95%)	p	pc
KIR2DL4*00102	109 (13.6%)	558 (17.8%)	1.38	1.10-1.74	0.004	0.228
KIR2DL4*00802	142 (17.7%)	466 (14.9%)	0.81	0.66-1.01	0.049	1.000
KIR2DS3*00103	94 (11.7%)	474 (15.1%)	1.34	1.06-1.72	0.014	0.798
KIR2DS3*00201	18 (2.2%)	124 (4.0%)	1.80	1.08-3.15	0.020	1.000
KIR2DS4*00101	120 (14.9%)	648 (20.7%)	1.48	1.20-1.85	0.0003	0.017
KIR2DS4*00601	148 (18.4%)	472 (15.0%)	0.78	0.64-0.97	0.019	1.000
KIR3DL2*00201	99 (12.3%)	526 (16.8%)	1.43	1.14-1.82	0.002	0.120
KIR3DL1*00201	80 (10.0%)	393 (12.5%)	1.30	1.00-1.69	0.045	1.000
KIR3DL1*00401	125 (15.5%)	384 (12.2%)	0.76	0.61-0.95	0.013	0.741
b) Genotype
2DS4*null/*null	289 (71.9%)	997 (63.5%)	Ref.
2DS4*full/*null	106 (26.4%)	477 (31.6%)	1.34	1.03 – 1.76	0.034	0.067
2DS4*full/*full	7 (1.7%)	76 (4.8%)	3.74	1.75 – 9.29	0.0017	0.003
c) Haplotypes
2DS4*00101~3DL2*00201	91 (12.2%)	523 (16.3%)	1.44	1.14-1.84	0.0019	0.019
2DL4*00102~3DL1*00201~2DS4*00101~3DL2*00201	72 (9.0%)	348 (11.1%)	1.27	0.97-1.68	0.080	1.000
2DL4*00102~3DL1*01502~2DS4*00101~3DL2*00201	23 (3.0%)	140 (4.5%)	1.52	0.97-2.47	0.061	1.000
2DL4*00102~3DL1*02901~2DS4*00101~3DL2*00201	0 (0.0%)	14 (0.4%)	-

F: allele or haplotype frequency, Ref. Genotype of reference, OR: odds ratio, CI: confidence interval, p_c: Bonferroni correction for 57 alleles; 2 KIR2DS4 genotypes, 9 KIR2DS4~KIR3DL2 haplotypes, or 10 2DL4~3DL1~2DS4~3DL2 haplotypes analyzed. In bold pc < 0.05

Table 4.

Logistic regression analysis of KIR gene candidates from screening study, using sex and age >60 years old as covariates.

	Non-hospitalized N=402	Hospitalized N=1569
Alleles	f (%)	f (%)	OR	CI (95%)	p
KIR2DS4*00101	113 (28.1%)	572 (36.5%)	1.48	1.14-1.93	0.003
KIR2DS4*006	133 (33.1%)	427 (27.2%)	0.74	0.57-0.96	0.021
KIR allotypes
KIR3DL1*00201/*01502/*02901	104 (25.9%)	520 (33.1%)	1.44	1.10-1.89	0.008
KIR2DS4*full	115 (28.6%)	574 (36.6%)	1.45	1.12-1.89	0.005
KIR2DS4*null	394 (98.0%)	1490 (95.0%)	0.32	0.14-0.66	0.004

f: carrier frequency. OR: odds ratio, CI: confidence interval. In bold p < 0.05. No p correction applied. The KIR2DS4*null genotype has 2 copies of alleles having a 22bp deletion (KIR2DS4*003; *004; *006, *010) or absence of the gene.

KIR2DS4*00101 encodes a receptor that is expressed on the cell surface (KIR2DS4-full allotype) whereas the remaining alleles observed in the cohort, KIR2DS4*003, *004, *006, and *010, share a 22bp deletion resulting in lack of cell surface expression (KIR2DS4-null allotype)^54,55. To evaluate if the presence or absence of functional KIR2DS4 molecules could impact disease severity, we divided KIR2DS4 alleles according to their expression ability. We observed an increase in susceptibility from none, through one, to two copies of a functional KIR2DS4, with two copies remaining significantly associated after multiple testing correction (OR = 3.74, 95% CI 1.75-9.29, p_c = 0.003, Table 3). This observation was repeated using logistic regression, where functional KIR2DS4 presence (risk) remained significant (OR=1.45, 95% CI=1.12-1.9, p=0.005, Table 4).

Because the KIR genes are densely packed and the receptors co-expressed⁷², we sought to identify any effect of LD on the association analysis. By example, KIR2DS4*00101 occurs in strong LD with KIR3DL2*00201⁷³, which was nominally associated with increased hospitalization risk (Table 3). Extending the analysis to KIR3DL1, which is located adjacently 5’ from KIR2DS4⁷², we observed that KIR3DL1 alleles *00201, *01502 and *02901 when linked to KIR2DS4*00101 were more frequent in hospitalized than non-hospitalized patients (Table 3). These three alleles belong to the 015-like lineage of KIR3DL1/S1⁶³, but no significant association was observed when we combined all alleles described for this lineage (Table S7). Importantly, the three distinct haplotypes associated with increased risk for hospitalization are characteristic to populations having distinct genetic ancestry background, 3DL1*00201 in Europeans, 3DL1*01502 in Asians and 3DL1*02901 in Amerindians⁹. This observation implies the 2DS4*00101 allotype association with severe COVID-19 is independent of genetic ancestry. Thus, these results suggest the KIR2DS4*00101 is the causal variant associated with increased risk of COVID-19 hospitalization, whereas the KIR3DL2 and KIR3DL1/S1 associations are driven by LD with this allele.

Ligation of activating KIR by HLA class I on the surface of infected cells can contribute to NK cell stimulation^6,15. Although KIR2DS4*001 remained significantly associated with severe COVID-19 in presence of either of its cognate ligands (OR = 1.41, 95% CI 1.04 - 1.91, p = 0.028, Table 5), we observed no change in the relative risk from determining this association without regard to ligands (OR = 1.48, Table 4). Indeed, we observed a similar odds ratio upon conditioning for the cognate ligands of KIR2DS4 (OR = 1.35, Table 5). Interestingly however, we observed that individuals possessing two or more activating KIR in combination with their respective HLA class I ligand were more frequent in hospitalized than non-hospitalized patients (OR = 1.94, 95% CI 1.36-2.80, p = 3.1x10⁻⁰⁴, Table 5). This latter finding implicates activating KIR in COVID-19 pathogenesis. Together, the findings suggest the association of KIR2DS4*001 with COVID-19 severity is independent of the known HLA class I ligands for KIR2DS4, or that the specificity of HLA class I ligation by KIR2DS4 is altered during infection, due to peptide repertoire changes ^74-76.

Table 5.

KIR + ligand combinations associating with severe COVID-19

Receptor + ligand combination	Non-hospitalized N=402	Hospitalized N=1569
	f (%)	f (%)	OR	CI (95%)	p
KIR2DS4*001
2DS4*00101 + ligand	73 (18.2%)	374 (23.8%)	1.41	1.04-1.91	0.028
2DS4*00101 + no ligand	40 (10.0%)	198 (12.6%)	1.35	0.92-2.01	0.133
KIR2DS1-4 + HLA ligands
Absent or 1 pair	357 (88.8%)	1278 (81.5%)	Ref
2 to 4 pairs	45 (11.2%)	291 (18.5%)	1.94	1.36-2.80	3.1x10 −04

By logistic regression association analysis, using sex and age >60 years old as covariates. f: carrier frequency. Ref: Allotype of reference, OR: odds ratio, CI: confidence interval. In bold p-values < 0.05. No p-value correction applied. HLA ligands for respective KIR are given in Methods.

3.4. Specific HLA and KIR variants may associate with IFN-α specific antibody production.

Autoantibodies against IFN-α can negatively impact infection clearance and contribute to severity of COVID-19^24,26. Among 156 severe COVID-19 patients analyzed in our cohort, 26% had autoantibodies recognizing one or more IFN-α subunits (Figure S2). No differences were found for any considered variables except for age, where antibody positive patients were younger (57.3 ± 15.0) than antibody negative patients (63.1 ± 14.2, p = 0.025; Table S1). Additionally, female patients produced autoantibodies that can recognize a greater diversity of IFN-α subunits when compared to male patients (p = 0.041; Table S2). Among these 156 patients, 150 had HLA and KIR genotype information.

In this cohort, we did not observe any HLA class II alleles to be associated with IFN-α specific antibody production (Table 6). Among HLA class I alleles, HLA-C*08:02 was enriched in patients who generated IFN-α antibodies, compared to those without them (OR= 5.09, 95% CI 1.16-25.1, p = 0.007). Additionally, the haplotype HLA-C*08:02 with HLA-B*14:02 was more frequent in IFN-α antibody positive than negative patients (8.3% vs 0.9%, p = 0.003, Table 6). Neither of these findings remained significant after correcting for multiple tests (p_c > 0.05). Of interest, KIR3DL1*01502 was present in 5.7 % patients lacking IFN-α antibody, but in none of the antibody positive patients (Table 6). The absence of this low frequency allele in the IFN-α antibody positive group could be due to sample size limitation (N=36). These results therefore suggest that specific HLA and KIR alleles may affect production of IFN-α antibodies, but the findings should be replicated in a larger cohort with multiple tests correction applied.

Table 6.

HLA and KIR alleles and haplotypes with presence of anti-Interferon-α levels in hospitalized COVID-19 patients.

	IFN-α negative antibody	IFN-α positive antibody
	N=114 2N=228	N=36 2N=72
	F (%)	F (%)	OR	CI (95%)	p	pc
Alleles
HLA-C*08:02	4 (1.7%)	6 (8.3%)	5.09	1.16-25.1	0.007	0.44
KIR3DL1*01502	13 (5.7%)	0 (0.0%)	-
Haplotype
HLA-B*14:02~HLA-C*08:02	2 (0.9%)	6 (8.3%)	10.97	1.87-116.48	0.003	0.05

OR: odds ratio, CI: confidence interval, pc: p-adjusted using Bonferroni correction for 88 HLA alleles and 17 HLA-B~HLA-C haplotypes analyzed.

4. Discussion

Presentation of pathogen-derived peptides by HLA, and their recognition by immune cell receptors, is critical for survival through driving both innate and adaptive immunity ^8,18,77. Although the exceptional polymorphism of HLA (and KIR) is driven by pathogen diversity, only rarely are specific alleles identified that protect from clinical infection by a given pathogen ⁷⁸. More commonly identified are those that, by comparison with other alleles, associate with differential disease severity following infection ^79,80. We therefore analyzed a cohort of individuals confirmed to be infected with SARS-CoV-2, to determine any impact of specific HLA or KIR alleles on the likelihood of developing severe COVID-19. We showed that HLA-DPB1*13:01 is associated with less severe disease outcome (not being hospitalized), whereas KIR2DS4*001 is associated with development of severe disease requiring hospitalization. These findings imply that specific variants of HLA and KIR have differential effect on the innate and adaptive responses to SARS-CoV-2 infection, as observed for other viral infections⁹.

As expressible forms of KIR2DS4, most commonly KIR2DS4*001, are found at relatively high frequency across populations worldwide^9,54, the findings in the Italian cohort are likely to be replicated in other populations. Indeed, we identify haplotypes containing KIR2DS4*001 that are rare in Europeans, but common in other populations, that are contributing to the disease risk. Although HLA-DPB1*13:01 allele frequency is low in the Italian cohort, this allele can reach up to 20% or 30% in East Asian and South American Amerindian populations, respectively⁸¹. Similar to our study, HLA-DRB1*15 has a well-established association with Multiple Sclerosis, and the frequency of this allele varies to the same degree as DPB1*013 across populations, but the association remains⁸². Thus, reporting associations with low worldwide allele frequency variants may stimulate replication studies in other cohorts, where these variants are more common and potentially have a higher impact.

We evaluated the HLA and KIR associations in a large cohort of patients from a single ancestry, whereas the majority of previous studies performed were (i) restricted to HLA-A, -B, -C, -DRB1, -DQB1⁷⁸, (ii) in small sample size cohorts⁴⁰, (iii) with distinct disease categories (i.e., patients versus population controls)³⁴, and/or (iv) in other ancestries⁸³. Moreover, no previous study analyzed KIR at high resolution. These factors render it difficult to compare across studies⁴ and, for these reasons, this is the first time that DPB1*13:01 and KIR2DS4*001 are being suggested to have roles in development of severe COVID-19. In other studies of Europeans, HLA-DRB1*04:01 associated with protection from severe disease³⁴ and from developing any symptoms⁷⁸. Although HLA-DPB1 variability was not evaluated in either study that identified DRB1*04:01, the lead SNP in the GWAS (rs9271609)³⁴ occurs in weak LD with DPB1*13:01 in a European ancestry population (D' = 0.032, r² < 0.001)⁸⁴. Thus, we tentatively suggest these other cohorts may replicate our findings.

Our finding that increased severity risk is associated with KIR2DS4 is consistent with the observed preferential expansion of NK cells expressing KIR2DS4 in severe COVID-19, which was observed in multiple studies, also irrespective of ligand presence^36,37,85. Together these results are compatible with a previous study suggesting individuals lacking specific inhibitory KIR/ligand combinations, whilst possessing activating KIR, were more prone to severe disease⁸³. Together with the correlation we observed of increasing activating receptor quantity with disease severity, these observations suggest the unusual expansion of KIR2DS4⁺ NK cells during severe disease could result in NK cells becoming impaired during SARS-CoV-2 infection⁸⁶, or that excess activation could lead to collateral damage.

Cytokine expression profiles have revealed distinct signatures of NK cells in COVID-19 severity, in which NK cell function, cytokine release, and pro-inflammatory pathways differ between moderate and severe cases^21,22,24. Among them, IFN-α pathway expression levels were enriched in NK cells of severe when compared to moderate COVID-19 patients. In addition, soluble IFN-α suppresses NK cell production of IFN-γ, affecting the immune response, even after decreasing IFN-α antibody titers in these patients²⁴. Antibodies neutralizing IFN-α, thereby preventing its activity towards SARS-CoV-2, were previously observed in 10.2% of COVID-19 patients with life-threatening disease²⁶. In our cohort, we observed that 26% of hospitalized COVID-19 patients had IFN-α specific antibodies in the plasma, albeit detected equally across severity subgroups. In addition, we observed the frequency of a haplotype containing HLA-C*08:02 and B*14:01 was elevated in patients having IFN-α autoantibodies. Although their role in autoantibody production is not clear, specific HLA class I allotypes could participate in presenting autoantigens to CD8+ T cells, indirectly activating B-cells, as observed in psoriasis⁸⁷. Due to the small sample size of the IFN-α autoantibody group, this result requires investigation in a larger study.

As a limitation of our study, we stratified our samples according to hospitalization status, dividing the patients with mild (or pauci-) symptoms that could be treated at home from those patients who needed hospital assistance. It is important to emphasize that the hospitalized group involves a variety of symptoms, as well presence of comorbidities, such as autoimmunity, cardiovascular diseases, diabetes type 2, and obesity. Due to lack of comorbidity data available for all the patients, we did not subdivide the cohort into smaller groups. It would be important to analyze a larger sample stratified by comorbidities to evaluate the impact of HLA and KIR polymorphism on severity. Moreover, we performed p-value correction using the Bonferroni method, considering all HLA and KIR alleles with frequency higher than 2%. In the HLA and KIR regions that both have high linkage disequilibrium this correction may be conservative, contributing to false-negative results (type II error) that might not fully resolve the complexity of HLA and KIR polymorphism in COVID-19 severity.

In summary, we show specific HLA and KIR variants have capacity to influence COVID-19 severity. In a large cohort from Italy, we observed that presence of HLA-DPβ1*13:01, which can interact with both CD4+ T cell and NK cell receptors, associates with protection from, whereas KIR2DS4, which stimulates NK cells, increases the risk of developing severe COVID-19. Our study highlights the complexity of HLA and KIR polymorphism in shaping the course of SARS-CoV-2 infection, through possible impact both on the innate and the adaptive immune response.