Interactions between amino-acid-defined MHC class II variants and smoking for seropositive rheumatoid arthritis

Abstract

Objective

This study aimed to refine the interaction between cigarette smoking and human leukocyte antigen (HLA) polymorphisms in seropositive rheumatoid arthritis (RA), in the context of a recent amino-acid based HLA model for RA susceptibility.

Methods

We imputed HLA amino acids and classical alleles from case-control Immunochip array data of 3,588 Swedish Epidemiological Investigation of Rheumatoid Arthritis (EIRA; case/control=1654/1934), 589 Nurses’ Health Study (NHS; 229/360) and 2,125 Korean (1390/735) subjects. We examined interaction effects between heavy smoking (>10 pack-years) and genetic risk score (GRS) from RA-associated amino-acid positions (11, 13, 71 and 74 in HLA-DRβ1; 9 in HLA-B; and 9 in HLA-DPβ1) and from HLA-DRβ1 four-amino-acid haplotypes with an attributable proportion due to interaction (AP) using the additive interaction model.

Results

Heavy smoking and all investigated HLA amino-acid positions and haplotypes were associated with RA susceptibility in all three populations. In the interaction analysis, we found a significant deviation from the expected additive joint effect between heavy smoking and the HLA-DRβ1 amino-acid haplotype in all three studies (0.416≤AP≤0.796). We further identified the key interacting variants as being located at amino-acid positions 11 and 13 of HLA-DRβ1 but not the other RA-risk amino-acid positions in all populations. At the positions 11 and 13, there were similar patterns between RA-risk effects and interaction effects of residues.

Conclusion

Our findings of significant gene-environment interaction effects implicate that a physical interaction between citrullinated auto-antigens produced by smoking and HLA-DR molecules is characterized by the HLA-DRβ1 four-amino-acid haplotype, primarily by the positions 11 and 13.

INTRODUCTION

Rheumatoid arthritis (RA; OMIM180300) is a complex chronic inflammatory disease associated with multiple genetic and environmental factors. The strongest RA susceptibility gene is for alleles at the HLA-DRB1 gene (1–3), the so-called “shared epitope”, and the influence of HLA genes are largely to the RA subset characterized by the presence of antibodies to citrullinated peptides (ACPAs) (4–7). Cigarette smoking is the most established RA environmental risk factor, with accumulating evidence demonstrating its risk effect as well as dose-response effect, also mainly for seropositive (ACPA and/or rheumatoid factor (RF)-positive) RA risk (8–13) (14–16). Since 2004, several studies conducted in multiple populations with distinct ethnic backgrounds have consistently identified a significant interaction between smoking and shared epitope alleles for seropositive RA susceptibility based on the additive interaction model (4,12,17–21). Analyses have further revealed a dose-dependency of this synergistic effect as the most pronounced association and interaction estimates have been repeatedly observed among heavy smokers (>10 pack-years) carrying two copies of shared epitope risk alleles, compared to non-carrier never smokers. This suggests the importance of considering the cumulative “dose” for both smoking and major histocompatibility complex (MHC) class II genes when testing for their interactions (12,14–16,19). These findings have led to the hypothesis that citrullinated peptides produced by smoking are accommodated by HLA-DR molecules in an allele-specific manner and that presenting the citrullinated peptides by HLA-DR with shared epitope-associated alleles is integrally related to seropositive RA pathogenesis (4,22).

The shared epitope hypothesis, originally described by Gregersen et al (1) defines causal alleles among classical 4-digit HLA-DRB1 alleles that encode a common conserved amino-acid sequences (QKRAA, QRRAA or RRRAA) at position 70 to 74 in the third hypervariable region of the HLA-DR β-chain. For many years, it was difficult, however, to separate effects from the structure defined by the amino acids constituting the shared epitope and other structures within MHC class II molecules, due to the strong linkage disequilibrium (LD) within the MHC region (23). Several studies have indicated that different MHC structures are involved in defining the susceptibility to ACPA-positive RA. In addition to effects outside of HLA-DRB1 (6,24), opposing (protective) effects have been described from some HLA-DRB1 non-SE alleles, such as *03, *07, *08, *11, *12, and *13, in the development of seropositive RA (25–30). Likewise, an increasing number of studies have demonstrated significant interactions between certain HLA-DRB1 non-SE alleles (*09:01 in Asians and *13:01 in Caucasians) and smoking (21,28,30). Taken together these studies demonstrate that the effects of MHC class II genetic variants as well as the interactions between smoking and MHC class II genes in conferring risk for RA are complex.

The redefinition of the structures of MHC class II molecules involved in the development of ACPA-positive RA in 2012, by Raychaudhuri et al (2) identified the strongest association signal for seropositive RA susceptibility at the HLA-DRβ1 amino-acid position 13 (and its tightly correlated position 11) whereas the traditional shared epitope (defined as positions from 70 to 74) as well as in HLA- DRβ1, position 9 in HLA-B and position 9 in HLA- DPβ1 all conferred independent but weaker effects. This amino-acid based HLA model was initially identified in Caucasian ancestries was later shown to be applicable also to Asian populations since the consistent association was observed in a large Asian study on ACPA-positive RA (3).

The use of the structure-based amino acid model (2) for analysis of how environmental risk factors interact with different variants of MHC class II molecules may allow better understanding of possible molecular mechanisms responsible for the development of different subsets of RA. In particular, theses analyses may provide us with clues to which MHC class II structures are involved in presentation of neoantigens, for example citrullinated peptides, which may be generated from environmental exposure to smoking in the lungs and elsewhere (31). In the present study, we aimed to explore the interaction of smoking with the MHC class II structures defined by the variable amino-acid positions in three HLA genes: HLA-DRB1, -B and -DPB1 by using three independent populations from Sweden, US and Korea.

PATIENTS AND METHODS

Subjects

We utilized genetic and epidemiologic data of 3,588 Swedish Epidemiological Investigation of Rheumatoid Arthritis (EIRA), 589 Nurses’ Health Study (NHS), and 2,125 Korean Observational Study Network for Arthritis subjects. The study subjects were surveyed for smoking behavior prior to RA onset and genotyped by Immunochip, the high-density genotyping array for immune disease loci including the HLA region.

EIRA is a population-based case-control study initiated in 1996 and is still actively enrolling subjects. In the current study, we included all individuals recruited until 2009, consisting of 1,654 incident ACPA-positive RA cases and 1,934 controls available for both genetic and environmental information. Cases were recruited from all rheumatology providers in defined areas in Sweden with more than 85% recruited within one year after first symptoms onset and at first visit with a rheumatologist. Controls, matched by age, gender and residential area, were randomly selected from the Swedish National Register. At baseline, participants were asked to fill out a self-administrated questionnaire, as well as to provide blood samples for serological and genetic examinations. For the current analysis, the cumulative quantity of cigarettes smoked (pack-years) before the RA onset (or up to the index year among controls) were used.

Nurses’ Health Study and Nurses’ Health Study II are prospective cohort studies in the United States in which data on lifestyle and behavioral factors and disease outcomes are collected every two years by questionnaire. In a nested case-control study, 229 Caucasian female, seropositive (positive for RF or ACPA) incident RA cases and 360 controls were selected for genotyping on Immunochip. For this study, we used smoking information about quantity (pack-year) surveyed just prior to RA onset for the cases and matched time for the controls. The Korean subjects consist of 1,390 RA cases and 735 controls recruited from Hanyang University Hospital for Rheumatic Diseases in Seoul, Korea (21) and genotyped by Immunochip. All cases were positive for ACPA (disease duration at enrollment: mean 8.43 years; standard deviation: 8.44). Controls were recruited from healthy volunteers by a campaign at public places and the hospital, enriched in females aged >30 years to get the similar age range and gender ratio observed among cases. Clinical information was collected directly by trained interviewers using questionnaires. All control volunteers were examined for past medical history and those with a personal or family history of autoimmune diseases including rheumatic diseases were excluded. Smoking information was retrospectively recorded at RA onset for cases or at enrollment for controls.

The institutional review boards of all participating sites approved these studies. The baseline characteristics including smoking information of each cohort are shown in Table 1. The Immunochip data have been reported by Eyre S et al (32) for EIRA, by Karlson EW et al (33) for NHS, and by Kim K et al (34) for Korean.

Table 1.

Baseline characteristics of RA cases and controls from the Epidemiologic Investigation of Rheumatoid Arthritis (EIRA), Nurses’ Health Study (NHS), and Korean ACPA-positive RA or seropositive RA case control studies

Characteristics	EIRA (n=3588)		NHS (n=589)		Korean (n=2125)
	ACPA-positive RA cases	Controls	Sero-positive RA cases	Controls	ACPA-positive RA cases	Controls
Subjects (n)	1654	1934	229	360	1390	735
Age, mean (SD)	51.08 (12.36)	54.22 (11.17)	54.33 (8.23)	54.42 (8.13)	44.02 (12.52)	34.67 (9.98)
Female (%)	70.44	73.47	100.00	100.00	84.17	93.74
Smoking characteristics
Ever cigarette smokers (%)	72.95	59.09	62.01	58.33	18.13	7.62
Pack-years among smokers, mean (SD)	17.37 (14.56)	14.59 (19.93)	23.56 (15.83)	22.10 (21.20)	21.82 (18.34)	8.78 (10.52)
Smokers with > 10 pack-years (%)	45,76	29.03	46.72	35.83	11.80	1.50

HLA imputation

HLA imputation was performed using SNP2HLA software according to published instructions (35). For Caucasian datasets (EIRA and NHS), the dosage of HLA classical alleles and amino acids were imputed from the Immunochip data with a publicly released reference panel generated from 5,225 Caucasians by the Type 1 Diabetes Genetics Consortium (35). For the Korean dataset, imputation was performed with a recently developed reference panel constructed from 413 Koreans (36). Long-range haplotypes across the extended MHC were obtained, where we extracted the haplotypes defined from HLA-DRβ1 amino-acid positions 11, 13, 71 and 74, as well as RA-associated amino-acid positions in HLA-DRβ1, -B and -DPβ1 for subsequent analyses.

Calculation of genetic risk score for HLA-DRβ1 amino-acid haplotypes and RA-associated amino-acid positions in HLA-DRβ1, -B and -DPβ1

To summarize the RA-risk effect of the previously characterized HLA-DRβ1 amino-acid haplotypes constructed by the amino-acid positions 11, 13, 71 and 74, we calculated a genetic risk score (GRS) for individual haplotypes by weighing each haplotype on its reported RA-risk effect size obtained from Raychaudhuri et al (2) for Caucasians and from Okada et al (3) for Koreans using the following equation:

$${\mathit{GRS}}_i=\sum _{h=1}^n(\text{ln}O{R}_h)·{\mathit{haplotype}}_{hi}$$

where GRS_i is a haplotype GRS of individual i, OR_h is the reported odds ratio (OR) of the h^th haplotype and haplotype_hi is the number of the h^th haplotype in individual i.

We also calculated GRS for each of the amino-acid positions 11, 13, 71 and 74 in HLA-DRβ1, position 9 in HLA-B and position 9 in HLA-DPβ1 by the equation:

$${\mathit{GRS}}_{ki}=\sum _{a=1}^n(\text{ln}O{R}_{ak})·{\mathit{dosage}}_{\mathit{aki}}$$

where GRS_ki is a GRS of amino-acid position k in individual i, OR_ak is the reported OR of the a^th residue at amino-acid position k and dosage_aki is an imputed dosage value of the a^th residue at amino-acid position k in individual i.

The calculated GRS for haplotypes and amino-acid positions (which were continuous) were dichotomized at the median value of the controls in each population for a subsequent interaction analysis.

RA association tests

Pack-years of smoking was dichotomized at 10 pack-years: heavy smoking > 10 pack-years vs. non-heavy or never smoking ≤ 10 pack-years (11,12,14–16,19). Heavy smokers were compared with non-heavy/never smokers. The main effect of heavy smoking was calculated by logistic regression models with adjustment for age, gender and residential area in EIRA, for age in NHS, and for age and gender in Koreans.

The RA associations of overall residue difference at each of the reported RA-risk amino-acid positions were evaluated by a log-likelihood ratio test that estimates log-likelihood difference between null and full models. The null logistic regression model included the top five principal components (PCs) as predictors for RA risk. The full logistic regression model included all residues, excluding the most frequent residue, and the same five PCs. To calculate the omnibus P values for the combined dataset from the three datasets, a dummy variable indicating the three cohorts was additionally included in both null and full logistic regression models. The association of dichotomized GRS was examined by logistic regression adjusting for the top five PCs.

Interaction tests

The interaction effects between heavy smoking and haplotype GRS as well as between heavy smoking and amino-acid position GRS were assessed using the attributable proportion due to interaction (AP) to determine if the joint effect between the two factors is significantly departed from additivity (19,37). For testing the interaction between heavy smoking and the dichotomized GRS for HLA-DRβ1 amino-acid haplotypes, we calculated disease odds ratios in three groups (with high GRS/heavy smoking, low GRS/heavy smoking, and high GRS/non-heavy smoking) with a reference group (with low GRS/non-heavy smoking) by using the following logistic regression model adjusting five PCs and continuous GRS at position 9 in HLA-B and -DPβ1 as covariates:

$$log(\mathit{odds})={\beta }_0+\sum _{i=1}^3{\beta }_{\mathrm{i}}{G}_{\mathrm{i}}+\sum _{\mathrm{j}=1}^5{\pi }_{\mathrm{j}}{PC}_{\mathrm{j}}+{\pi }_6{\mathit{GRS}}_{B\_9}+{\pi }_7{\mathit{GRS}}_{\mathit{DPB}1\_9}+\epsilon$$

Where G₁, G₂, and G₃ are indicate variables for high GRS/non-heavy smoking, low GRS/heavy smoking and high GRS/heavy smoking groups. PC_j is the j^th PC, and GRS_{B_9} and GRS_{DPB1_9} are continuous GRS of HLA-B and -DPβ1 amino-acid position 9, respectively. β₀ is the logistic regression intercept. β_i and π are the effects of G_i and covariates (PC_j, GRS_{B_9} and GRS_{DPB1_9}). Similarly, to test the interactions between heavy smoking and the dichotomized GRS for the RA-associated amino-acid positions in HLA-DRβ1, -B and -DPβ1, we conditioned on five PCs and other amino-acid position GRS to remove GRS correlation effects generated by linkage disequilibrium (LD) among the amino-acid positions. For the two tightly correlated amino-acid positions at 11 and 13 of HLA-DRβ1, we did not adjust each position for the other position in the AP model. For example, when testing for HLA-DRβ1 amino-acid position 11, we conditioned on GRS from position HLA-DRβ1 amino-acid positions 71 and 74, HLA-B amino-acid position 9 and HLA-DPβ1 amino-acid position 9.

As an alternative method to remove the possible LD bias between the amino-acid positions, we residualized each continuous GRS of the tested amino-acid position by the other continuous GRS of the other amino-acid positions and the top five PCs using linear regression by the following equation:

$$R_k={\mathit{GRS}}_k-\widehat{{\mathit{GRS}}_k}$$

where GRS residual of amino-acid position k (R_k) is calculated by subtracting the predicted GRS ( $\widehat{{\mathit{GRS}}_k}$; e.g. for HLA-DRβ1 amino acid position 11, ${\widehat{\mathit{GRS}}}_{\mathit{DRB}1\_11}={\beta }_0+{\sum }_{j=1}^5{\beta }_j{PC}_j+{\beta }_6{\mathit{GRS}}_{\mathit{DRB}1\_71}+{\beta }_7{\mathit{GRS}}_{\mathit{DRB}1\_74}+{\beta }_8{\mathit{GRS}}_{B\_9}+{\beta }_9{\mathit{GRS}}_{\mathit{DPB}1\_9}$) from the actual GRS (GRS_k), considered as a proportion not explained by the RA-risk effect (GRS) at the other positions in a regression model. We performed the same interaction analysis without covariates by using the dichotomized GRS residuals as a complement

Since each amino acid position harbors different residues, we finally investigated the residue effects of HLA-DRβ1 amino-acid positions 11 and 13 in the interaction with heavy smoking in RA. Each of the two amino-acid positions can be occupied by one of 6 possible residues. We coded each residue for each individual as a binary variable according to its presence (one residue or two residues) or absence (no residue). Since some residues have protective effect on RA susceptibility, we alternatively recoded these RA-protective residues to set up a correct reference combination of both genetic and epidemiologic factors in AP calculation, as previously suggested (38), and their APs were converted to the opposite direction to interpret them appropriately. All analyses were performed by PLINK, R and SAS.

RESULTS

Association of smoking with RA in the study subjects

Smoking status and intensity (pack-years of exposure) until the time of RA onset were obtained from the study populations. The smoking behavior varied considerably by population, especially, comparing ever-smoking rates among the Korean population where overall rates (7.6% in Korean controls) were much lower than the Caucasian populations (59.1% in EIRA controls and 58.3% in NHS controls) (Table 1). The overall low smoking rates in the Korean study reflects the high percentage of females who tend to be non-smokers for cultural reasons (Supplementary Table 1). For pack years of smoking, we saw similar population differences. In EIRA, we identified 754 heavy smoker and 893 non-heavy/never smokers among cases and 556 heavy smokers and 1,359 non-heavy/never smokers among the controls. In NHS, we identified 107 heavy smokers and 122 non-heavy/never smokers among RA cases and 129 heavy smokers and 231 non-heavy smokers among the controls. In Korean, we identified 164 heavy smokers and 1,226 non-heavy/never smokers among RA cases and 11 heavy smokers and 724 non-heavy smokers among controls.

Heavy smoking, a strong epidemiologic risk factor, was consistently associated with RA susceptibility in EIRA, NHS and Korean subjects (1.62 ≤ OR ≤ 4.73; 6.48×10⁻³ ≤ P ≤ 1.60×10⁻³⁴) (Table 2). Although rates of heavy smoking were much lower in the Korean subjects than the Caucasian subjects, the heavy smoking effect on RA was larger in the Koreans (Table 2). We note three important observations about the association of smoking with RA susceptibility on our study, in order to address potential concerns. First, to investigate whether there was a selection bias resulting in low ever-smoking rate among the Korean controls compared to Korean cases, we obtained the smoking status in Health2 Study (39), a community-based cohort study where the 2,987 participants aged 40 – 69 were recruited by Korea National Institute of Health (KNIH). We found very similar ever-smoking rates (3.2% in females and 76.4% in males vs. (3.0% in females and 76.1% in males) in this independent cohort of general subjects in Korea (Supplementary Table 1). Second, heavy smoking effects upon RA risk were similar in males and females in EIRA and Koreans (NHS includes females only). Third, when smoking was categorized as never and ever smokers, the association became weaker in EIRA and Koreans, and disappeared in NHS (data not shown).

Table 2.

Association between heavy smoking and ACPA-positive or sero-positive RA risk in EIRA, NHS and Korean cohorts

Population	Smoking intensity (pack-years)		OR* (95%CI)	p-value
	0–10	>10
EIRA
Controls	1359 (70.97)	556 (29.03)	1.0	ref.
ACPA-positive RA	893 (54.22)	754 (45.78)	2.52 (2.18–2.92)	1.60E-34
NHS
Controls	231 (64.17)	129 (35.83)	1.0	ref.
Sero-positive** RA	122 (53.28)	107 (46.72)	1.62 (1.14–2.29)	6.48E-03
Korean
Controls	724 (98.50)	11 (1.50)	1.0	ref.
ACPA-positive RA	1226 (88.20)	164 (11.80)	4.73 (2.27–9.85)	3.28E-05

^*OR were adjusted for age, gender and residential area in EIRA; for age in NHS; and for age and gender in Koreans.

^**NHS cases defined as seropositive based on positive rheumatoid factor in charts since 1976, positive anti-CCP2 in charts since 1990, or positive anti-CCPs in subset of subjects with plasma samples.

Associations of the reported HLA amino-acid positions with RA in the study subjects

We imputed HLA classical alleles, amino-acid residues and SNPs within the extended MHC region from the Immunochip data of 3,588 EIRA (case/control=1,654/1,934), 589 NHS (229/360) and 2,125 Korean (1,390/735) subjects, and achieved high concordance rates between the imputed and previously genotyped HLA-DRB1 alleles in all three populations (EIRA, concordance rate for 2-digit *04 alleles: 97.3%, 4-digit *04 alleles: 95.0%; NHS, 2-digit: 93.8%, 4 digit: 91.8%; Koreans, 2-digit: 97.4%, 4-digit: 91.4%), indicating a high quality of the imputed data.

From the imputed results, the most significant association for RA risk was found within HLA-DRB1 in the extended MHC region in all three populations (Supplementary Figure 1). We assessed the association of amino-acid positions 11, 13, 71 and 74 in HLA-DRβ1, position 9 in HLA-B and position 9 in HLA-DPβ1 that were reported to have independent RA-risk effect in Caucasian and Asian populations (2,3). Consistent with previous reports, the most significant associations were found at the amino-acid positions 11 and 13 before and after conditioning on other amino-acid positions in each of the three populations and a single analysis (P_omnibus = 7.17×10⁻¹⁶⁹ at the position 11 and P_omnibus = 2.67×10⁻¹⁶⁶ at the position 13 in an unconditional analysis) (Table 3). Relatively weak or no associations at the other positions in HLA-DRβ1, -B and -DPβ1 were found in each population. However, independent associations at these positions were observed after conditioning on other position’s effects in the combined analysis (P_omnibus ≤ 2.74×10⁻³; Table 3).

Table 3.

Association of the known RA-risk amino-acid positions with ACPA-positive or sero-positive RA susceptibility in the EIRA, NHS and Korean cohorts

HLA genes	Amino acid Positions	Number of residues	EIRA		NHS		Korean		All**
			Pomnibus	Conditional Pomnibus*	Pomnibus	Conditional Pomnibus*	Pomnibus	Conditional Pomnibus*	Pomnibus	Conditional Pomnibus*
HLA-DRB1	11	6	3,43E-97	5,47E-22	1,03E-11	5,59E-04	1,41E-43	4,53E-29	7,17E-169	5,32E-58
HLA-DRB1	13	6	6,45E-95	3,82E-19	1,74E-11	4,04E-04	4,35E-42	8,83E-26	2,67E-166	1,15E-49
HLA-DRB1	71	4	1,38E-44	3,27E-05	2,76E-04	0,035	5,98E-06	0,282	4,12E-54	3,35E-08
HLA-DRB1	74	5	5,63E-19	9,44E-04	3,84E-03	0,266	1,36E-10	1,23E-03	8,49E-34	7,88E-13
HLA-B	9	3	1,68E-08	4,47E-03	0,397	0,209	2,03E-03	0,040	0,507	2,74E-03
HLA-DPB1	9	3	2,96E-09	1,03E-05	0,409	0,844	0,394	0,938	4,62E-09	1,67E-04

^*The effect of each amino acid position was conditioned on other amino acid positions. HLA-DRB1 positions 11 and 13 were not conditioned on each other due to their tight correlation. All the analyses were adjusted for top 5 PCs from each population.

^**For the combined dataset, a dummy variable indicating the three cohorts was additionally included in the logistic regression.

Interactions between HLA-DRβ1 amino-acid haplotypes and heavy smoking

The haplotype defined by the amino-acid positions 11, 13, 71 and 74 is the best model to explain the HLA-DRB1 association with RA susceptibility (2). We found that RA susceptibility was strongly associated with the dichotomized haplotype GRS in all three populations (Supplementary Table 2).

We evaluated interaction effect between the dichotomized haplotype GRS and heavy smoking and found significant interaction effects between the two factors in all three populations (AP = 0.416, 95%CI = 0.306 to 0.526, P_AP = 1.16×10⁻¹³ in EIRA; AP = 0.467, 95%CI = 0.187 to 0.748, P_AP = 1.07×10⁻³ in NHS; and AP = 0.796, CI = 0.535 to 1.058, P_AP = 9.73×10⁻¹⁴ in Koreans; Table 4). We note that the number of heavy smokers in the Korean control group subjects was too small to get a stable AP value and might cause an overestimated AP in the Korean population.

Table 4.

interaction effect between HLA-DRβ1 haplotype genetic risk score (GRS) and heavy smoking in ACPA-positive or sero-positive RA susceptibility in EIRA, NHS and Korean cohorts

Population	Pack-year	GRS level*	Cases	Controls	OR (95%CI)	AP(95%CI)** and P
EIRA	0–10	Low	161 (19.05)	684 (80.95)	1.0 ref.	0.416 (0.306–0.526) p=1.16E-13
	>10	Low	128 (31.30)	281 (68.70)	1.96 (1.49–2.57)
	0–10	High	732 (52.03)	675 (47.97)	4.38 (3.51–5.46)
	>10	High	626 (69.48)	275 (30.52)	8.49 (6.62–10.89)
NHS	0–10	Low	38 (25.50)	111 (74.50)	1.0 ref.	0.467 (0.187–0.748) p=1.07E-3
	>10	Low	26 (27.37)	69 (72.63)	1.09 (0.60–1.99)
	0–10	High	84 (41.18)	120 (58.82)	1.80 (1.11–2.92)
	>10	High	81 (57.45)	60 (42.55)	4.42 (2.59–7.57)
Korean	0–10	Low	232 (39.19)	360 (60.81)	1.0 ref.	0.796 (0.535–1.058) p=2.41E-9
	>10	Low	43 (84.31)	8 (15.69)	8.09 (3.71–17.63)
	0–10	High	994 (73.20)	364 (26.80)	3.66 (2.92–4.59)
	>10	High	121 (97.58)	3 (2.42)	61.49 (19.05–198.48)

^*HLA-DRβ1 Haplotype was constructed from the amino-acid positions 11, 13, 71 and 74. Haplotype GRS was calculated by weighing each haplotype on its reported RA-risk effect size. The calculated GRS were dichotomized at the median in controls of each population.

^**Dichotomized two main factors, GRS and smoking pack-year, were used in the AP model adjusting for HLA-B amino-acid position 9 GRS, HLA-DPB1 amino-acid position 9 GRS and five principal components.

Interactions between RA-risk HLA amino-acid positions and heavy smoking

We further evaluated interaction effects between heavy smoking at the HLA amino-acid position level using the dichotomized GRS for each RA-risk amino-acid positions in HLA-DRβ1, -B and -DPβ1. To minimize potential correlation between the amino-acid positions due to their LD, we added continuous GRS of non-tested amino-acid positions as covariates. Significant and consistent synergy was mapped to HLA-DRβ1 amino-acid positions 11 and 13 in all three populations (0.324 ≤ AP ≤ 0.721, 1.64×10⁻⁴ ≤ P_AP ≤ 1.45×10⁻⁵; Table 5). In contrast, inconsistent interaction effects were found at the other positions, including HLA-DRβ1 amino-acid positions 71 and 74 within the traditional shared epitope, in all three study populations. When we performed the same interaction analysis using the dichotomized GRS residual, we also observed significant AP at either HLA-DRβ1 position 11 or 13 in both EIRA and NHS (Supplementary Table 3).

Table 5.

Additive interaction between HLA-DRβ1 amino acid genetic risk score (GRS) and smoking in ACPA-positive or sero-positive RA susceptibility in EIRA, NHS and Korean cohorts

Amino acid GRS	Pack-year	GRS level*	Cases	Controls	OR (95%CI)	AP(95%CI)** and P
EIRA
DRB1_P11	0–10	Low	123 (17.65)	574 (82.35)	1.0 ref.	0.333 (0.183–0.484) p=1.45E-05
	>10	Low	98 (28.99)	240 (71.01)	1.92 (1.41–2.62)
	0–10	High	770 (49.52)	785 (50.48)	2.30 (1.76–3.01)
	>10	High	656 (67.49)	316 (32.51)	4.26 (3.16–5.73)
DRB1_P13	0–10	Low	115 (17.14)	556 (82.86)	1.0 ref.	0.324 (0.172–0.477) p=2.98E-05
	>10	Low	95 (28.79)	235 (71.21)	2.00 (1.45–2.75)
	0–10	High	778 (49.21)	803 (50.79)	2.39 (1.82–3.12)
	>10	High	659 (67.24)	321 (32.76)	4.30 (3.18–5.80)
DRB1_P71	0–10	Low	237 (27.56)	623 (72.44)	1.0 ref.	0.158 (−0.046–0.361) p=0.129
	>10	Low	232 (45.49)	278 (54.51)	2.13 (1.67–2.71)
	0–10	High	656 (47.13)	736 (52.87)	1.37 (1.07–1.75)
	>10	High	522 (65.25)	278 (34.75)	2.84 (2.13–3.80)
DRB1_P74	0–10	Low	315 (31.50)	685 (68.50)	1.0 ref.	0.26 (0.065–0.456) p=9.09E-03
	>10	Low	233 (46.05)	273 (53.95)	1.80 (1.42–2.27)
	0–10	High	578 (46.17)	674 (53.83)	1.16 (0.90–1.49)
	>10	High	521 (64.80)	283 (35.20)	2.32 (1.76–3.07)
B_P9	0–10	Low	358 (39.51)	548 (60.49)	1.0 ref.	0.031 (−0.206–0.268) p=0.796
	>10	Low	535 (39.75)	811 (60.25)	2.22 (1.76–2.80)
	0–10	High	349 (58.46)	248 (41.54)	1.19 (0.97–1.45)
	>10	High	405 (56.80)	308 (43.20)	2.78 (2.18–3.55)
DPB1_P9	0–10	Low	293 (33.76)	575 (66.24)	1.0 ref.	0.244 (0.054–0.433) p=0.012
	>10	Low	222 (49.12)	230 (50.88)	1.93 (1.49–2.50)
	0–10	High	600 (43.35)	784 (56.65)	1.31 (1.08–1.59)
	>10	High	532 (62.00)	326 (38.00)	2.97 (2.38–3.71)
NHS
DRB1_P11	0–10	Low	32 (23.02)	107 (76.98)	1.0 ref.	0.534 (0.257–0.812) p=1.60E-04
	>10	Low	20 (22.99)	67 (77.01)	0.89 (0.46–1.73)
	0–10	High	90 (42.06)	124 (57.94)	1.52 (0.85–2.70)
	>10	High	87 (58.39)	62 (41.61)	3.31 (1.81–6.04)
DRB1_P13	0–10	Low	30 (22.39)	104 (77.61)	1.0 ref.	0.529 (0.254–0.805) p=1.64E-04
	>10	Low	19 (22.35)	66 (77.65)	0.89 (0.45–1.76)
	0–10	High	92 (42.01)	127 (57.99)	1.60 (0.89–2.86)
	>10	High	88 (58.28)	63 (41.72)	3.48 (1.88–6.44)
DRB1_P71	0–10	Low	53 (31.93)	113 (68.07)	1.0 ref.	0.449 (0.043–0.855) p=0.030
	>10	Low	38 (36.19)	67 (63.81)	1.39 (0.78–2.47)
	0–10	High	69 (36.90)	118 (63.10)	0.99 (0.55–1.77)
	>10	High	69 (52.67)	62 (47.33)	2.19 (1.17–4.11)
DRB1_P74	0–10	Low	26 (22.22)	91 (77.78)	1.0 ref.	−0.442 (−1.217–0.333) p=0.264
	>10	Low	32 (41.56)	45 (58.44)	2.54 (1.29–4.98)
	0–10	High	96 (40.68)	140 (59.32)	1.55 (0.87–2.74)
	>10	High	75 (47.17)	84 (52.83)	2.32 (1.23–4.38)
B_P9	0–10	Low	61 (36.53)	106 (63.47)	1.0 ref.	−0.084 (−0.776–0.609) p=0.812
	>10	Low	49 (47.12)	55 (52.88)	1.76 (1.02–3.03)
	0–10	High	61 (32.80)	125 (67.20)	0.92 (0.55–1.52)
	>10	High	58 (43.94)	74 (56.06)	1.55 (0.89–2.70)
DPB1_P9	0–10	Low	47 (28.48)	118 (71.52)	1.0 ref.	−0.580 (−1.429–0.269) p=0.180
	>10	Low	46 (47.42)	51 (52.58)	2.54 (1.43–4.53)
	0–10	High	75 (39.89)	113 (60.11)	1.51 (0.93–2.45)
	>10	High	61 (43.88)	78 (56.12)	1.85 (1.11–3.10)
Korean
DRB1_P11	0–10	Low	223 (41.14)	319 (58.86)	1.0 ref.	0.709 (0.357–1.060) p=7.79E-05
	>10	Low	40 (85.11)	7 (14.89)	8.76 (3.80–20.19)
	0–10	High	1003 (71.24)	405 (28.76)	2.87 (2.28–3.62)
	>10	High	124 (96.88)	4 (3.13)	44.12 (15.70–123.96)
DRB1_P13	0–10	Low	266 (43.68)	343 (56.32)	1.0 ref.	0.721 (0.383–1.060) p=2.89E-05
	>10	Low	37 (84.09)	7 (15.91)	6.88 (2.99–15.83)
	0–10	High	960 (71.59)	381 (28.41)	2.48 (1.98–3.11)
	>10	High	127 (96.95)	4 (3.05)	34.30 (12.40–94.92)
DRB1_P71	0–10	Low	403 (53.03)	357 (46.97)	1.0 ref.	0.588 (0.016–1.160) p=0.044
	>10	Low	65 (89.04)	8 (10.96)	7.72 (3.54–16.83)
	0–10	High	823 (69.16)	367 (30.84)	1.17 (0.91–1.50)
	>10	High	99 (97.06)	3 (2.94)	19.22 (5.75–64.20)
DRB1_P74	0–10	Low	493 (57.73)	361 (42.27)	1.0 ref.	−0.290 (−1.964–1.384) p=0.734
	>10	Low	63 (94.03)	4 (5.97)	12.55 (4.42–35.65)
	0–10	High	733 (66.88)	363 (33.12)	1.17 (0.93–1.47)
	>10	High	101 (93.52)	7 (6.48)	7.93 (3.44–18.27)
B_P9***	0–10	Low	563 (57.16)	422 (42.84)	1.0 ref.	0.628 (0.116–1.139) p=0.016
	>10	Low	75 (90.36)	8 (9.64)	7.75 (3.56–16.86)
	0–10	High	663 (68.70)	302 (31.30)	1.37 (1.12–1.67)
	>10	High	89 (96.74)	3 (3.26)	20.74 (6.39–67.32)
DPB1_P9	0–10	Low	602 (63.57)	345 (36.43)	1.0 ref.	NC
	>10	Low	82 (97.62)	2 (2.38)	30.55 (7.18–130.07)
	0–10	High	624 (62.21)	379 (37.79)	0.99 (0.81–1.22)
	>10	High	82 (90.11)	9 (9.89)	6.85 (3.28–14.33)

^*Amino acid GRS was calculated by weighing each amino acid on its reported RA-risk effect size. The calculated GRS were dichotomized at the median in controls of each population (High GRS group was defined by GRS ≥ median GRS).

^**Dichotomized two main factors, GRS and smoking pack-year, were used in the AP model adjusting for the top 5 principal components and GRS from other amino acid positions.

^***The high GRS group for HLA-B amino-acid position 9 in the Korean population was defined by GRS > median GRS because the median GRS is equivalent to the lowest GRS.

Significant P_AP values after the Bonferroni correction are in bold.

Correlation between HLA amino-acid residue/heavy smoking effects and RA-risk effects

We observed a positive correlation between the interaction effect and RA-risk effect at most residues at the positions 11 and 13 (Figure 1) in all three populations. For example, the most significant interaction effect with positive direction was found at Val11 and His13, which are the same residues associated with the strongest risk of RA in previous studies (2,3). Similarly, the most RA-protective residues (Ser11 and Ser13) showed the strongest AP with negative direction.

Figure 1.

Correlation between RA-risk effects and interaction effects of each residue at HLA-DRβ1 amino-acid positions 11 and 13. For each residue at the HLA-DRβ1 amino-acid positions 11 and 13, attributable proportion due to interaction (AP) with smoking for RA was plotted against to the reported odds ratio (OR) for RA susceptibility in Caucasians (Raychaudhuri S et al. Nature genetics 2012, 44(3):291) and Asians (Okada Y et al. Human Molecular Genetics 2014, 23(25):6916). Error bars display their 95% confidence intervals. Unstable AP with large 95% confidence intervals spanning from less than -1 to more than 1 were excluded from the plots. There are 6 correlation plots for the HLA-DRβ1 amino-acid position 11 in (a) EIRA, (b) NHS and (c) Korean, and HLA-DRβ1 amino acid position 13 in (d) EIRA, (e) NHS and (e) Korean.

DISCUSSION

In this study, we applied novel HLA models for RA risk in evaluating gene-environment interactions for RA risk. We found consistent and strong additive interactions between heavy smoking and the HLA-DRβ1 amino-acid haplotype model in Swedish, US, and Korean datasets, and further identified the key interacting polymorphisms by testing the interaction at the HLA amino-acid position level and amino-acid residue level.

In addition to the synergistically additive interaction between heavy smoking and HLA-DRβ1 amino-acid haplotype, we report two other novel observations in this study relevant to RA pathogenesis. First, our case-control study mapped the strongest interaction effect to the positions 11 and 13 in HLA-DRβ1 among the four amino acids comprising the haplotype interacting with smoking, whereas weaker or inconsistent effects were seen when interaction analysis was performed for HLA-DRβ1 positions 70 and 74, HLA-B amino-acid position 9 and HLA-DPβ1 amino-acid position 9 in our study populations. Second, we found a similar pattern between interaction effects of amino-acid residues at position 11 and 13 with heavy smoking and main effects on RA risk of those residues that are known to explain a majority of RA risk effect in the HLA-DRβ1 amino-acid haplotype (2,3).

We took advantage of a weighted GRS method (40,41) in the interaction analysis for the highly polymorphic HLA variants (e.g. 15 possible HLA-DRβ1 amino-acid haplotypes, 3 to 6 possible residues at each RA-risk amino-acid position). Because diverse amino-acid haplotypes or residues with different effect sizes could be aggregated into a single GRS value, all subsequent analyses became simplified, powerful and easily interpretable. Furthermore, in contrast to the traditional shared epitope model, our approach could accommodate effects of non-SE RA-risk alleles (e.g., HLA-DRB1*09:01) in the interaction tests.

The variety of the study designs included in our analysis adds to the strength of the consistency of the associations we described. Strengths of all three cohort studies include the enrollment of incident or recently diagnosed RA cases in NHS and EIRA, the collection of detailed smoking data regarding both smoking status and intensity over time, and the availability of genotyping data for HLA classical alleles and Immunochip SNPs. The EIRA is a population-based study recruiting individuals aged 18–70 years with both men and women included, and had the largest sample size. Since 85% of subjects were enrolled within a year of diagnosis, there is less potential for recall bias. The NHS dataset, a nested case-control dataset, was generated from two large prospective cohorts comprised of middle to older aged women with high education levels where exposure information was collected prior to onset of RA, thus without recall bias. However, due to the low incidence of RA onset in a prospective cohort design, the sample size was limited. The NHS lacks data on ACPA status in cases diagnosed before the widespread clinical usage of this test, however, RF status was available from the medical record reviews, and RF and ACPA are significantly correlated (4,42). Studies have also demonstrated similar relationships for RF-positive and ACPA-positive RA phenotypes in terms of gene-smoking interactions (4,17). Whereas the EIRA and NHS subjects are primarily of Caucasian heritage, the Korean population was composed of only Asian subjects recruited for a retrospective study. We note that Korean cases had a longer disease duration raising more potential for recall bias. The Korean female subjects, the major portion of the Korean subjects, were characterized by very low rate of smoking, which might result in instability of estimates for smoking in both association and interaction tests. Although we found relatively higher point estimates and their wider confidence intervals, Korean results showed similar patterns to EIRA and NHS in terms of statistical significance and direction of the effects of heavy smoking-RA association, haplotype GRS-smoking interaction, as well as amino-acid GRS-smoking interaction.

Our study was initiated with the aim of investigating whether the interaction between smoking and MHC class II and class I genetic variants was restricted to the variants that had previously been associated with risk for ACPA-positive RA. The results indicate indeed that the major interaction is with the variants defined by presence of certain amino acids in the 11 and 13 amino acid positions of HLA-DRβ1 molecule. However, the power of the investigation, despite being based on large cohorts, does not allow us to exclude weaker interactions with other MHC class II or class I structural variants. The primary interaction appears to be restricted to the amino acid positions 11 and 13, which are in high LD, and position 13 lies in the p4 pocket of the HLA-DR molecule, is important as it has recently been described that citrullinated peptides may differ from their arginine counterparts in their binding to the p4 pocket (43–45). It is thus tempting to suggest that presentation of epitopes of autoantigens that are citrullinated due to the influence of smoking may be specifically dependent on the presence of certain amino acid residues in position 13 of the HLA-DRβ1 chain. Such a scenario would be consistent with several previous studies that have reported that interactions between MHC class II variants and smoking is more pronounced for subpopulations of RA characterized by the presence of certain specific ACPA autoantibodies, notably to enolase and vimentin (46–50). All these previous studies have been performed using the shared epitope as a proxy for smoking-MHC class II interactions, but the present results suggest that functional studies should now be extended to use the amino-acid based model for detailed investigations of which particular MHC structures are involved in these interactions. Taken together, the opportunity to use the new amino-acid based model for detailed analysis of the contribution of variants of MHC to the emergence of potential pathogenic MHC class II restricted and environmentally triggered immune reactions, offers novel opportunities to define and eventually modify specific immune reactions that drive defined subsets of RA.