The influence of HLA genotype on the development of metal hypersensitivity following joint replacement.

Background: Over five million joint replacements are performed across the world each year. Cobalt chrome (CoCr) components are used in most of these procedures. Some patients develop delayed-type hypersensitivity (DTH) responses to CoCr implants, resulting in tissue damage and revision surgery. DTH is unpredictable and genetic links have yet to be definitively established.
Methods: At a single site, we carried out an initial investigation to identify HLA alleles associated with development of DTH following metal-on-metal hip arthroplasty. We then recruited patients from other centres to train and validate an algorithm incorporating patient age, gender, HLA genotype, and blood metal concentrations to predict the development of DTH. Accuracy of the modelling was assessed using performance metrics including time-dependent receiver operator curves.
Results: Using next-generation sequencing, here we determine the HLA genotypes of 606 patients. 176 of these patients had experienced failure of their prostheses; the remaining 430 remain asymptomatic at a mean follow up of twelve years. We demonstrate that the development of DTH is associated with patient age, gender, the magnitude of metal exposure, and the presence of certain HLA class II alleles. We show that the predictive algorithm developed from this investigation performs to an accuracy suitable for clinical use, with weighted mean survival probability errors of 1.8% and 3.1% for pre-operative and post-operative models respectively. Conclusions: The development of DTH following joint replacement appears to be determined by the interaction between implant wear and a patient's genotype. The algorithm described in this paper may improve implant selection and help direct patient surveillance following surgery. Further consideration should be given towards understanding patient-specific responses to different biomaterials.

Introduction

Hip joint replacement surgery (hip arthroplasty) has proven to be extremely successful in the treatment of end-stage hip arthritis. As a result, there are now approximately 2 million hip arthroplasties carried out in countries of the Organisation for Economic Co-operation and Development (OECD) alone[1].

Conventional total hip replacements (THRs) are composed of a metal femoral head which articulates against a polyethylene (plastic) cup or liner[2]. The lifespan of these so-called metal on polyethylene (MoP) prostheses may be limited in younger, more active patients. This is because during activities of daily living, the harder metal head wears away the plastic component. The release of greater amounts of wear debris over time increases the probability of a macrophage-driven, adverse immune response developing in the periprosthetic tissue[2]. The result of this is wear-induced osteolysis, in which the bony architecture surrounding the implant becomes compromised and the component/components loosen[3]. In this situation, revision surgery must be undertaken and a new device implanted.

Metal on metal (MoM) hip resurfacing prostheses were reintroduced at the turn of the century to address this problem[4]. In hip resurfacing surgery, the damaged articular surface is removed from the native femoral head and is replaced by a hollow CoCr femoral component, which articulates against a CoCr acetabular component. It was hoped that removal of the softer polyethylene from the bearing combination would lead to a reduction in wear debris and, therefore, an increase in implant longevity.

The initial early success of the Birmingham Hip Resurfacing (BHR) in young males[4], saw a rapid expansion in the market, with ever-widening patient eligibility criteria and several new prostheses released from competing manufacturers[5]. The technology was then adapted for use in THRs, so that patients without sufficient bone quality to accommodate a resurfacing might benefit from the perceived advantages of decreased wear and increased stability afforded by the large diameter metal bearings[6].

From 2005, there began to emerge an increasing number of case reports which described MoM hip patients returning to clinic with delayed onset groin pain[7]. At revision surgery, large, sterile fluid collections were encountered in the joint capsule[8]. Histopathological examination of excised periprosthetic tissues identified a macrophagic infiltration - as previously encountered in MoP prostheses - but frequently the macrophage response was accompanied by a perivascular T lymphocyte infiltrate. In the most severe cases, the perivascular cuffs had expanded in circumference to coalesce, leading to the formation of germinal centres and destruction of the synovial surface[9]. In a seminal paper, Willert et al. coined the term aseptic lymphocyte-dominated vasculitis association lesion (ALVAL) to describe these histopathological features[10]. ALVAL can frequently be associated with the destruction of local tissues including bone, muscle and neurovascular structures[11]. The lesions are progressive, and if revision surgery is delayed, the incidence of major post operative complications increases[12]. Reimplantation with CoCr components may lead to rapid recurrence of symptoms[13]. The overall clinical and histopathological picture is consistent with a delayed-type hypersensitivity (DTH) response to the metal debris shed from the prostheses[10]. Initially thought to be an idiopathic, rare phenomenon, failure of MoM THRs secondary to ALVAL has been reported to reach over 30% at six years[14].

Note: In this paper we use the terms DTH and ALVAL interchangeably, with ALVAL the preferred term when referring specifically to MoM hip patients in the current study.

Studies have shown that the risk of tissue damage is increased when prostheses shed greater volumes of metal debris[15]. Consequently, in 2012, the Medicines and Healthcare Products Regulatory Agency (MHRA, UK) issued an alert regarding the management of patients with MoM implants. In it they recommended the monitoring of metal concentrations in blood, establishing a threshold of 7 micrograms per litre (µg/l) of Co or Cr as an indicator of an adverse tissue reaction. Although these guidelines were based on a small study involving only 26 patients[16], the guidance has not been substantially modified since its first release. (http://www.mhra.gov.uk/home/groups/dts bs/documents/medicaldevicealert/con155767.pdf) However, patients display varying tolerances to metallic debris[17], with female patients apparently at greater risk of developing hypersensitivity[18]. There is also evidence to indicate that debris release from the taper junction of THRs exhibits greater immunogenicity than bearing surface debris[19]. This is reflected in the greater failure rates of MoM THRs which has led to their withdrawal from clinical use.

There are two phases of DTH: sensitisation and elicitation. During the sensitisation phase, antigen-presenting cells (APCs) take up, process and display an antigen. APCs migrate to regional lymph nodes where the displayed antigen may activate T4 cells and the production of memory T cells, which migrate to the original site. In the elicitation phase, a subsequent exposure to the antigen leads to its re-presentation to memory T cells with the release of T cell chemokines and cytokines such as interferon-gamma, which enhance the inflammatory response.

A critical factor in the development of DTH, therefore, is the presentation of a specific peptide/antigen at the peptide binding groove of an APC; a competitive process.

Metals are capable of provoking a variety of T cell-mediated, HLA-linked diseases, such as chronic beryllium disease[20], Co hard metal lung disease[21], and contact hypersensitivities[22]. Three pathogenic mechanisms have been described: self peptides held in the binding groove of an MHC molecule form complexes with metal ions, with the resulting complexes acting as antigens[23]; T cells recognize metal-induced changes to the MHC molecule itself[24]; metals directly affect the processing of self-peptides, resulting in T cells reactive to cryptic self-peptides[25].

Which mechanism may be the most important in the initiation of the ALVAL response? Previous research and clinical experience indicated that the first mechanism was the most likely, and that the N terminal sequence (NTS) of albumin was the prime candidate peptide sequence to investigate[26]. We, therefore, hypothesized that individuals developing ALVAL may have greater frequencies of (HLA gene encoded) peptide binding grooves with greater affinities for the NTS of albumin.

In this investigation, we demonstrate that variation in HLA class II genotype influences an individual’s susceptibility to DTH following implantation with a CoCr hip prosthesis. We go on to describe the development and validation of a machine learning algorithm to investigate the possibility that a patient’s genotype and basic clinical parameters may be used to predict the development of DTH.

Methods

Patients and hospital centres

Following Health Research Authority ethical approval (IRAS reference 227785), the study commenced at a single centre (centre 1, United Kingdom) where a large number of MoM hip arthroplasties were performed between 2002 and 2010. These patient cohorts have been described in full in previous publications[27]. The patients have been kept under surveillance with annual clinical review and blood metal ion testing. As part of an ethically approved project (IRAS reference 14119), patients who undergo revision of their MoM hip prostheses have: undergone metal ion testing to assess Co and Cr concentrations in their blood, serum, and hip joint synovial fluid samples; their explanted prostheses analysed to determine their volumetric wear; and tissue samples excised at revision surgery assessed by a specialist histopathologist (SN) (Fig. 1). The total number of revision cases in the database at commencement of the current study was 420. All patients included in the study gave informed consent.

Fig. 1

Explant analysis and tissue responses.

MoM hip components are manufactured from standard, medical-grade CoCr alloy (ASTM-75 or ASTM-1537), which is composed of approximately 65% Co and 30% Cr by weight. Explanted prostheses can be reverse engineered using coordinate measuring machines (CMMs) to quantify the volumetric material loss through wear and corrosion (shown in red in the wear maps below). Serum or whole blood Co and Cr concentrations provide a reliable in-vivo surrogate measure of the rate of this material loss[80]. Panels a,b,c and d relate to the same patient, whose blood Co concentration was elevated at 20.1 µg/l just prior to revision. a The explanted 46 mm diameter Birmingham Hip Resurfacing. b The corresponding CMM generated wear map (volumetric wear rate of 25mm3/year). c and d The synovial tissue sections (hematoxylin and eosin (H&E) stained), (X2 magnification and x15 magnification respectively) showing heavy macrophage infiltration, no lymphocytes. Panels e–h relate to a second patient, whose blood Co concentration was 1.5 µg/l just prior to removal of her ASR XL THR. e The explanted 45 mm diameter prosthesis. f The corresponding CMM generated wear map (volumetric wear rate of 1.5mm3/year). g and h The synovial tissue sections (H&E stained), (x2 magnification and x15 magnification respectively) showing a heavy perivascular lymphocyte infiltrate and extensive synovial necrosis (severe ALVAL).

Blood/serum metal ion testing

We have carried out a substantial amount of work detailing the relationships between volumetric wear of implants and the corresponding concentrations of Co and Cr ion in the blood, serum, and synovial fluid fractions[28]. Samples were tested using the generally accepted method of inductively coupled plasma mass spectrometry (ICP-MS) at accredited laboratories[19,29].

Wear analysis

Explanted prostheses were analysed using a coordinate measuring machine (Legex 322; Mitutoyo Ltd, Halifax, United Kingdom) to calculate the total amount of material that had been lost from the components in vivo: ‘total volumetric wear’, measured in mm.3 The total volumetric wear was divided by the number of years in vivo to calculate a mean ‘volumetric wear rate’ (expressed in mm3/year) which was the value used in the statistical analyses. The accuracy of the volumetric wear analysis performed on these types of explanted components has been validated and is of the order of 0.5 mm3 for a bearing surface and 0.2 mm3 for a female taper surface[30]. In this paper, wear rates refer only to CoCr material loss. For resurfacings, therefore, the wear rates refer only to the bearing surface wear rates (combined femoral head and acetabular component volumetric wear rates). For THAs, ‘total volumetric wear rates’ include the bearing wear as well as the wear from the female taper surface (Fig. 2). THAs in the study were used with titanium stems. We have previously demonstrated that titanium release is small in comparison with CoCr[31].

Fig. 2

The structure of total hip replacements (THRs).

a In total hip arthroplasty, the femoral neck is sectioned and a femoral stem (titanium, uncemented for the patients in this study) placed down the femoral canal. The femoral head is press fit on to the stem, creating the taper junction. b Metal debris can be generated from the taper junction, the great majority of which is released from the CoCr head. For the explanted THRs in this study, material loss from the bearing and tapers was quantified using a coordinate measuring machine. c A taper wear map is shown, with red areas indicating areas of material loss greater than 50 microns in depth.

Histopathological tissue assessment

This was carried out as has previously been described in greater detail[32]. Samples were taken from between two and four periprosthetic sites. Up to ten paraffin blocks were processed per site. Samples were also sent for microbiological testing to exclude sepsis. A single consultant histopathologist (SN) examined the slides independently of the clinical findings, blinded to the results of the wear or metal ion analyses. Note: Adverse reaction to metal debris (ARMD) is an umbrella term which refers to clinical signs and symptoms association with metal debris exposure[33]. The typical immunological response to metal debris is limited to a macrophage infiltrate[34]. ALVAL is a subset of ARMD, referring to the additional lymphocyte infiltrate and histological features of DTH. The hallmark of ALVAL/DTH is the development of a perivascular lymphocytic cuffs which increase in thickness as the recruitment of lymphocytes is further stimulated. In more severe cases, these cuffs can expand to develop into aggregates or coalesce into one another, forming larger aggregates. These higher-grade ALVAL responses are associated with the development of tertiary lymphoid organs in the local tissue. As part of routine clinical practise, the ALVAL response in the tissue samples in this study was graded from 0 (absent) to 3 (severe) according to the integrity of the synovial membrane and the extent of lymphocytic infiltration (Fig. 1), a classification system which has shown good intra and interobserver reliability[32].

Investigation of genetic associations using extreme phenotype group comparison

From the hospital database, we identified four groups of patients, to represent the different phenotypes: patients with joint failure who developed moderate/severe ALVAL in association with prostheses wearing at lower than the median wear rate of the total revision cohort; patients with joint failure who developed moderate/severe ALVAL in association with prostheses wearing at greater than the median wear rate of the total revision cohort; patients with joint failure with a pathological response limited to macrophage infiltration, no lymphocyte infiltration identified; patients with joints remaining in situ who were pain-free and satisfied with the results of their hip arthroplasties at a minimum of ten years post surgery. We wrote to these patients explaining the nature of the study and invited them to submit a sample for DNA analysis.

DNA sample collection and processing

A combination of ORAcollect OCR-100 buccal swabs and Oragene DNA OG-610 saliva collection kits (both DNA Genotek Inc, Ontario, Canada) were used to collect samples for DNA extraction. DNA was extracted using a Roche MagnaPure Compact automated platform (Roche Holding AG, Switzerland). DNA was then quantified using a Thermo Fisher Qubit dsDNA BR Assay kit (Thermo Fisher, Massachusetts, United States) with standardisation to 25 ng/μl. HLA genotyping was then performed using One Lambda AllType NGS kits (One Lambda, USA), with the Illumina MiSeq platform (Illumina, USA). Full gene sequencing was carried out for HLA-A, -B, -C, -DQA1 and -DPA1, and partial gene sequencing for HLA-DRB1, -DRB345, -DQB1 and -DPB1 (with omission of exon 1). HLA genotypes were analysed using One Lambda TypeStream Visual 1.3 software (One Lambda, USA).

Global locus-wise association for each HLA gene was performed using UNPHASED v 3.0.13. Haplotypes were estimated for DRB1-DQA1-DQB1 also in UNPHASED[35], and then the distribution of the HLA class I and II alleles were compared between groups using a standard approach[36]. The genotypes for each HLA gene were transformed into dosages of each individual allele from the patient population, where 2 denoted two copies of an allele, 1 denoted one, and 0 denoted zero copies. These values were then entered as predictor variables in a logistic regression analysis. Multiple models were tested, comparing the extreme phenotype groups described above, and these were also compared to a background population from the United Kingdom. All models were also tested with sex as an additional covariate and also age plus sex as covariates.

In silico analysis of peptide-HLA class II binding affinity and Cox proportional hazards modelling

Peptide binding analysis

We used validated software to model the peptide-binding grooves encoded by an individual’s HLA genotype and to determine the resulting binding affinity between these binding grooves and an array of naturally occurring peptides[37]. Using this approach, we sought to: identify HLA genes associated with the development of ALVAL; determine whether HLA genes are associated with the development of ALVAL at low rates of wear encode for peptide binding grooves with higher affinities for the N terminal metal-binding sites of albumin.

All HLA-DQA1, -DQB1, and DRB1 alleles were selected to assess the peptide binding affinity of their corresponding peptide-binding proteins. HLA-DR is represented by HLA-DRA/DRB1 dimer. Since HLA-DRA is considered monomorphic, we only used HLA-DRB1. HLA-DQ is represented by the HLA-DQA1/DQB1 dimer. A basic schematic of the HLA-DQ structure and how it relates to peptide binding is shown in Fig. 3.

Fig. 3

MHC structures and peptide presentation.

The HLA-DQ molecule is an αβ heterodimer of MHC class type II. The three-dimensional shape of the peptide-binding groove is formed by the combination of α and β chains, which are genetically encoded by the HLA-DQA1 and HLA-DQB1 alleles respectively[81]. The structure of the peptide-binding groove determines which peptides (foreign or self) are presented at the cell’s surface[81], as is shown in the schematic. As an example, in coeliac disease, patients possess HLA alleles that encode for peptide binding grooves with structures suited to the presentation of gluten-derived peptides[82].

FASTA-formatted protein sequence data were retrieved from the UniProt database (www.uniprot.org) for human serum albumin (P02768). We extracted the first 15 amino acids of the N terminal (DAHKSEVAHRFKDLG), a sequence which includes two recognised Co binding sites. Predictions for HLA binding to this sequence were performed using NetMHCIIpan4.0[37]. The rank binding affinities were calculated for all the possible DQ and DRB1 combinations. We used the %EL rank score as the primary binding metric, as advised by the software developers[38]. We investigated whether the binding scores influenced the risk of developing ALVAL over time using Cox proportional hazards modelling. Multiple survival models were constructed to explain the development of time dependent prosthetic failure associated with mild/moderate or severe ALVAL, using the following independent variables: NTS binding affinity; pre-revision blood Co concentrations; pre-revision blood Cr concentrations; patient sex; patient age at the time of primary surgery; the presence of bilateral prostheses; type of prosthesis (THR versus resurfacing arthroplasty).

Expansion of data set, the inclusion of patients from other centres and development of machine learning algorithm

We then invited all remaining patients in the database who had undergone revision surgery for whom there was a full complement of clinical data, including explanted components available for analysis. We also invited all remaining patients under regular follow up who were recorded to be asymptomatic at greater than ten years follow up. Concurrently, we expanded the study to include two other units. Centre 2 is a major specialist orthopaedic unit in New York, United States. Centre 3 is a teaching hospital and tertiary referral centre in Western Australia. The units manage the follow up of MoM patients in a similar way and also routinely carry out analysis of explanted components. A similar research protocol was followed, with patients who were asymptomatic as well as those who had experienced failure of their joints invited to give a sample for DNA analysis. Relevant national and local ethical approvals were sought and granted (Protocol 2020-208, IRB approval for the United States; study RGS0000003851 Human Research Ethics Committee approval for Australia). The same parameters were recorded as at centre 1, with all patients giving informed consent. When all samples had been analysed, the data set was randomly split 70/30, with the larger set used to train a machine learning algorithm for the prediction of the development of ALVAL. The remaining data was held back, blinded from the analysts and used to test the algorithm when it was finalised.

Two models were trained to predict hazard ratios and survival functions up to ten years after implantation of a MoM prosthesis for pre-operative and post-operative patients. The first was a model to preoperatively predict the development of ALVAL. For this model, metal exposure was divided into two groups: low wear (Co concentrations stabilise to <2 µg/l) and increased wear (Co concentrations stabilise to ≥2 and ≤4 µg/l). 4 µg/l equates to approximately three times the wear rate of a well-functioning device. It was therefore not felt necessary to provide a preoperative prediction for metal concentrations above this level. A second model was developed to predict the development of ALVAL in the post-operative period, in which actual measured Co and Cr concentrations could be used in the modelling.

Statistics and machine learning approach

As the training and test set are assumed to be drawn from the same probability distribution, they should be identically distributed[39]. We, therefore, formulated our test set by randomly sampling the full dataset (without replacement) stratified on the event indicator. The training data was composed of the remaining samples.

Feature engineering was carried out on the training data to identify features that best predicted risk of failure due to ARMD and ALVAL within ten years of implantation of a MoM prosthesis. Boruta[40], a random forest feature selection algorithm was applied to 2939 features, generated from a combination of: patient features; binding affinities of cis and trans haplotypes; binary presence of cis and trans haplotypes; cis and trans haplotype gene dosage; thresholding binding affinities of cis and trans haplotypes to generate categorical features; polynomial and interaction features. The algorithm removed features that were identified as being less relevant than random features in an iterative supervised fashion to avoid overfitting. Features that were identified as being associated with ALVAL were used to train gradient boosted survival analysis machine learning models with a Cox proportional hazards loss function and a regression tree base learner[41,42]. Regularisation was employed to reduce overfitting on the training data. Nested 5-fold cross-validation (CV) was used on the training data to enable better estimation of generalisation error and reduce model selection bias.[43,44]. Hyper-parameters of both models were optimised using a successive halving random search[45,46]. Integrated Brier Score (IBS) was chosen as the scoring function. Cross-validated probability calibration did not yield improvements in IBS and Integrated Calibration Index (ICI).

After selecting the best-performing model based on the IBS assessed on the training data, the model was then used to predict on the test set. IBS, Uno’s c-index[47], time-dependent AUROC (ROC(t))[48], and ICI performance statistics were computed[49]. We used Austin et al.’s adaptation of ICI for survival analysis problems[50]. Confidence intervals were estimated using the Bootstrap method. After completion of performance evaluation, each model was refit on the training and test data and hyper-parameters were returned. The models were then serialised and integrated into a cloud-hosted pipeline for inference via a web app.

Table 1

Clinical details of all the patients from all centres in the study, divided by clinical outcome.

	Total	Asymptomatic	Failed
Total number of patients	606	430	176
Total number of hips	711	535	176
Follow up (years)	10 (1–20)	12 (3–20)	6 (1–15)
Age (range)	55 (25–85)	54 (25–78)	58 (25–85)
% male patients	66% (397:209)	74% (320:110)	44% (77:99)
Resurfacings vs THRs	468 vs 138 (77%)	43 vs 430 (90%)	46% (81:75)
% patients with bilateral prostheses	24%	24%	24%
BMI	26.6	26.7	26.3
Median (range) Co (µg/l)	2.00 (0.1–271.0)	1.50 (0.1–34.4)	7.60 (0.7–271.0)
Median (range) Cr (µg/l)	2.50 (0.2–108.4)	2.00 (0.2–18.6)	7.01 (0.7–108.4)

Table 2

Cox proportional hazards modelling, all 606 patients involved in the study included.

Variable	Coeff	Standard error	P-value	Hazard ratio (HR)	HR Lower CI (95%)	HR Upper CI (95%)
Model 1: Survival based on ALVAL severity of mild, moderate, or severe
Rank binding affinity for NTS	−1.463	0.404	<0.001	0.232	0.105	0.511
Log normalised cobalt concentration	1.649	0.136	<0.001	5.202	3.982	6.797
Age	0.005	0.011	0.667	1.005	0.984	1.026
Sex-M	−0.571	0.190	0.003	0.565	0.389	0.820
Type-THR	0.779	0.201	<0.001	2.180	1.471	3.230
Model 2: Survival based on ALVAL severity of moderate or severe
Rank binding affinity for NTS	−2.532	0.544	<0.001	0.079	0.027	0.231
Log normalised cobalt concentration	1.656	0.178	<0.001	5.236	3.696	7.417
Age	0.013	0.014	0.359	1.013	0.986	1.041
Sex-M	−0.631	0.247	0.011	0.532	0.328	0.863
Type-THR	0.728	0.264	0.006	2.070	1.235	3.470

Patients were censored based on a minimum ALVAL grade of “mild” and above, then a second model was constructed with patients censored only if they had “moderate” ALVAL and above.