Objective To find autoantibodies (AAbs) in serum that could be useful to predict incidence of radiographic knee osteoarthritis (KOA).
Design A Nucleic-acid Programmable Protein Arrays (NAPPA) platform was used to screen AAbs against 2125 human proteins in sera at baseline from participants free of radiographic KOA belonging to the incidence and non-exposed subcohorts of the Osteoarthritis Initiative (OAI) who developed or not, radiographic KOA during a follow-up period of 96 months. NAPPA-ELISA were performed to analyse reactivity against methionine adenosyltransferase two beta (MAT2β) and verify the results in 327 participants from the same subcohorts. The association of MAT2β-AAb levels with KOA incidence was assessed by combining several robust biostatistics analysis (logistic regression, Receiver Operating Characteristic and Kaplan-Meier curves). The proposed prognostic model was replicated in samples from the progression subcohort of the OAI.
Results In the screening phase, six AAbs were found significantly different at baseline in samples from incident compared with non-incident participants. In the verification phase, high levels of MAT2β-AAb were significantly associated with the future incidence of KOA and with an earlier development of the disease. The incorporation of this AAb in a clinical model for the prognosis of incident radiographic KOA significantly improved the identification/classification of patients who will develop the disorder. The usefulness of the model to predict radiographic KOA was confirmed on a different OAI subcohort.
Conclusions The measurement of AAbs against MAT2β in serum might be highly useful to improve the prediction of OA development, and also to estimate the time to incidence.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
What is already known about this subject?
Autoantibodies (AAbs) are used as biomarkers in autoimmune diseases such as rheumatoid arthritis or systemic lupus erythematosus. In these and other plethora of disorders, they can be detected at asymptomatic stages.
Although the presence of AAbs has been reported in the serum of patients with osteoarthritis (OA), they had not been previously associated with the incidence or progression of this disease.
What does this study add?
A specific panel of AAbs has been detected at baseline in individuals developing incident radiographic knee OA (KOA) during a 96-month follow-up period, compared with those who remained healthy.
Reactivity levels of AAbs against the beta subunit of the methionine adenosyltransferase (MAT2β-AAb) II enzyme are positively correlated with the time to OA incidence.
How might this impact on clinical practice or future developments?
The addition of MAT2β-AAb to a prognostic clinical model of incident radiographic KOA might significantly improve the identification at baseline of those individuals who will develop the disorder during a follow-up period of 96 months.
Osteoarthritis (OA) is the most common arthritic disease involving movable joints and it is increasingly important in current ageing populations, leading to patient chronic disability.1 2 The current diagnostic methods are insensitive to detect the small changes occurring at early stages, when OA is characterised as an asymptomatic disease.1 To solve this problem, a molecular level of interrogation is hypothesised as the only alternative to detect the earliest phases of the disease process.2
Although OA is not considered an autoimmune disease, cell stress and extracellular matrix degradation may activate maladaptive repair responses, including pro-inflammatory pathways of innate immunity.3 Activation of the immune response usually involves the production of immunoglobulins against self-proteins or autoantibodies (AAbs), which can be detected in sera and used as biomarkers for early diagnosis.4 5 In this field, the Nucleic-Acid Programmable Protein Array (NAPPA) strategy has been widely used to detect AAbs in a high-throughput manner in many diseases,6 7 and has been employed in an exploratory study on sera from patients with OA.8 The NAPPA arrays are generated by printing full-length cDNAs encoding the target proteins with a tag on the surface of the array.9 Proteins are then transcribed and translated by a mammalian cell-free system and captured in situ by immobilised antibodies specific for the tag encoded at the carboxy-terminus of the amino acid sequence.10
The Osteoarthritis Initiative (OAI) is an ideal target population to detect relevant biomarker characteristics of earlier stages of the disease. It is a multi-centre, longitudinal and observational cohort study that has enrolled 4796 individuals which have been followed during 96 months.11 12 Among all these subjects, the OAI comprises participants without clinically significant knee osteoarthritis (KOA) at baseline, but selected on the basis of having specific characteristics that give them an increased risk of developing incident symptomatic KOA (incidence subcohort), and a reference control group whose participants did not have neither symptomatic KOA nor risk factors at baseline (non-exposed subcohort).
In the present study, serum samples at baseline from the incidence and non-exposed subcohorts of the OAI were analysed using NAPPA technology for the discovery of an AAb profile that could be associated with an early and asymptomatic stage of the disease. The objective was to detect AAbs useful to identify those asymptomatic individuals who will develop radiographic KOA before 96 months, and examine the putative relationship between their levels in serum and the time for OA incidence.
Materials and methods
Definition of incident radiographic KOA
In this case–control study, two main outcomes group, with one study knee per subject (target knee), were defined: the incident (n=146) and the non-incident group (n=181), both without clinically relevant radiographic KOA at baseline (Kellgren and Lawrence (KL) grade=0–1) in at least one knee (target knee). Incident radiographic KOA was defined by KL grade ≥2 in the target knee at some point between 12 and 96 months of follow-up.
A two-stage discovery approach (screening and verification) was designed to analyse the presence and putative usefulness of KOA-associated AAbs to predict the incidence of the disorder. The sera were blindly analysed and belonged to Caucasian participants from the incidence and non-exposed subcohorts of the OAI at the baseline visit (181 non-incidents and 146 incidents at 96 months). The prognostic clinical model generated to predict KOA development was later replicated in a total of 108 participants (65 non-incidents and 43 incidents at 96 months) from the progression subcohort of the OAI at the baseline visit. Detailed information about the workflow of the study is summarised and illustrated in figure 1.
NAPPA profiling of serum AAbs
The NAPPA core Centre for Personalised Diagnostics (CPD) at the Biodesign Institute (Arizona State University, USA) had all human genes from the DNASU (www.dnasu.org) on six different array sets: HC1−HC6. The HC5 set was selected for the screening on the basis of having the greatest number of genes that could be related with OA pathogenesis according to bibliography (listed in online supplementary table S1). The quality of DNA printing, protein expression and detection of the NAPPA slides were performed using the standard procedure of CPD,13 but incubating the expressed slides overnight with 150 µL of 1:20 (v/v) diluted serum. A titration assay was performed using serum dilutions from 1:20 down to 1:200 to identify an optimal dilution factor that provided an acceptable background without overwhelming the true signals.
The signal intensities obtained in the assay were normalised as described.14 To determine positive AAb response, a cut-off level was calculated by median intensity absolute deviation rule from all the spots through all the serum pools. The mean and SD of the mean were obtained for the incident and non-incident groups. The antigens that did not exhibit intensities over the cut-off were eliminated. A differential spot analysis was performed with the remaining antigens by Wilcoxon Rank-Sum test (p value<0.05) and the area under the curve (AUC) at 95% specificity were calculated using the pROC package in R. In addition, AAb candidates were qualitatively examined by visual analysis through all the slides by adjusting in identical black and full colour threshold scale.
In all, 500 ng of full-length methionine adenosyltransferase two beta (MAT2β) human recombinant protein fused to glutathione S-transferase (GST) were synthesised in vitro using the HeLa cell lysate-based protein in vitro transcription/translation (IVTT) system (Thermo Fisher Scientific). NAPPA-ELISA assay was performed as previously described8 with some variations. 96-well plates coated with 40 ng of anti-GST antibody were blocked 4 hours at room temperature with 5% milk-1×phosphate-buffered saline with tween (PBST) (0.02% Tween20). 50 µL of the IVTT-expressed recombinant human protein was transferred to each well and incubated at 4°C overnight on shaker. Plates were then washed and incubated with 1:20 (v/v) diluted sera. The presence of specific MAT2β-AAb was detected by incubation with horseradish peroxidase (HRP)-linked anti-Human IgG (Jackson ImmunoResearch Laboratories) diluted 1:1000 (v/v) in blocking buffer. After addition of tetramethyl-benzidine substrate, the absorbance signals at 450 nm were read on a Biotek Synergy four plate reader (Winooski, VT, USA). Levels of MAT2β-AAb were expressed in arbitrary units (a.u.) of absorbance.
The biological context network of MAT2β was analysed with the STRING (https://string-db.org/) bioinformatics webtool, using the K-means clustering method. Differences in the baseline reactivity levels of MAT2β-AAb were assessed by the Mann-Whitney U test and the association of this potential biomarker with OA incidence was evaluated with the OR. In addition, after assessment of cut-off values (tertiles) for MAT2β-AAb, patients with OA were categorised into high-level, medium-level and low-levels groups. Kaplan-Meier (KM) analyses were used to estimate and represent the survival probability, explained as the probability of not developing KOA in specific periods of time (12, 24, 36, 48, 72 and 96 months) depending on the tertiles of MAT2β-AAb reactivity levels of the participants.
To define prognostic models of OA, clinical data at baseline were obtained from the OAI database (https://data-archive.nimh.nih.gov/oai). Candidate non-radiographic clinical variables that may have prognostic value were selected based on the specific eligibility risk factor criteria for the incident subcohort of the OAI and prior published evidence suggesting a risk factor role in KOA incidence (online supplementary table S2). For all variables concerning the joint, knee-value predictors were recoded to indicate they were for the target knee. When neither or both knees have incident KOA, one of them was randomly selected and used in the analysis. In a primary step, univariable logistic regression analyses were employed to assess association between each variable with incident radiographic KOA. In a secondary step, a stepwise multivariable logistic regression analysis was performed to define a prognostic model of incident radiographic KOA.
The capacity of the models to predict OA incidence was evaluated using the AUC. The utility of the measurement of the reactivity levels of the potential biomarker was assessed by comparing the AUC of the covariates-only model with the AUC of the biomarker plus covariates model. Sensitivity, specificity and positive predictive value and negative predictive value were also estimated by the Youden Index to determine the validity and security of the models, and receiver operating characteristic (ROC) curves were evaluated. A nomogram was developed to facilitate application of the proposed prognostic model in a clinical setting. Finally, the validity of the proposed biomarker plus covariates model was evaluated by a replication analysis in a different set of participants from the OAI progression subcohort.
All regression analyses and KM curves were carried out using SPSS V.25 for Mac. Metrics were calculated using the pROC package in R.
Patient and public involvement
This research was done without patient involvement. Patients were not invited to comment on the study design and were not consulted to develop patient relevant outcomes or interpret the results. Patients were not invited to contribute to the writing or editing of this document for readability or accuracy.
Identification of AAbs associated with the incidence of KOA
To search for AAbs in the serum that could be associated with a future development of KOA, a comprehensive AAb profiling against 2125 full-length proteins was performed by NAPPA. It was carried out comparing pools of serum samples at baseline from two groups: the incident group, which contains participants belonging to the incidence subcohort of the OAI who did develop radiographic KOA during the 96-month follow-up (n=100, 10 pools), and the non-incident group, which contains participants from the non-exposed subcohort that remained radiographically healthy (n=100, 10 pools).
A signal cut-off >1.1 was employed to assure a sufficient margin between positive and negative AAbs reactivities. Among the 2125 screened proteins, a total of 1031 proteins showed positive immunoreactivity (online supplementary table S3). Mean and SD values for all the proteins expressed on the array for the incident and non-incident groups are summarised in online supplementary table S3. From all the proteins over the cut-off, the Wilcoxon Rank-Sum test identified a total of six AAbs that reacted with six different proteins on the array (online supplementary table S4). Furthermore, as shown in online supplementary figure S1, visually discernible differences demonstrated that the normalisation criteria employed did neither create signal differences that do not exist, nor destroy true signal differences.
Verification of MAT2β-AAb levels as potential prognosis marker of OA incidence
To confirm the putative ability of any of these AAbs to predict the incidence of OA, anti-MAT2β was selected to be technically verified by NAPPA-ELISA in 327 individual serum samples, 200 of which (100 incidents and 100 non-incidents) were previously used at the screening phase. This selection was based on the role of MAT2β as negative regulatory subunit of the production of S-adenosylmethionine (SAMe), a main methyl donor in the human body and a widely used dietary supplement for OA management (AdoMet). In addition, its biological context network (online supplementary figure S2) suggests a bottleneck role of this protein in metabolic pathways that are known to be related with OA pathogenesis. After the subtraction of the negative control, we found positive reactivity against MAT2β in all the analysed sera. The higher baseline reactivity levels of MAT2β-AAb found in the incident group (0.58±0.22 vs 0.49±0.23 a.u., p=3.140E-04) verified our previous findings. The association of this potential biomarker with the clinical outcome and its ability to predict incidence of radiographic KOA were assessed and are summarised in table 1.
Furthermore, we detected a significant decrease in anti-MAT2β reactivity levels with the time to KOA onset (p value=0.002), as it is shown in online supplementary figure S3. The association between the baseline reactivity levels of MAT2β-AAb in sera and the time to OA incidence were inferred by KM curves (figure 2). Individuals with low AAb levels at baseline (range=2.00E-3–0.39 a.u.) had a significant lower risk to develop KOA sooner in time than those with high (range=0.60–1.58 a.u., p=2.440E-04) or medium (range=0.39–0.60 a.u., p=5.000E-06) baseline levels. There were no significant differences in the time to KOA incidence when the high-level and medium-levels groups were compared (p=0.267).
Predictive modelling of KOA incidence with the combination of clinical variables and MAT2β-AAb levels
A unique covariates-only model including age, sex, body mass index (BMI), frequent knee bending activity, history of knee injury and the Western Ontario and McMaster Universities Osteoarthritis Index pain score was defined by stepwise multivariable logistic regression analysis. The clinical characteristics finally included in the model of the selected participants are presented in table 2. The results from the univariable logistic regression model of all the clinical variables analysed are shown in online supplementary table S5.
The clinical model defined herein yielded an AUC (95% CI) of 0.81 (0.76–0.86). The addition of MAT2β-AAb to this model significantly improved the capacity to predict radiographic KOA development in the target knee (p=0.048), yielding an AUC of 0.83 (0.78–0.87). Figure 3A,B shows the results from this regression analysis, together with the metrics and the ROC curves obtained when comparing the covariates-only model with the MAT2β-AAb plus covariates model.
To facilitate the use of the proposed prognostic model in a clinical routine, a nomogram was developed (figure 3C) to determine the probability of a certain individual to develop KOA in the next 96 months.
Replication analysis in an independent OAI subcohort
The proposed MAT2β-AAb plus covariates model was replicated in an independent set of sera at baseline from participants without KOA belonging to the progression subcohort of the OAI. The clinical characteristics of this population are summarised in table 2. The AUC observed in this replication analysis was 0.76 (table 3), showing no significant differences with that obtained in the verification phase (p=0.218). In addition, the sensitivity, specificity and predictive values remained very similar in both verification and replication cohorts.
One of the features of a prognostic marker is the ability to predict the future occurrence of a certain disease among people who do not have it.15 The production of antibodies against self-proteins is a characteristic feature of many diseases.16 Considering the fact that AAbs can often be detected at asymptomatic stages,4 they might have the potential to identify susceptible individuals or populations and facilitate prognosis. The idea that AAbs can be used to predict a disease state has been extensively studied in different disorders, such as cancer17–19 or type 1 diabetes.20 21 In the field of rheumatic diseases, AAbs have a fundamental value in the diagnosis of those with an autoimmune pathogenesis, such as systemic lupus erythematosus22 and rheumatoid arthritis.23
Although OA is not considered an autoimmune disorder, the immune system is highly related with early disease.24 However, existing literature related to the presence of AAbs in patients with OA is limited.8 25–27 Indeed, this is the first study that evaluates the usefulness of AAbs to stratify patients with asymptomatic OA. The use of a large-scale approach (NAPPA) enabled the search of a massive number of putative AAbs, which is the largest screening performed to date in the OA field. With this approach, we have identified significant levels of reactivity from AAbs against six different proteins that are associated with the future incidence of the disease. At this point, a characteristic of this approach should be taken into consideration when interpreting the findings presented herein: The low sera dilutions employed in this work lead to the primary detection of class M immunoglobulins (IgM), which, in contrast to IgGs, have no immune memory. However, IgMs are not subjected to immunoregulation28 and are formed early in the immune response. Therefore, specific antibodies of the IgM class might be important in the diagnosis of chronic diseases.29
The results for MAT2β-AAb have been verified on 327 individual samples at baseline from the OAI cohort, which provides a robust evaluation of its ability to classify patients at baseline as incident or non-incident during a 96-month period. Moreover, KM curves showed a significant association of the baseline reactivity levels of MAT2β-AAb with the time of KOA appearance. This statistical approach has been widely used in cancer biomarkers30–32 and it has been recently introduced in the rheumatology field.33 34 Interestingly, our results showed that higher baseline reactivity levels of this AAb result in a sooner development of radiographic KOA.
MAT2β is the regulatory subunit responsible of enhancing or inhibiting the synthesis of SAMe. This latter compound plays a vital role in methylation, transsulfuration and aminopropylation pathways,35 and it has been employed as dietary supplement for OA management.35–38 Although there is no evidence of the direct involvement of MAT2β in OA, its fundamental role in key biological processes for the pathogenesis of this disease (online supplementary figure S2) turns it into a potential marker of interest. Curiously, this protein was not included in either the planar arrays or the NAPPA developed previously by our group for the screening of OA-associated AAb.8
The MAT2β-AAb-only model showed a modest ability to predict radiographic KOA development, yielding an AUC of 0.62 with 39% specificity and 86% sensitivity. In the OA field, this modest predictive capacity is in agreement with that obtained for different biomarkers that have been evaluated in the last years to predict relevant OA progression. For example, Eckstein and collaborators examined the relationship of 15 molecular markers with structural progression based on femorotibial cartilage loss.39 The strongest predictors of longitudinal thinning were serum C-terminal telopeptide of collagen type I (CTX-I) and plasma N-terminal propeptide of type II procollagen, yielding AUCs of 0.65 and 0.64, respectively, but all the remaining biomarkers showed AUCs<0.60. Furthermore, in the meta-analysis published by Valdes and collaborators,40 urine CTX-II also showed a limited predictive capacity (AUC ≤0.63). Recently, Kraus and collaborators have investigated a target set of 18 biochemical markers (baseline and time-integrated concentrations (TICs) over 12 and 24 months) as predictors of symptomatic and radiographic KOA progression.41 Among all of them, the best single biomarker was the 24 M TIC CTX-II measured in urine, yielding an AUC=0.58.
With the aim of defining a useful non-radiographic prognostic model focused on KOA, different non-radiographic clinical factors related with risk of incident KOA in the literature42–44 have been analysed in this study by univariable logistic regression analysis to look for significant predictors. It is important to specify that this study is based on a Caucasian US population, which may not comprise all the factors that enhance predisposition to OA. Among the clinical variables finally included in the model, the history of knee injury showed the highest OR, which was markedly lower than the reactivity levels of MAT2β–AAb (OR 2.57 vs 5.99) for being incident. As shown herein, the use of stepwise multivariable regression analysis resulted in one prognostic model of KOA incidence with an AUC=0.81. Using data from individuals in the Rotterdam study, a prediction model using clinical factors that yields an AUC of 0.66 was defined.45 In another study, Zhang and collaborators defined a model of incidence of radiographic KOA with data from the Nottingham cohort, the OAI cohort and the Genetics of Osteoarthritis and Lifestyle (GOAL) study.46 This model, including variables such as age, gender, BMI, occupational risk, family history and knee injury yielded the greatest AUC (0.74) in the GOAL population, compared with the OAI (AUC=0.60) and the Nottingham (AUC=0.69). In our case, the remarkably high ability to predict the appearance of radiographic KOA using this covariates-only model could be due to the high prevalence of the disease in our verification cohort (45% of incident participants). Indeed, the application of this model into the whole OAI database yielded a lower AUC (AUC <0.70 (data not shown)).
Despite the rather low AUC of MAT2β-AAb, the addition of this potential biomarker to the prognostic covariates-only model led to an increase in its discriminative ability (AUC=0.83), being this increase statistically significant (p=0.048). A similar improvement, but in this case not significant, was previously reported by including uCTX-II levels in a clinical prediction model for KOA.45 At this point, it is important to take into account that among all the patients involved in this study, 42.5% of them had already developed radiographic OA in the off-target knee (KL ≥2). Nevertheless, we observed that the prognostic ability of MAT2β-AAb levels is maintained after removing subjects with contralateral OA at baseline (AUC=0.61 (0.54–0.68)). In addition, the inclusion of the presence of contralateral OA to the proposed MAT2β-AAb plus covariates model did not improve its predictive capacity (p=0.093). This strengthens the utility of this biomarker to predict incidence of KOA without the need of any radiographical information from the patients, avoiding their exposure to harmful radiation. Finally, the prognostic model combining baseline levels of MAT2β-AAb with clinical variables was replicated in an independent population, which confirms its putative utility to predict the appearance of radiographic KOA and strongly suggests it can be generalisable to the wider population.
In summary, the present study shows that a high antibody reactivity against MAT2β protein in serum is associated at baseline with individuals who will develop incident radiographic KOA, and also with an earlier appearance of the disease. Our results suggest that the inclusion of MAT2β-AAb in a clinical prognostic model for radiographic KOA could improve the identification of individuals who will develop the disease before 96 months.
The authors acknowledge the sample and data set provided from the OAI, a public–private partnership comprised of five contracts (N01-AR-2–2258; N01-AR-2–2259; N01-AR-2–2260; N01-AR-2–2261 and N01-AR-2–2262) funded by the NIH.
Handling editor Josef S Smolen
Contributors Conception and design: MC-E, JvD, JQ, JLB, CR-R and FJB. Acquisition, analysis and interpretation of data: MC-E, VB-B, FP, IR-P, JvD, JQ, MF, JLB, CR-R and FJB. Drafting the article: MC-E, VB-B, IR-P, CR-R and FJB. Final approval of the article: All authors.
Funding The Proteomics Unit belongs to ProteoRed, PRB3- ISCIII, supported by grant PT17/0019/0014. This work has been funded by grants from Fondo Investigación Sanitaria-Spain (PI14/01707, PI16/02124, PI17/00404, DTS17/00200, CIBER-BBN CB06/01/0040, CIBER-ONC CB16/12/00400, RETIC-RIER-RD12/0009/0018), a part of the National Plan for Scientific Program Development and Technological Innovation 2013-2016, funded by the ISCIII-General Subdirection of Assessment and Promotion of Research - European Regional Development Fund (FEDER) "A way of making Europe" . MC-E is supported by the Xunta de Galicia and the European Union (European Social Fund – ESF) through a predoctoral fellowship (IN606A-2016/012). IR-P and CR-R were supported by the Miguel Servet II programme from Fondo Investigación Sanitaria-Spain (CPII17/0026 and CPII15/00013, respectively).
Competing interests None declared.
Patient consent for publication Not required.
Provenance and peer review Not commissioned; externally peer reviewed.
Data availability statement All data relevant to the study are included in the article or uploaded as supplementary information.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.