Autoantibodies identify primary Sjögren’s syndrome in patients lacking serum IgG specific for Ro/SS-A and La/SS-B

Abstract

Objective

Identify autoantibodies in anti-Ro/SS-A negative primary Sjögren’s syndrome (SS).

Methods

This is a proof-of-concept, case-control study of SS, healthy (HC) and other disease (OD) controls. A discovery dataset of plasma samples (n=30 SS, n=15 HC) was tested on human proteome arrays containing 19,500 proteins. A validation dataset of plasma and stimulated parotid saliva from additional SS cases (n=46 anti-Ro⁺, n=50 anti-Ro^–), HC (n=42), and OD (n=54) was tested on custom arrays containing 74 proteins. For each protein, the mean+3SD of the HC value defined the positivity threshold. Differences from HC were determined by Fisher’s Exact test and random forest machine learning using 2/3 of the validation dataset for training and 1/3 for testing. Applicability of the results was explored in an independent rheumatology practice cohort (n=38 Ro⁺, n=36 Ro^–, n=10 HC). Relationships among antigens were explored using STRING interactome analysis.

Results

Ro⁺ SS parotid saliva contained autoantibodies binding to Ro60, Ro52, La/SS-B, and muscarinic receptor 5 (MR5). SS plasma contained 12 novel autoantibody specificities, 11 of which were detected in both the discovery and validation datasets. Binding to ≥1 of the novel antigens identified 54% of Ro^– SS and 37% of Ro⁺ SS cases, with 100% specificity in both groups. Machine learning identified 30 novel specificities showing receiver operator characteristic area under the curve of 0.79 (95% CI 0.64–0.93) for identifying Ro^– SS. Sera from Ro^– cases of an independent cohort bound 17 of the non-canonical antigens. Antigenic targets in both Ro⁺ and Ro^– SS were part of leukemia cell, ubiquitin conjugation, and antiviral defense pathways.

Conclusion

We identified antigenic targets of the autoantibody response in SS that may be useful for identifying up to half of Ro seronegative SS cases.

INTRODUCTION

Primary Sjögren’s syndrome (SS) is a systemic rheumatic autoimmune disorder characterized by focal lymphocytic infiltrates in the salivary and lacrimal glands, chronic severe dry mouth and eyes, pain, fatigue, and reduced quality of life.[1, 2] The etiology of SS includes genetic predisposition, epigenetic and environmental factors.[3, 4]

Cardinal features of SS diagnosis or classification require objective measures of ocular and/or oral dryness and either i) the presence of serum autoantibodies to Ro/SS-A[5, 6] and/or La-SS-B[7] or ii) focal lymphocytic sialoadenitis as defined by a focus score ≥1.[5–8] Approximately one third of SS cases lack serum Ro and La autoantibodies and thus require a minor salivary gland lip biopsy for definitive classification.[9] The requirement for labial salivary gland biopsy to classify SS in Ro/La antibody negative individuals is an important contributing factor to delayed diagnosis of SS, as this procedure is not readily available at many clinical sites. One promising new approach for classifying Ro antibody negative (Ro^–) SS cases is salivary gland ultrasound (SGUS), which has the added benefit of enabling the measurement of disease-associated changes in the major salivary glands. While SGUS has shown good correlation with labial salivary gland biopsy results, a barrier to widespread implementation in the clinic is high inter-observer variability.[10] Thus, additional approaches for confirming SS in Ro^– patients are needed.

Previous work has shown that autoantibodies secreted by minor salivary gland plasmablasts of SS patients are also present in serum.[11] The presence of salivary gland focal lymphocytic infiltrates[5–7] and positive serum anti-nuclear autoantibody (ANA) titers[12] in many Ro^– SS cases suggests that autoantibodies of unknown specificity may be present in the circulation and/or saliva of these individuals. Non-Ro/La autoantibodies have been described in SS. Autoantibodies to muscarinic receptor 3 (MR3) that may contribute to salivary gland dysfunction have been identified by a number of groups,[13] but these antibodies generally occur in Ro antibody seropositive (Ro⁺) individuals.[14] A trio of SS antigens were first identified in mouse models and subsequently observed in human patients, including mouse salivary protein 1 (SP-1), carbonic anhydrase 6 (CA6) and parotid secretory protein (PSP).[15] Antibodies to mouse SP-1 were initially reported to occur in over 40% of Ro/La seronegative SS patients and less than 5% of healthy controls.[15] Evaluation in an independent cohort of 123 SS cases showed SP-1 antibodies co-occurred with anti-Ro/La in 34% of cases, in isolation in 19%, and present in 18% of matched regional healthy controls.[16] However, the human parotid antigen recognized by these antibodies remains unknown. Other autoantibodies reported in SS were recently summarized.[17, 18]

Few studies have attempted unbiased, systematic searches for autoantigens in Ro/La seronegative SS. Yuan, et al. explored novel autoantibody specificities in SS using Phage Immuno Precipitation technology[19] that detects antibodies to peptides up to 90 amino acids long.[20] This meeting abstract reported two novel antigens; however, some conformational B cell epitopes may not be readily detected by this approach. In the present study, we used human proteome arrays containing mostly full-length, yeast-expressed recombinant proteins covering >75% of the human proteome to search for novel autoantibody specificities in Ro⁺ and Ro^– SS cases. We hypothesized that we would identify novel autoantibody specificities useful for identifying Ro^– SS patients without an invasive lip biopsy and/or that may be biomarkers of aberrant biological pathways.

METHODS

Study population and samples

Biospecimens and clinical data were derived from subjects previously enrolled and evaluated in the Oklahoma Sjögren’s Research Clinic at the Oklahoma Medical Research Foundation (OMRF, Oklahoma City, USA).[21] Data and samples of 126 self- or clinician-referred subjects, collected from 2004–2020 and meeting 2002 American European Consensus Group (AECG) classification criteria for primary SS [7] were split into a discovery cohort (n=30) and a validation cohort (n=96). The disease comparison group included subjects evaluated for SS at the same Oklahoma Sjögren’s Research Clinic but who met classification criteria for other rheumatic diseases (OD, n=45; Table 1 and SI Table 1). Plasma from demographically matched healthy controls (HC; discovery n=15, validation n=50) was obtained from the Oklahoma Rheumatic Disease Research Cores Center. The HC underwent collection of stimulated parotid saliva, and their unaffected status was corroborated by completion of the Connective Tissue Disease Screening Questionnaire;[22] eight with possible or probable rheumatic disease were excluded from further study. Parotid saliva was stimulated with 2% citric acid and collected for 5 min using Lashley cups as described.[23] Plasma samples from additional OD patients with multiple sclerosis (MS; n=9) were collected at the Oklahoma Multiple Sclerosis Center of Excellence. Diagnosis of MS was performed by a board-certified neurologist using the McDonald criteria.[24] Data and serum samples from 74 patients meeting the 2016 American College of Rheumatology/European League Against Rheumatism criteria for SS in the setting of an independent rheumatology practice [25] and 10 HC were obtained from Johns Hopkins University Medical Institute (JHU). Membership in the OMRF Ro⁺ and Ro^– cohorts was determined by Ro60 and/or Ro52 positivity using the Bioplex 2200 ANA test for combined SSA antigens (Bio-Rad) or slot blot for Ro60 and Ro52 (Inno-Lia). Membership in the JHU Ro⁺ and Ro^– groups was determined by Ro60 and/or Ro52 positivity using commercial ELISA and immunoprecipitation as described.[25] Descriptive features of all participants are shown in Tables 1 and 2 and SI Tables 1 and 2. The study was approved by the Oklahoma Medical Research Foundation and Johns Hopkins Medicine Institutional Review Boards, and all participants gave informed consent.

Table 1.

Demographics and Clinical Features of Cases and Controls.

	Discovery Array Study			Validation Array Study
Demographics	HC (n=15)	Ro+ (n = 15)	Ro– (n = 15)	HCg (n = 42)	Ro+g (n=46)	Ro– (n=50)	Other Diseaseg (n=54)
Age (Mean (SD))	56 (12)	57 (12)	57 (12)	50(14)	49(14)	50(14)	49(14)
Race (%)
White	9/15 (60)	9/15 (60)	9/15 (60)	27/39 (69)	31/46 (67)	36/50 (72)	32/51 (63)
NatAm/>one a	6/15 (40)	6/15 (40)	6/15 (40)	11/39 (28)	14/46 (30)	14/50 (28)	16/51 (31)
Black/Asian	N/A	N/A	N/A	1/39 (3)	1/46 (2)	0/50 (0)	3/51 (6)
Gender (% Female)	14/15 (93)	14/15 (93)	14/15 (93)	39/42 (93)	43/46 (93)	47/50 (94)	49/54 (91)
Clinical Features
FS+ (%) b		14/15 (93)	15/15 (100)		28/46 (61)	36/50 (72)	5/38 (13)
Anti-Ro+ (%) c		15/15 (100)	0/15 (0)		46/46 (100)	0/50 (0)	5/45 (11)
Anti-La+ (%) c		5/15 (33)	1/15 (7)		22/46 (48)	18/50 (36)	3/45 (7)
WUSF+ (%) d		6/15 (40)	9/15 (60)		25/46 (54)	30/50 (60)	13/45 (29)
Schirmer’s+ (%) e		4/15 (27)	10/15 (67)		13/46 (28)	20/50 (40)	7/45 (16)
vBS+ (%) f		8/15 (53)	12/15 (80)		31/46 (67)	25/50 (50)	10/45 (22)
ESSDAI
Low (0–4) (n, (median(range))		15, 1(0–3)	15, 1(0–4)		40, 2(0–4)	38, 0(0–4)	N/A
High (≥5) (n, (median(range))		N/A	N/A		6, 8(6–17)	12, 8(5–15)	N/A

^aNative American and more than one race

^bfocus score ≥1

^cPositive by Bioplex value ≥1, InnoLia, and/or medical record

^dwhole unstimulated salivary flow ≤1.5mL/15 min

^eSchirmer’s test≤5mm/5min

^fvan Bijsterveld score, maximum value ≥4.

^gData unavilable for: race (n=3 HC, n=3 Other disease), FS (n=1 Ro+, n=7 Other disease), clinical features for subjects with multiple sclerosis (n=9), ESSDAI for Other disease group.

Table 2.

Demographics and Clinical Features of Independent Cohort

Demographics	HC (n=10)	Ro+ (n = 38)f	Ro− (n = 36)f
Age (Mean (SD))	51 (9)	52 (14)	54 (10)
Race (%)
White	9/10 (90)	35/38 (92)	34/36 (94)
Black/Asian/Pacific Islander	1/10 (10)	3/38 (8)	2/36 (6)
Gender (% Female)	9/10 (90)	34/38 (89)	33/36 (92)
Clinical Features
FS+ (%) a		25/35 (71)	30/30 (100)
Anti-Ro+ (%) b		38/38 (100)	0/36 (0)
Anti-La+ (%) b		19/38 (50)	2/36 (6)
WUSF+ (%) c		15/16 (94)	19/20 (95)
Schirmer’s+ (%) d		27/36 (75)	23/35 (66)
OSS+ or vBS+ (%) e		8/14 (57)	3/14 (21)
ESSDAI
Low (0–4) (n, (median(range))		20, 2(0–4)	21, 2(0–4)
High (≥5) (n, (median(range))		18, 9(5–28)	15, 7(5–22)

^afocus score ≥1

^bPositive by ELISA or immunoprecipitation

^cwhole unstimulated salivary flow ≤0.5mL/5 min

^dSchirmer’s test≤5mm/5min

^eAbnormal ocular staining score or van Bijsterveld score

^fData unavailable for: FS (n=3 Ro+, n=6 Ro−), WUSF (n=22 Ro+, n=16 Ro−), Schirmer’s (n=2 Ro+, n=1 Ro−), vBS (n=24 Ro+, n=22 Ro−)

Patient and public involvement

SS patients and the public were not involved in the planning or design of this study.

Proteome arrays

HuProt v. 3.2 (>19,500 proteins), and custom arrays (74 proteins) from CDI Laboratories, Inc. (Baltimore, MD, USA) were used according to manufacturer’s instructions. Yeast-expressed recombinant proteins on HuProt arrays are printed in duplicate and contain N-terminal GST and 6x-His tags. Plasma (1:500) and saliva (1:20) were diluted in blocking solution (5% bovine serum albumin/1X-Tris buffered saline-0.1% Tween-20, CDI Labs, catalog number CDIHPK-007) and washed with 1X-Tris buffered saline-0.1% Tween-20 (CDI Labs, catalog number CDIHPK-008). AlexaFluor647-conjugated anti-human IgG H+L secondary antibody (Jackson ImmunoResearch, catalog number 109–606-003) was used at 1:2000, which did not bind intact human IgM array control proteins. Although this reagent did not recognize purified IgA heavy chains by capillary Western blot, it did recognize purified human IgA and IgG light chains equally but at substantially (>10-fold) reduced intensity compared to IgG heavy chains (not shown). Arrays were scanned using a Genepix 4000B scanner, aligned using Genepix Pro 7 software, and results exported as .gpr and .txt files. To assist with array alignment, discovery arrays were additionally probed with anti-GST antibody at 1:5000 (Millipore, catalog number ABN116) or 1:2000 (Proteintech, catalog number 10000–0-AP), followed by Cy3-conjugated anti-rabbit IgG H+L secondary antibody (Invitrogen, catalog number A10520) at 1:1000, and scanned a second time. Resulting alignment settings were then applied to the initial scans using patient samples. Each version of array was probed with the secondary antibody alone to determine background binding.

Capillary Western blot

Proteins selected for validation studies were purchased as yeast-expressed proteins from CDI Laboratories. RPS29 was purchased from Novus Biological (wheat germ expressed, catalog number H00006235-P01). Capillary western blot assays were performed using the Wes system from Protein Simple, with proteins tested at 0.2–10 µg/mL, patient plasma tested at 1:10–1:500 dilution(s), and anti-human IgG secondary HRP antibody (ready to use, Protein Simple, DM-005). Monoclonal or polyclonal positive control antibodies were purchased from Novus (CBX3: NBP2–92962, FUT8: H00002530-B01P, GMNN: NBP2–16639, RPS: NBP1–57477, SnoN (SKIL): NBP2–45851, SOX5: NBP2–03766, and ZBTB46: H00140685-B01P) or Abcam (KLHDC8A: ab235419, MAPRE1: ab232970, NUP50: ab151567).

Statistical analysis

Plasma data from the discovery cohort (HuProt 3.2 arrays) were normalized to 27 pairs of control Igκ proteins using the Robust Linear Model[26] within the PAA Bioconductor Package[27] in R and log₂ intensity values generated (SI Figure 1A). Proteins bound by plasma Ig >3SD above the mean of HC were catalogued, and those bound by more SS cases than HC in either Ro⁺ or Ro^– groups, defined by Fisher’s exact p<0.1, were selected for inclusion on custom arrays (equating to ≥4 positive cases in each case group). Antigens recognized by ≥3 SS cases were also included if enriched in salivary glands (Human Protein Atlas). Plasma and saliva data from the validation cohort (custom arrays) were examined for high background by determining the mean of the Background 635nm Median value for all proteins in each subarray. Samples with mean>300 (n=9 saliva) were removed and remaining data were normalized to 3 pairs of duplicate control IgG spots (25, 6.25, and 1.5625ng/µL) using quantile normalization with the PAA Bioconductor package[27] in R and log₂ intensity values generated (SI Figure 1B). Common autoantibody specificities were defined as those that bound a given antigen >3SD above the mean of HC samples (using n=15 and n=42 HC in the discovery and validation datasets, respectively) and showed common binding in SS groups versus HC by Fisher’s Exact Test at p<0.1. Custom array data from the independent JHU cohort were quantile normalized with the OMRF validation cohort data and batch-corrected using Harmony.[28] Contingency tests used to compare numbers of antigens or numbers of subjects were chi-square or 2-tailed Fisher’s exact tests, respectively. Comparison of continuous clinical data between groups was assessed using the Mann-Whitney U test.

Machine Learning approach for Human Proteome data

The validation cohort proteome data were coded as binary where 0 indicated normalized expression value <3SD + mean of the validation cohort HC (n=42), and 1 if >3SD + mean of HC. The independent JHU cohort proteome data were similarly coded, using combined JHU and validation cohort HC (n=52) to establish positivity thresholds. The data were randomly split into 2/3 for training and 1/3 for testing sets in R. The RandomForest (RF) package in R[29] was used to select variable importance features and build a random model, regressing top 30 variable importance features, and other default parameters. The model was then used to predict groups in the test data set, with and without the canonical Ro and La autoantigens. The areas under the receiver opening characteristics (ROC) curves (AUCs) for multi-class classifications were calculated and compared using multiROC package in R for both the analyses.

HLA Imputation

Classical HLA alleles were imputed from genotyped SNPs using the R package HIBAG,[30] with default parameters used for all settings, or SNP2HLA.[31] The reference data for the European samples provided with the software was used to impute alleles in HLA-A, -B, -C, -DQA1, -DQB1, and -DRB1, -DPA1 (SNP2HLA only), and -DPB1.

RESULTS

Autoantibodies detected in SS plasma using human proteome arrays

Plasma from 30 individuals meeting criteria[7] for primary Sjögren’s syndrome (SS) was tested on human proteome arrays displaying >19,500 human proteins to identify autoantibodies. SS patients were divided into Ro⁺ (n=15) and Ro^– (n=15) groups, and demographically matched with each other and HC (n=15). Using normalized intensity values, the cases were examined for specificities bound by plasma immunoglobulin (Ig) above the positive threshold, and commonly bound antigens, defined by differences between each case group and HC at Fisher’s exact p<0.1,were selected for further validation testing. In addition to the canonical plasma Ro and La autoantibodies, we identified 63 novel antigens bound by plasma immunoglobulin (Ig) of Ro⁺ SS cases (mean of 18 specificities/case) or Ro^– SS cases (mean of 12 specificities/case), of which 8 novel antigens were identified in both groups: CCDC155, DDB1, MUM1L1, NFU1, RPS29, SOX5, TCP10 and ZNF655 (Fig 1). Furthermore, the presence of antibodies to at least one of the 8 shared novel antigens identified 87% of the Ro^– SS cases in this discovery dataset, compared to 7% with clinical Bioplex testing for La antibodies (SI Table 3).

Figure 1. Binding of novel and canonical antigens by plasma Ig of anti-Ro positive and anti-Ro negative cases from the discovery dataset.

Heatmap indicates specificities with normalized intensity values above the positive threshold (mean+3SD of HC values, green), identified in significant number of cases (≥4 cases/group, p<0.1, Fisher’s exact test, gene symbols used to refer to protein products).

Discovered SS autoantibodies replicated with validation studies

For validation studies, custom arrays displaying 74 proteins were tested with plasma and stimulated parotid saliva from larger groups of SS cases (n=46 Ro⁺, n=50 Ro^–), HC (n=42), and OD controls (n=54). Of the 74 proteins, 73 were selected from the discovery dataset and were either bound by at least 4 subjects in either Ro+ or Ro^– groups (Fisher’s exact p<0.1 compared to HC, 66 proteins) or bound by 3 subjects and enriched in the salivary gland according to the Human Protein Atlas (7 proteins) (SI Figure 2). In addition, muscarinic receptor (MR)5 was included based on separate, early preliminary data (not shown). The normalized data were evaluated for antibody specificities as described with the discovery dataset.

Overall, the majority of cases bound ≤10 specificities in plasma and saliva (SI Table 4), with a mean of 3 and 4 in Ro⁺ and Ro^– plasma, respectively, and a mean of 2 and 3 in Ro⁺ and Ro^– saliva, respectively. Individual-level plasma antibody results, as well as clinical serologic test results for antibodies to the canonical SS antigens, are summarized in SI Table 5. Plasma antibodies of a few individuals bound to 10 or more antigens, which could reflect polyreactivity. In plasma, we identified 16 specificities excluding the Ro and La antigens, three of which were commonly bound by plasma Ig in the OD group only (ARFGAP1, NFU1, and PML) (Fig 2A). Of the remaining 13 specificities, 11 (GMNN, GRAMD1A, KLHDC8A, MAPRE1, NUP50, POLR3H, RCAN3, RPAP3, SKIL, TCP10, and ZBTB46) were commonly bound by antibodies in ≥4 cases in the discovery dataset, while ISG15 was recognized by antibodies from 3 cases and enriched in the salivary gland, thus confirming these antibodies in larger groups of subjects. Except for 9 proteins, all antigens identified in the discovery dataset were bound by plasma Ig from at least one SS case, with more than 20 proteins bound by at least 5 SS cases (SI Figure 3). Surprisingly, five Ro^– cases bound TROVE2 (Ro60) on the array, which may be due to differences in assay sensitivity or source species of proteins. Only the Ro and La antigens and MR5 were commonly bound by salivary Ig in SS cases (Fig 2B).

Figure 2. (A-B) Binding of novel and canonical antigens by plasma and saliva Ig of anti-Ro positive and anti-Ro negative cases of validation dataset.

(A) Plasma, (B) Stimulated parotid saliva. Green indicates specificities with normalized intensity values above the positive threshold (mean+3SD) of HC values. Gray indicates saliva samples not available or excluded due to high background. In each figure, upper panels indicate specificities significantly bound by SS cases, lower panels indicate specificities significantly bound by OD controls only (p<0.1, Fisher’s exact test, gene symbols used to refer to protein products).

We used capillary western blot to confirm the binding of plasma antibodies to select proteins, including 8 proteins commonly bound by the SS group in the validation dataset and/or identified by random forest analyses (see below): CBX3, FUT8, GMNN, KLHDC8A, MAPRE1, NUP50, SKIL, and ZBTB46. Two other proteins (RPS29 and SOX5) selected from the discovery dataset did not replicate in the validation dataset when testing nine samples in parallel on both sets of arrays (data not shown), suggesting array-dependent differences in protein preparations. Each protein was tested with plasma from three subjects that bound the selected protein on the arrays and two HC that did not react to any of the selected proteins. With this assay, we confirmed plasma IgG antibodies to SOX5, FUT8, and GMNN exclusively in the SS cases. (Fig 3).

Figure 3. Capillary western blot of plasma IgG binding SOX5, FUT8, and GMNN.

Binding of patient IgG antibodies to select proteins (yeast expressed, CDI). Commercial monoclonal (SOX5) or polyclonal (FUT8, GMNN) antibodies were used as positive controls (PC). Capillaries were loaded with 0.2µg/mL (SOX5) or 10µg/mL (FUT8, GMNN) protein solutions. Subject IDs and plasma dilutions listed at the top of each panel. HC=healthy control. Gene symbols used to refer to protein products.

Machine Learning-Identified Specificities for Predictive Model of SS

Using the specificities identified in the validation dataset, we selected commonly recognized antigens (Fisher’s exact p<0.1) in SS cases that were not commonly recognized by plasma Ig from OD or HC groups to create a panel to identify SS in cases without Ro and La autoantigens. This panel of 12 antigens (Fig 2a upper panel, excluding MR5) correctly identified 54% of Ro^– SS cases with 100% specificity, compared to 32% with Bioplex clinical testing for La antibodies (SI Table 3). Though subjects were selected using the 2002 AECG criteria, we also applied the panel to the subset of 35 Ro^– SS cases who also met the 2016 ACR/EULAR criteria, excluding all of the Ro^–La⁺ cases meeting AECG but not 2016 ACR/EULAR criteria, and found that the panel identified 46% of Ro^– SS cases with 100% specificity, compared to 9% by Bioplex (SI Table 3).

Next, we applied machine learning to identify specificities for a predictive model of SS, using 2/3 of the validation dataset for training and 1/3 for testing. A random forest analysis was performed including (Fig 4 A) and excluding (Fig. 4B) canonical Ro and La antigen data from the proteome arrays and receiver operating curves (ROCs) generated. In both analyses, thirty features were identified that could predict SS (Fig 4 and SI Table 6), several of which were also commonly recognized in the discovery dataset. Ro^– SS could be distinguished from HC and OD controls with an area under the curve (AUC) of 0.88 (95% CI 0.78–0.96) with Ro and La array binding data and AUC of 0.79 (95% CI 0.64–0.93) without the canonical antigens. Unsurprisingly, the heterogeneous OD group could not be predicted as an entity.

Figure 4. (A-B) Receiver Operating Curves including (A) and excluding (B) canonical Ro and La autoantigen proteome array binding data.

Machine learning predicted specificities that distinguished the Ro^– SS case group from the Ro⁺ SS case group, Healthy controls (HC) , and Other Disease (OD) controls. Areas under the curve (AUC) and 95% confidence intervals are shown.

Assessment of an independent rheumatology practice cohort

We next assessed applicability of the discovered non-canonical specificities to a previously described [25] rheumatology practice cohort (Table 2) by screening serum antibodies with the custom arrays. Individual-level results, as well as clinical serologic test results for antibodies to the canonical SS antigens, are summarized in SI Table 7. The independent JHU Ro^– case group (n=36) had serum antibodies to 17 of the 70 non-canonical antigens, compared to binding to only 2 of the 70 antigens in 10 HC from the same site (χ²=13.7, p=0.0002; Fig 5). Of the 17 non-canonical antigens bound, 9 were among those identified by random forest machine learning to predict the Ro^– cases in the validation cohort shown in Fig 4B and SI Table 6B. The independent Ro⁺ case group (n=38) had serum antibodies binding to 39 of the 70 non-canonical antigens compared to binding to 6 of the 70 antigens by 10 HC from the same site (χ²=35.6, p<0.0001; Fig 5). On an individual case basis, 50% (18 of 36) of the independent Ro^– case group had serum antibodies binding to one or more of the 17 non-canonical antigens recognized by this group, similar to the frequency with which non-canonical antigens identify the Ro^– validation cohort, while 2 of 10 independent HC bound at least one of the antigens (p=0.15, Fisher’s exact test). Of the n=10 JHU and n=42 OMRF HC cohorts combined, 12 had serum or plasma antibodies binding to at least one of the 17 antigens, distinguishing HC from the JHU Ro^– cases (p=0.012). Comparison of the frequency with which the independent Ro⁺ cases recognized one or more of the 17 non-canonical antigens (39%) to the combined HC group did not distinguish the groups (p=0.108). However, the frequency with which the independent Ro⁺ cases recognized one or more of the 39 non-canonical antigens identified by that group (55%) did differ from that of the combined HC group (29%; p=0.016). When considered individually, none of the antigens recognized by the JHU cases were more commonly bound compared to the JHU HC group, as defined by Fisher’s exact p<0.1.

Figure 5. Binding of novel and canonical antigens by plasma Ig of anti-Ro positive and anti-Ro negative cases from an independent rheumatology practice cohort.

Heatmap indicates specificities with normalized intensity values above the positive threshold (mean+3SD of HC values, green).

Application of the random forest models described in Fig 4B and SI Table 6B failed to predict the JHU Ro^– case group (not shown), indicating differences in non-canonical antigen recognition between the cohorts. However, training of a new random forest machine learning model using 2/3 of the JHU SS cases and the combined JHU and OMRF HC, followed by testing on 1/3 of the independent cases and combined HC in the absence of Ro and La proteome array binding data resulted in little but detectable predictive power for the JHU Ro^– cases (ROC AUC 0.61 (95% CI 0.31–0.58); SI Fig 4A; SI Table 8A) that was further enhanced by inclusion of the canonical Ro60, Ro52 and La binding data (ROC AUC 0.71 (95% CI 0.37–0.62; SI Fig 4B; SI Table 8B). Thus, while an independent SS cohort had serum antibodies to some of the newly discovered non-canonical antigens, they were not sufficient to predict SS status, and further discovery studies incorporating multiple SS cohorts are required.

Non-canonical SS antibody positivity associated with elevated serum IgM levels in Ro^– subjects

Next, we examined the data to determine if clinical features of SS were associated with antibody binding to the novel antigens. Association of ANA positivity with positivity for novel specificities in aggregate was explored in the subset of 26 Ro^– cases with plasma antibody reactivity to one or more non-canonical antigens and lacking reactivity to clinically important nuclear antigens by the Bioplex 2200 ANA screening test, of which 18 were considered ANA⁺ (titer ≥40); however, no association was found (not shown). Commercial antibodies to CBX3 showed a homogenous nuclear pattern, while commercial antibodies to GMNN, NUP50 and SOX5 showed nuclear speckled patterns (SI Table 9). Although limited in number, a few individuals with reactivity to these specificities but lacking reactivity to clinically important nuclear antigens showed matching patterns, possibly explaining their positive ANA tests (SI Table 9).

Using the specificities identified in the random forest analysis excluding Ro and La (Fig 4B and SI Table 6B), subjects were split into positive and negative groups, based on whether they bound ≥1 specificity in the panel. Clinical features were then compared between the positive and negative groups. We observed increased serum IgM levels in the combined Ro⁺/Ro^– subjects with plasma antibody binding to ≥1 specificity of the panel (p=0.015). Upon further examination, we found increased serum IgM levels in the Ro^– group (p=0.0074, Fig 6A), but not in the Ro⁺ group (Fig 6B). Ro⁺ cases binding to one or more of the same 30 antigens had elevated serum IgG (SI Fig 5). No other clinical associations were observed except increased frequency of anti-La positivity in the Ro^– SS cases (SI Fig 5). We also evaluated potential clinical associations with antibody reactivity to MR5 in the Ro⁺ group and observed association of this antibody specificity with increased disease severity by multiple clinical measures, including increased focus score, van Bijsterveld’s score, and activity in the ESSDAI biological domain (Fig 6C). In the independent JHU cohort, 10 of the 38 Ro⁺ cases had serum antibodies to MR5, and this was associated with an abnormal Schirmer tear flow test (p=0.039) and higher ANA titers (p=0.051) compared to the Ro⁺ patients lacking this specificity (not shown). Insufficient data precluded the evaluation of the association of MR5 positivity with salivary flow in this cohort. In both cohorts, the vast majority of patients with MR5-binding antibodies (and definitive Ro52 and Ro60 clinical test results), also showed antibody reactivity to both Ro52 and Ro60 antigens (89% [17/19] and 100% [10/10] in the OMRF validation and JHU cohorts, respectively; SI Tables 5 and 7). However, there was no difference in the incidence of MR5 antibody binding between individuals with antibodies to Ro52, Ro60 and La (87% [13/15] and 36.8% [7/19] in the OMRF and JHU cohorts, respectively) compared to those with antibodies to Ro52 and Ro60 without anti-La antibodies (80% [4/5] and 25% [3/12] in the OMRF and JHU cohorts, respectively).

Figure 6. (A-C) Clinical correlations with non-canonical antibody specificities.

SS validation cohort patients with plasma Ig binding to at least 1 of 30 antigens identified by machine learning (see Fig 4B and SI Table 6B) exhibit increased serum IgM levels which were found to be elevated in Ro^– patients (A), but not in Ro⁺ patients (B). Ro⁺ SS patients with Ig binding to MR5 exhibited more severe SS by multiple measures (C). (Mann-Whitney or Fisher’s exact tests, p<0.05).

Because HLA is strongly associated with SS,[32] with the highest signals accounted for by MHC class II alleles, we explored potential HLA associations with reactivity to the newly identified antigens in a subset of subjects from whom HLA data were available (n=40 Ro⁺ and n=41 Ro^– cases) and observed positive associations between HLA Class II alleles and several specificities (SI Fig 6). Interestingly, antibodies to TCP10, Ro52 and Ro60 were associated with the DR3-DQ2 haplotype (SI Fig 6).

Protein interactions and tissue expression of specificities determined with STRING and Protein Atlas

Finally, we explored protein interactions and expression of the antigens using the STRING and Human Protein Atlas databases, respectively. Important specificities from the random forest analyses (SI Table 6), as well as those commonly recognized by the SS, but not HC or OD groups (Fig 2), were entered into the STRING database to uncover pathways enriched for the identified antigens. Analyses were performed with (SI Fig 7, SI Table 10) and without (Table 3, SI Fig 8, SI Table 11) the Ro and La autoantigens. Leukemia cell, ubiquitin conjugation, and antiviral defense pathways were among the most significant results in both analyses. A query of the Human Protein Atlas revealed biased expression of certain antigens in particular tissues or cells, with EGFLAM, ISG15, TPD52L1, and TPI1 being enriched in salivary gland or tongue and RCAN3 involved in oral antimicrobial defense (SI Table 12). Other proteins showed expression in particular cell types of the immune system, brain, testis, prostate, ovary, or pancreas (SI Table 12).

Table 3.

STRING Interactome Analysis

Category	Term description	FDR	Antigens
TISSUES	Leukemia cell	0.00097	GAPDH, TPI1, GMNN, CRBN, DDB2, SMAD2, PML, CBX3, PDPK1, HNRNPAB, MAPRE1, NFU1
KEGG	Ubiquitin mediated proteolysis	0.002	RBX1, DDB2, ERCC8, PML, DDB1
Reactome	Post-translational protein modification	0.0095	USP18, RBX1, LMAN1, DDB2, SMAD2, ERCC8, PML, DDB1, ARFGAP1, NUP50, FUT8, SEC23IP, MCFD2
UniProt Keywords	Ubiquitin conjugation	0.0256	RPAP3, GAPDH, TPI1, CRBN, DDB2, SMAD2, PML, DDB1, CBX3, PDPK1, NUP50, HNRNPAB, ZBTB46, ZNF655, SKIL
UniProt Keywords	Antiviral defense	0.0394	POLR3B, PML, POLR3H, ISG15
WikiPathways	TGF-beta signaling pathway	0.0451	RBX1, SMAD2, PML, SKIL

Novel specificities are indicated in bold italics. Only pathways including at least two novel specificities are shown.

DISCUSSION

Under the current criteria, diagnosis of SS may be hindered or delayed for those lacking anti- Ro/SS-A antibodies, owing in part to lack of access to or lack of consent for the minor salivary gland biopsy procedure. In our validation cohort, one-third of the subjects (n=32/96) had a positive lip biopsy (focus score ≥1), yet were negative for Ro and La antibodies, suggesting other antigens are playing a role in the SS autoimmune response. Here, we identified novel autoantibodies present in multiple SS cases, including seronegative SS. Excluding the Ro and La canonical antigens and MR5, 55 of our 70 discovery dataset specificities (79%) were identified in >1 SS case in validation studies with larger groups of subjects. Of these 55 specificities, 33 (60%) were commonly bound by SS cases compared to HC and/or distinguished from HC by random forest analyses. Previous studies identifying non-Ro/La autoantibodies in SS were recently reviewed, [17]^,[18] with varying degrees of prevalence. Although we cannot comment on alpha-fodrin, aquaporin 3, or salivary protein 1, as they were not present on the discovery proteome arrays used, many of these antibodies were also identified in the present study, with anti-Cofilin 1, anti-PUF60, and anti-STMN4 occurring in the highest percentage of cases ranging from 10–17% (SI Table 13). However, these antibody specificities did not meet our criteria for common expression in SS cases in our cohort. Conversely, we failed to observe antibodies to at least 13 other specificities reported in the literature, including PSP/BPIFA2 and CA6. This is in agreement with the failure to detect antibodies to PSP by luciferase immunoprecipitation in another cohort of SS patients.[33]

A limitation of the current study is that proteins printed on the array may not display the proper conformation. This likely varies for each protein, as Ro60/TROVE2 and La/SSB reactivity on arrays showed high agreement with clinical tests (87–96%, SI Table 14), and select proteins were validated by capillary western blot. Notably, antibodies to Ro52/TRIM21 were poorly identified with custom arrays used for the validation (18% positive agreement, SI Table 14) and JHU datasets (0% positive agreement, not shown), apparently owing to a high mean+3SD threshold value in the HC group, suggesting a custom array protein conformation permissive for the known FcR binding activity of Ro52.[34] Membrane proteins may be particularly susceptible to conformational changes. Despite this, we observed a relatively high prevalence of antibodies binding to MR5 in the Ro⁺ SS group, and patients with this reactivity displayed more severe SS by multiple measures. Both of these features, namely association with anti-Ro reactivity and more severe disease, have been reported for SS patients having serum antibody reactivity to MR3.[35–39] We noted that the vast majority of patients with MR5-binding antibodies also produced serum autoantibodies to both Ro52 and Ro60, and this patient group has been shown to have increased prevalence of abnormal ocular surface staining, positive focus scores, anti-nuclear autoantibodies ≥1:320, positive La antibodies, rheumatoid factor, and elevated IgG.[25] Further studies will be required to determine if the MR5 reactivity observed in the present study represents cross-reactivity with MR3 or is specific to MR5. Though MR5 expression is restricted to the central nervous system, mice lacking MR5 expression showed defects in saliva production.[40]

A second limitation of this study is the relatively small number of SS cases (n=30) used for autoantibody discovery, making it unlikely that we captured the full range of SS heterogeneity. Though multiple non-canonical specificities were recognized by Ro^– SS cases from two independent cohorts, additional studies incorporating multiple cohorts are needed to identify additional non-canonical antigens that are broadly applicable across cohorts. Additional studies will also be required to address post-translational modifications, which are known to be involved in autoimmune responses.

With the exception of MR5 and ISG15, the non-canonical antigens identified herein are exclusively expressed intracellularly, suggesting that most of the discovered autoantibodies are unlikely to be directly involved in disease pathogenesis and are probably exposed to the immune system through cellular death/tissue damage.[41] Biased expression of a subset of the antigens in salivary, oral, or exocrine tissue, suggests that some of these antigens may participate in the targeting of salivary or lacrimal glands in SS. Interestingly, many of the novel antigenic targets and STRING database-identified pathways were found in both Ro^– and Ro⁺ SS, suggesting common mechanisms of their generation in both groups.

The most significantly enriched tissue in both SS groups was “leukemia cell”, suggesting the interesting possibility that some of these responses could arise through successful cancer cell immune surveillance, a phenomenon for which there is precedence in systemic sclerosis and myositis. In systemic sclerosis, antibodies to the POLR3A-encoded RPC1 protein arise along with the development of cancer in a subset of patients, with many patients showing POLR3A somatic mutations and loss of heterozygosity in their matching tumors.[42] Some patients with these antibodies do not develop cancer,[42] suggesting successful immunosurveillance. In dermatomyositis, autoantibodies to transcriptional intermediary factor 1-γ are associated with cancer diagnoses, and a recent study identified novel autoantibodies in this group that may protect against development of cancer.[43] Patients with SS have a substantially increased risk of developing non-Hodgkins lymphoma.[44] Although ISG15 was not among the discovered non-canonical SS antigens associated with “leukemia cell” by STRING analysis, a recent study has identified ISG15 gene expression as a biomarker for lymphoma in SS.[45] Thus, further examination of these specificities in SS patients who develop lymphoma is warranted.

Enrichment of the term “post-translational modification” among the newly described non-canonical antigens is also of potential interest, as post-translational modification of proteins can lead to loss of immunological tolerance, just as somatic mutations can lead to loss of tolerance to self-antigens. Though not identified as involved in post-translational modification by the STRING analysis, ISG15 is itself a ubiquitin-like post-translational modification that can be conjugated to intracellular proteins in a process known as ISGylation.[46] An important role for ISG15 is regulation of type 1 interferon signaling by stabilizing Ubl carboxy-terminal hydrolase 18, a negative regulator of the interferon α/β receptor 2 [47] and by promoting the autophagic degradation of retinoic acid-inducible gene I protein.[48] Further, an extracellular form of ISG15 can signal through LFA-1 to potentiate IL-12-induced production of interferon gamma and IL-10.[49] Moreover, ISG15 shows enhanced expression in salivary gland tissue (SI Table 12), and increased ISG15 in salivary glands of SS patients has been reported.[45] It is possible that autoantibodies to ISG15 could modulate its extracellular function, either by stabilizing or blocking its effects.

In conclusion, using large-scale, unbiased screens, we identified and validated a panel of non-canonical autoantigens in Ro^– SS cases. This study provides proof-of-concept for being able to use autoantibody biomarkers to recognize primary SS in Ro^– individuals without a lip biopsy in the context of current classification criteria.

Key Messages.

What is already known about this subject?

• Knowledge of non-canonical SS antigens, particularly in anti-Ro antibody negative (Ro^–) SS is limited.

What does this study add?

• This proof-of-concept study used a research cohort to identify 12 novel antigens commonly bound by plasma IgG of individuals with Ro^– primary SS compared to healthy controls.

• Of the 12 novel antigens, 9 were commonly bound in the Ro^– SS group, 1 was commonly bound in the Ro⁺ SS group, and 2 were commonly bound by both groups.

• Plasma autoantibodies to at least one of the 12 novel antigens identified Ro^– SS with 100% specificity and 54% sensitivity, while 30 novel autoantibodies identified by machine learning distinguished the Ro^– SS group with ROC AUC 0.79 (95% CI 0.64–0.93).

• Many of the novel antigens were bound by serum antibodies from an independent SS rheumatology practice cohort.

How might this impact on clinical practice or future developments?

• Plasma autoantibody profiling is a promising approach for classifying a substantial proportion of Ro^– SS cases without a minor salivary gland lip biopsy.

Data availability statement

Data are available upon reasonable request.