Individuals with APS1 develop autoantibodies to many known protein targets, some of which exhibit tissue-restricted expression and have been shown to correlate with specific autoimmune disease manifestations. However, the target proteins for many of the APS1 tissue-specific manifestations remain enigmatic. To this end, we employed a high-throughput, proteome-wide programmable phage display approach (PhIP-Seq) to query the antibody target identities within serum of people with APS1 (Larman et al., 2011; O’Donovan et al., 2018). The PhIP-Seq technique leverages large scale oligo production and efficient phage packaging and expression to present a tiled-peptide representation of the proteome displayed on T7 phage. Here, we utilize a phage library that we previously designed and deployed for investigating paraneoplastic autoimmune encephalitis (Mandel-Brehm et al., 2019; O’Donovan et al., 2018). The library itself contains approximately 700,000 unique phage, each displaying a 49 amino acid proteome segment. As previously described, phage were immunoprecipitated using human antibodies bound to protein A/G beads. In order to increase sensitivity and specificity for target proteins, eluted phage were used for a further round of amplification and immunoprecipitation. DNA was then extracted from the final phage elution, amplified and barcoded, and subjected to Next-Generation Sequencing (Figure 1A). Finally, sequencing counts were normalized across samples to correct for variability in sequencing depth, and the fold-change of each gene was calculated (comprised of multiple unique tiling phage) as compared to mock IPs in the absence of human serum (further details of the protocol can be found in the methods section).

From a cohort of 67 APS1 serum samples, a total of 39 samples were subjected to PhIP-Seq investigation, while the remaining 28 samples were obtained at a later time point and reserved for downstream validation experiments (for clinical data, refer to Supplementary file 1). In addition, 28 non-APS1 anonymous blood donor serum samples were subjected to PhIP-Seq, and an additional group of 61 non-APS1 plasma samples were used for downstream validation experiments (Supplementary file 2).

PhIP-Seq results were first cross-referenced with previously reported APS1 autoantibody targets (Alimohammadi et al., 2008; Alimohammadi et al., 2009; Clemente et al., 1997; Fishman et al., 2017; Hedstrand et al., 2001; Husebye et al., 1997; Kluger et al., 2015; Kuroda et al., 2005; Landegren et al., 2015; Landegren et al., 2016; Leonard et al., 2017; Meager et al., 2006; Meyer et al., 2016; Oftedal et al., 2015; Pöntynen et al., 2006; Sansom et al., 2014; Shum et al., 2009; Shum et al., 2013; Söderbergh et al., 2004). To avoid false positives, a conservative set of criteria were used as follows. We required a minimum of 2/39 APS1 samples and 0/28 non-APS1 control samples to exhibit normalized gene counts in the immunoprecipitation (IP) with greater than 10-fold enrichment as compared to the control set of 17 mock-IP (beads, no serum) samples. This simple, yet stringent criteria enabled detection of a total of 23 known autoantibody specificities (Figure 1B). Importantly, many of the well-validated APS1 antigens, including specific members of the cytochrome P450 family (CYP1A2, CYP21A1, CYP11A1, CYP17A1), lung disease-associated antigen KCRNG, as well as IL17A, IL17F, and IL22, among others were well represented (Figure 1B). In contrast, the diabetes-associated antigens GAD65 and INS did not meet these stringent detection criteria and only weak signal was detected to many of the known interferon autoantibody targets known to be present in many people with APS1, perhaps due to the conformational nature of these autoantigens (Figure 1B and Figure 1—figure supplement 1; Björk et al., 1994; Meager et al., 2006; Meyer et al., 2016; Wolff et al., 2013; Ziegler et al., 1996).

Three known autoantigens that were prevalent within our cohort were selected to determine how PhIP-Seq performed against an orthogonal whole protein-based antibody detection assay. A radioligand binding assay (RLBA) was performed by immunoprecipitating in vitro transcribed and translated S35-labeled proteins CYP11A1, SOX10, and NLRP5 with APS1 serum (Alimohammadi et al., 2008; Berson et al., 1956; Hedstrand et al., 2001; Winqvist et al., 1993). Importantly, and in contrast to PhIP-Seq, this assay tests for antibody binding to full-length protein (Figure 1C). By RLBA, these three antigens were present in and specific to both the initial discovery APS1 cohort (n = 39) as well as the expanded validation cohort (n = 28), but not the non-APS1 control cohort (n = 61). Together, these results demonstrate that PhIP-Seq detects known APS1 autoantigens and that PhIP-Seq results validate well in orthogonal whole protein-based assays.

To determine whether the PhIP-Seq APS1 dataset could yield higher resolution information on antigenic peptide sequences with respect to previously reported targets, the normalized enrichments of all peptides belonging to known disease-associated antigens CYP11A1 and SOX10 were mapped across the full length of their respective proteins (Figure 1—figure supplement 2). The antigenic regions within these proteins were observed to be similar across all samples positive for anti-CYP11A1 and anti-SOX10 antibodies, respectively (Figure 1—figure supplement 2) suggesting peptide-level commonalities and convergence among the autoreactive antibody repertoires across individuals. These data suggest that people with APS1 often target similar, but not identical protein regions.

Having confirmed that PhIP-Seq analysis of APS1 sera detected known antigens, the same data were then investigated for the presence of novel, previously uncharacterized APS1 autoantigens. We applied the same positive hit criteria as described for known antigens, and additionally increased the required number of positive APS1 samples to 3/39 to impose a stricter limit on the number of novel candidate autoantigens. This yielded a list of 81 genes, which included 12 known antigens and 69 putative novel antigens (Figure 2).

The most commonly held hypotheses regarding the nature and identity of proteins targeted by the aberrant immune response in APS1 are that targeted proteins (1) tend to exhibit AIRE-dependent thymic expression and (2) have restricted expression to one or few peripheral organs and tend not to be widely or ubiquitously expressed. We investigated whether our novel antigens were also preferentially tissue-restricted. In order to systematically address this question, tissue-specific RNA expression was assessed using a consensus expression dataset across 74 cell types and tissues (Uhlén et al., 2015). For each gene, the ratio of expression in the highest tissue as compared to the sum of expression across all tissues was calculated, resulting in higher ratios for those mRNAs with greater degrees of tissue-restriction. Using this approach, the mean tissue-specificity ratio of the 81 PhIP-Seq positive antigens was increased by approximately 1.5-fold (p=0.0017) as compared to the means from iterative sampling of 81 genes (Figure 2—figure supplement 1).

Identified autoantigens were ranked by frequency within the cohort. Five antigens were positive in ten or more APS1 samples, including two novel antigens. In addition, the majority of antigens found in 4 or more APS1 sera were novel (Figure 3A). Five of the most frequent novel antigens were selected for subsequent validation and follow-up. These included RFX6, a transcription factor implicated in pancreatic and intestinal pathology (Patel et al., 2017; Smith et al., 2010); ACP4, an enzyme implicated in dental enamel hypoplasia (Choi et al., 2016; Seymen et al., 2016; Smith et al., 2017); KHDC3L, a protein with oocyte-restricted expression (Li et al., 2008; Zhang et al., 2018; Zhu et al., 2015); NKX6-3, a gastrointestinal transcription factor (Alanentalo et al., 2006); and GIP, a gastrointestinal peptide involved in intestinal motility and energy homeostasis (Adriaenssens et al., 2019; Jörnvall et al., 1981; Moody et al., 1984; Pederson and McIntosh, 2016). Several less frequent (but still shared) novel antigens were also chosen for validation, including ASMT, a pineal gland enzyme involved in melatonin synthesis (Ackermann et al., 2006; Rath et al., 2016); and PDX1, an intestinal and pancreatic transcription factor (Holland et al., 2002; Stoffers et al., 1997; Figure 3A). Of note, this group of seven novel antigens all exhibited either tissue enriched, tissue enhanced, or group enhanced expression according to the Human Protein Atlas database (https://www.proteinatlas.org/about/download; Uhlén et al., 2015; Supplementary file 3). Using a whole-protein radiolabeled binding assay (RLBA) for validation, all seven proteins were immunoprecipitated by antibodies in both the PhIP-Seq APS1 discovery cohort (n = 39), as well as in the validation cohort of APS1 sera that had not been interrogated by PhIP-Seq (n = 28). Whereas an expanded set of non-APS1 controls (n = 61) produced little to no immunoprecipitation signal by RLBA as compared to positive control antibodies (low antibody index), APS1 samples yielded significant immunoprecipitation signal enrichment for each whole protein assay (high antibody index) (Figure 3B and Supplementary file 4).

The comparison of PhIP-Seq data to the results from the RLBAs (n = 39, discovery cohort only) yielded positive correlations between the two datasets (r = 0.62–0.95; Figure 3—figure supplement 1). Notably, for some antigens, such as NLRP5, and particularly for ASMT, the RLBA results revealed additional autoantibody-positive samples not detected by PhIP-Seq (Figure 3B and Figure 3—figure supplement 1 and Figure 1—figure supplement 2).

Because the individuals in this APS1 cohort have been extensively phenotyped for 24 clinical manifestations, the PhIP-Seq APS1 data was queried for phenotypic associations. Several autoantibody specificities, both known and novel, were found to possess highly significant associations with several clinical phenotypes (Figure 4 and Figure 4—figure supplement 1). Among these were the associations of KHDC3L with ovarian insufficiency, RFX6 with diarrheal-type intestinal dysfunction, CYP11A1 (also known as cholesterol side chain cleavage enzyme) with adrenal insufficiency (AI), and SOX10 with vitiligo (Figure 4). Strikingly, anti-CYP11A1 antibodies are present in AI and are known to predict disease development (Betterle et al., 2002; Petra et al., 2000; Winqvist et al., 1993). Similarly, antibodies to SOX10, a transcription factor involved in melanocyte differentiation and maintenance, have been previously shown to correlate with the presence of autoimmune vitiligo (Hedstrand et al., 2001).

Primary ovarian insufficiency is a highly penetrant phenotype, with an estimated 60% of females with APS1 progressing to an early, menopause-like state (Ahonen et al., 1990; Ferre et al., 2016). Interestingly, a set of 5 proteins (KHDC3L, SRSF8, PNO1, RASIP1, and MORC2) exhibited a significant association with ovarian insufficiency in this cohort (Figure 4). A publicly available RNA-sequencing dataset from human oocytes and supporting granulosa cells of the ovary confirmed that of these 5 genes, only KHDC3L exhibited expression levels in female oocytes comparable to the expression levels seen for the known oocyte markers NLRP5 and DDX4 (Zhang et al., 2018; Figure 4—figure supplement 2). We therefore chose to further investigate the relationship between anti-KHDC3L antibodies and ovarian insufficiency in our cohort (Figure 5).

KHDC3L is a well-studied molecular binding partner of NLRP5 within the ovary (Li et al., 2008; Zhu et al., 2015). Together, NLRP5 and KHDC3L form part of a critical oocyte-specific molecular complex, termed the subcortical maternal complex (SCMC) (Bebbere et al., 2016; Li et al., 2008; Liu et al., 2016; Zhu et al., 2015). Furthermore, knockout of the NLRP5 and KHDC3L in female mice results in fertility defects, and human genetic mutations in these genes of the SCMC have been linked to infertility and molar pregnancies (Akoury et al., 2015; Li et al., 2008; Reddy et al., 2013; Wang et al., 2018; Zhang et al., 2019). Interestingly, previous work established NLRP5 as a parathyroid-specific antigen in APS1, with potential for additional correlation with ovarian insufficiency (Alimohammadi et al., 2008; Brozzetti et al., 2015). However, anti-NLRP5 antibodies lack sensitivity for ovarian insufficiency. Importantly, unlike NLRP5, KHDC3L is expressed primarily in the ovary, and thus potentially represents a more oocyte-specific autoantigen (Liu et al., 2016; Virant-Klun et al., 2016; Zhang et al., 2018). Using the dataset from Zhang et. al, we confirmed that KHDC3L, as well as NLRP5 and the known oocyte marker DDX4, are highly expressed within the oocyte population, but not in the supporting granulosa cell types (Zhang et al., 2018; Figure 5A). Interestingly, the majority (64%) of APS1 sera had a concordant status for antibodies to KHDC3L and NLRP5 (Figure 5B). Although previous reports did not find a strong gender prevalence within samples positive for anti-NLRP5 antibodies, the mean anti-NLRP5 and anti-KHDC3L antibody signals were increased in females in this cohort (by 1.6- and 2.1-fold, respectively; Figure 5C). Finally, all 10 females in the expanded APS1 cohort with diagnosed ovarian insufficiency were also positive for anti-KHDC3L antibodies (Figure 5D).

Similar to known antigens CYP11A1, SOX10, and LCN1, the novel antigen ACP4 was found to occur at high frequencies in this cohort (Figure 3A). ACP4 (acid phosphatase 4) is highly expressed in dental enamel, and familial mutations in the ACP4 gene result in dental enamel hypoplasia similar to the enamel hypoplasia seen in ~90% of this APS1 cohort (Seymen et al., 2016; Smith et al., 2017). Strikingly, 50% of samples were positive for anti-ACP4 antibodies by RLBA, with excellent correlation between RLBA and PhIP-Seq data (Figure 3B and Figure 3—figure supplement 1A). Consistently, samples from individuals with enamel hypoplasia exhibited a trend towards higher anti-ACP4 antibody signal by RLBA (Figure 3—figure supplement 1B, p=0.064).

In this cohort, 82% (55/67) of APS1 sera exhibited an RFX6 signal that was at least 3 standard deviations above the mean of non-APS1 control signal due to the extremely low RFX6 signal across all non-APS1 controls by RLBA (Figure 3B). Using a more stringent cutoff for RFX6 positivity by RLBA at 6 standard deviations above the mean, 65% of APS1 samples were positive for anti-RFX6 antibodies. RFX6 is expressed in both intestine and pancreas, and loss of function RFX6 variants in humans lead to both intestinal and pancreatic pathology (Gehart et al., 2019; Patel et al., 2017; Piccand et al., 2019; Smith et al., 2010). Interestingly, across all samples with anti-RFX6 antibodies, the response targeted multiple sites within the protein, suggesting a polyclonal antibody response (Figure 6—figure supplement 1A).

The extent and frequency of intestinal dysfunction in people with APS1 has only recently been clinically uncovered and reported, and therefore still lacks unifying diagnostic markers as well as specific intestinal target antigen identities (Ferre et al., 2016). This investigation of APS1 sera revealed several antigens that are expressed in the intestine, including RFX6, GIP, PDX1, and NKX6-3. We chose to further study whether autoimmune response to RFX6+ cells in the intestine was involved in APS1-associated intestinal dysfunction. Using a publicly available murine single-cell RNA sequencing dataset of 16 different organs and over 120 different cell types, RFX6 expression was confirmed to be present in and restricted to pancreatic islets and intestinal enteroendocrine cells (Schaum et al., 2018; Figure 6—figure supplement 1B and C). Serum from an individual with APS1-associated intestinal dysfunction and anti-RFX6 antibodies was next tested for reactivity against human intestinal enteroendocrine cells, revealing strong nuclear staining that colocalized with ChromograninA (ChgA), a well-characterized marker of intestinal enteroendocrine cells (Goldspink et al., 2018; O'Connor et al., 1983; Figure 6A, right panel and inset). In contrast, enteroendocrine cell staining was not observed from APS1 samples that lacked anti-RFX6 antibodies or from non-APS1 control samples. (Figure 6A, center and left panels). Furthermore, serum from samples with anti-RFX6 antibodies stained transfected tissue culture cells expressing RFX6 (Figure 6B, Figure 6—figure supplement 2). These data support the notion that there exists a specific antibody signature, typified by anti-RFX6 antibodies, associated with enteroendocrine cells in APS1.

Both mice and humans with biallelic mutation of the gene encoding RFX6 have enteroendocrine cell deficiency and intestinal malabsorption (Mitchell et al., 2004; Piccand et al., 2019; Smith et al., 2010), and humans with other forms of genetic or acquired enteroendocrine cell deficiency also suffer from chronic malabsorptive diarrhea (Högenauer et al., 2001; Mitchell et al., 2004; Oliva-Hemker et al., 2006; Posovszky et al., 2012; Wang et al., 2006). In this cohort, 54/67 (81%) of individuals have intestinal dysfunction defined as the presence of chronic diarrhea, chronic constipation or an alternating pattern of both, without meeting ROME III diagnostic criteria for irritable bowel syndrome, as previously described (Ferre et al., 2016). When the cohort was subsetted by presence or absence of intestinal dysfunction, the anti-RFX6 RLBA signal was significantly higher when intestinal dysfunction was present (Figure 6C). Further subsetting of the cohort by subtype of intestinal dysfunction revealed that individuals with anti-RFX6 antibodies belonged preferentially to the diarrheal-type (as opposed to constipation-type) group of intestinal dysfunction (Figure 6D and Figure 6—figure supplement 3A). In contrast, antibodies to the known intestinal antigen TPH1 were distributed less specifically across both types (diarrheal- and constipation-type) of ID, despite a high frequency of both anti-TPH1 and anti-RFX6 in the cohort (55/67 for anti-RFX6, 45/67 for anti-TPH1) (Figure 6—figure supplement 3C and D). These observations are consistent with a previous report that anti-TPH1 antibodies show an association with ID, but not specifically with the diarrheal subtype (Kluger et al., 2015).

Given that RFX6 is also expressed in the pancreas, we also examined the association of anti-RFX6 antibodies with APS1-associated type 1 diabetes. We observed that 6/7 APS1-associated type 1 diabetes samples had positive signal for anti-RFX6 antibodies by RLBA (Figure 6—figure supplement 3B). However, due to small sample size, an expanded cohort would be needed to determine the significance of this observation. Together, these data suggest that RFX6 is a common, shared autoantigen in APS1 that may be involved in the immune response to intestinal enteroendocrine cells as well as pancreatic islets. Future studies will help to determine whether testing for anti-RFX6 antibodies possesses clinical utility for prediction or diagnosis of specific APS1 autoimmune disease manifestations as well as for non-APS1 autoimmune disease.

All patient cohort data was collected and evaluated at the NIH, and all APECED/APS1 patients were enrolled in a research study protocols approved by the NIAID, NIH Clinical Center, and NCI Institutional Review Board Committee and provided with written informed consent for study participation. All NIH patients gave consent for passive use of their medical record for research purposes (protocol #11-I-0187). The majority of this cohort data was previously published by Ferre et al. (2016) and Ferré et al. (2019).

For PhIP-Seq, we adapted a custom-designed phage library consisting of 731,724 49AA peptides tiling the full protein-coding human genome including all isoforms (as of 2016) with 25AA overlap as previously described (O’Donovan et al., 2018). 1 milliliter of phage library was incubated with 1 microliter of human serum overnight at 4C, and human antibody (bound to phage) was immunoprecipitated using 40 ul of a 1:1 mix of protein A/G magnetic beads (Thermo Fisher, Waltham, MA, #10008D and #10009D). Beads were washed 4 times and antibody-bound phage were eluted into 1 ml of E. coli at OD of 0.5–0.7 (BLT5403, EMD Millipore, Burlington, MA) for selective amplification of eluted phage. This library was re-incubated with human serum and repeated, followed by phenol-chloroform extraction of DNA from the final phage library. DNA was barcoded and amplified (Phusion PCR, 30 rounds), gel purified, and subjected to Next-Generation Sequencing on an Illumina MiSeq Instrument (Illumina, San Diego, CA).

Sequencing reads from fastq files were aligned to the reference oligonucleotide library and peptide counts were subsequently normalized by converting raw reads to percentage of total reads per sample. Peptide and gene-level enrichments for both APS1 and non-APS1 sera were calculated by determining the fold-change of read percentage per peptide and gene in each sample over the mean read percentage per peptide and gene in a background of mock-IP (A/G bead only, n = 17). Individual samples were considered ‘positive’ for genes where the enrichment value was 10-fold or greater as compared to mock-IP. For plotting of multiple genes in parallel (Figures 1 and 2), enrichment values were z-scored and hierarchically clustered using Pearson correlation.

For comparison of distribution of PhIP-Seq gene enrichment between APS1 patients with and without specific disease manifestations, a (non-parametric) Kolmogorov-Smirnov test was used.

For radioligand binding assays, antibody index for each sample was calculated as follows: (sample value – mean blank value) / (positive control antibody value – mean blank value). Comparison of antibody index values between non-APS1 control samples and APS1 samples was performed using a Mann-Whitney U test. Experimental samples that fell 3 standard deviations above of the mean of non-APS1 controls for each assay were considered positive, except in the case of RFX6, where a cutoff of 6 standard deviations above the mean of non-APS1 controls was used.

To determine tissue-specificity and tissue-restriction of Rfx6 expression in mice, we used publicly available Tabula Muris data (tabula-muris.ds.czbiohub.org) (Schaum et al., 2018). For investigation of KHDC3L expression in human ovary, we downloaded publicly available normalized FPKM transcriptome data from human oocytes and granulosa cells (GSE107746_Folliculogenesis_FPKM.log2.txt) (Zhang et al., 2018). With this data, we performed principle component analysis, which clustered the two cell types correctly according to their corresponding sample label, and plotted log2(FPKM) by color for each sample.

Human kidney embryo 293T (ATCC, Manassas, VA, #CRL-3216, RRID:CVCL_0063) cells were plated at 30% density in a standard 24-well glass bottom plate in complete DMEM media (Thermo Fisher, #119651198) with 10% Fetal Bovine Serum (Thermo Fisher, #10438026), 292 ug/ml L-glutamine, 100 ug/ml Streptomycin Sulfate, and 120Units/ml of Penicillin G Sodium (Thermo Fisher, #10378016). 18 hr later, cells were transiently transfected using a standard calcium chloride transfection protocol. For transfections, 0.1 ug of sequence-verified pCMV-insert-MYC-FLAG overexpression vectors containing either no insert (Origene #PS100001; ‘mock’ transfection) or RFX6 insert (Origene #RC206174) were transfected into each well. 24 hr post-transfection, cells were washed in 1X PBS and fixed in 4% PFA for 10 min at room temperature. 293T cells were tested as negative for mycoplasma contamination.

Fixed 293T cells were blocked for 1 hr at room temperature in 5% BSA in PBST. For primary antibody incubation, cells were incubated with human serum (1:1000) and rabbit anti-FLAG antibody (1:2000) in 5% BSA in PBST for 2 hr at room temperature (RT). Cells were washed 4X in PBST and subsequently incubated with secondary antibodies (goat anti-rabbit IgG 488, Invitrogen, Carlsbad, CA; #A-11034, 1:4000; and goat anti-human 647, Invitrogen #A-21445, 1:4000) for 1 hr at room temperature. Finally, cells were washed 4X in PBST, incubated with DAPI for 5 min at RT, and subsequently placed into PBS for immediate imaging. All images were acquired with a Nikon Ti inverted fluorescence microscope (Nikon Instruments, Melville, NY).

Human fetal small bowels (21.2 days gestational age) were processed as previously described (Berger et al., 2015). Individual APS1 sera (1:4000 dilution) were used in combination with rabbit antibodies to human Chromogranin A (Abcam, Cambridge, MA; #ab15160, 1:5000 dilution). Immunofluorescence detection utilized secondary Alexa Fluor secondary antibodies (Life Technologies, Waltham, MA; 488 goat anti-human IgG, #A11013, RRID:AB_141360; and 546 goat anti-rabbit IgG, #A11010, RRID:AB_2534077). Nuclear DNA was stained with Hoechst dye (Invitrogen, #33342). All images were acquired with a Leica SP5 White Light confocal laser microscope (Leica Microsystems, Buffalo Grove, IL).

DNA plasmids containing full-length cDNA under the control of a T7 promoter for each of the validated antigens (Supplementary file 4) were verified by Sanger sequencing and used as DNA templates in the T7 TNT in vitro transcription/translation kit (Promega, Madison, WI; #L1170) using [35S]-methionine (PerkinElmer, Waltham, MA; #NEG709A). Protein was column-purified on Nap-5 columns (GE healthcare, Chicago, IL; #17-0853-01) and immunoprecipitated on Sephadex protein A/G beads (Sigma Aldrich, St. Louis, MO; #GE17-5280-02 and #GE17-0618-05, 4:1 ratio) in duplicate with serum or control antibodies in 96-well polyvinylidene difluoride filtration plates (Corning, Corning, NY; #EK-680860). Each well contained 35’000 counts per minute (cpm) of radiolabeled protein and 2.5 ul of serum or appropriately diluted control antibody (Supplementary file 4). The cpms of immunoprecipitated protein was quantified using a 96-well Microbeta Trilux liquid scintillation plate reader (Perkin Elmer).