Abstract
Article Figures and data Abstract Editor’s evaluation Introduction Results and discussion Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Phage immunoprecipitation sequencing (PhIP-seq) allows for unbiased, proteome-wide autoantibody discovery across a variety of disease settings, with identification of disease-specific autoantigens providing new insight into previously poorly understood forms of immune dysregulation. Despite several successful implementations of PhIP-seq for autoantigen discovery, including our previous work (Vazquez et al., 2020), current protocols are inherently difficult to scale to accommodate large cohorts of cases and importantly, healthy controls. Here, we develop and validate a high throughput extension of PhIP-seq in various etiologies of autoimmune and inflammatory diseases, including APS1, IPEX, RAG1/2 deficiency, Kawasaki disease (KD), multisystem inflammatory syndrome in children (MIS-C), and finally, mild and severe forms of COVID-19. We demonstrate that these scaled datasets enable machine-learning approaches that result in robust prediction of disease status, as well as the ability to detect both known and novel autoantigens, such as prodynorphin (PDYN) in APS1 patients, and intestinally expressed proteins BEST4 and BTNL8 in IPEX patients. Remarkably, BEST4 antibodies were also found in two patients with RAG1/2 deficiency, one of whom had very early onset IBD. Scaled PhIP-seq examination of both MIS-C and KD demonstrated rare, overlapping antigens, including CGNL1, as well as several strongly enriched putative pneumonia-associated antigens in severe COVID-19, including the endosomal protein EEA1. Together, scaled PhIP-seq provides a valuable tool for broadly assessing both rare and common autoantigen overlap between autoimmune diseases of varying origins and etiologies. Editor’s evaluation This is an important paper that is methodologically compelling. The work presents a series of enhancements to the PhIP-seq method of autoantibody discovery, improving scaling to larger cohorts and control populations, and increasing the ability to discover disease-specific immune responses. The approach is used to discover a novel, frequent autoantibody response (BTNL8) in IPEX patients, and will be an accessible approach to investigate the presence and specificity of autoantibodies in diseases where these have been difficult to define. https://doi.org/10.7554/eLife.78550.sa0 Decision letter Reviews on Sciety eLife’s review process Introduction Autoantibodies provide critical insight into autoimmunity by informing specific protein targets of an aberrant immune response and serving as predictors of disease. Previously, disease-associated autoantigens have been discovered through candidate-based approaches or by screening tissue specific expression libraries. By contrast, the development of proteome-wide screening approaches has enabled the systematic and unbiased discovery of autoantigens. Two complementary approaches for proteome-wide autoantibody discovery include printed protein arrays and phage-immunoprecipitation sequencing (PhIP-seq) (Jeong et al., 2012; Larman et al., 2011; Sharon and Snyder, 2014; Zhu et al., 2001). While powerful, printed protein arrays can be cost- and volume-prohibitive and are not flexible to adapting or generating new antigen libraries. On the other hand, PhIP-seq, originally developed by Larman et al., 2011, uses the economy of scale of arrayed oligonucleotide synthesis to enable very large libraries at comparatively low cost. However, phage-based techniques have remained hindered by labor-intensive protocols that prevent broader accessibility and scaling. Recently, we and others have adapted and applied PhIP-seq to detect novel, disease-associated autoantibodies targeting autoantigens across a variety of autoimmune contexts (Larman et al., 2011; Mandel-Brehm et al., 2019; O’Donovan et al., 2020; Vazquez et al., 2020). However, both technical and biological limitations exist. From a technical standpoint, PhIP-seq libraries express programmed sets of linear peptides, and thus this technique is inherently less sensitive to detect reactivity to conformational antigens, such as Type I interferons (IFNs) and GAD65 (Vazquez et al., 2020; Wang et al., 2021). Nonetheless, the technique allows hundreds of thousands to millions of discrete peptide sequences to be represented in a deterministic manner, including the multiplicity of protein isoforms and variants that are known to exist in vivo. In this sense, PhIP-seq is complementary to mass spectrometry and other techniques that leverage fully native proteins. Given that many forms of autoimmunity exhibit significant phenotypic heterogeneity, the true number of patients with shared disease-associated autoantibodies may be low (Ferre et al., 2016). Therefore, the screening of large cohorts is an essential step for identifying shared antigens and would benefit from a scaled PhIP-seq approach for high throughput testing. Beyond the benefits of detecting low-frequency or low-sensitivity antigens, a scaled approach to PhIP-seq would also facilitate increased size of healthy control cohorts. Recently, PhIP-seq has been deployed to explore emerging forms of autoimmune and inflammatory disease, including COVID-19-associated multisystem inflammatory syndrome in children (MIS-C) (Gruber et al., 2020). However, these studies suffer from a relative paucity of control samples, resulting in low confidence in the disease specificity of the suggested autoantigens. Questions of disease specificity, rare antigens, and antigen overlap can be addressed in larger, scaled experiments. Here, we develop a high-throughput PhIP-seq protocol with markedly increased accessibility (not requiring robotics) and scale (enabling 600–800 samples to be run in parallel), with minimal plate-to-plate variability and low contamination potential, and without sacrificing data quality. We demonstrate the utility of this protocol in the context of an expanded, multi-cohort study, including APS1, patients with immune dysregulation, polyendocrinopathy, X-linked (IPEX), RAG1/2 deficiency with autoimmune features, a KD patient cohort, and emerging COVID-19 patient phenotypes with possible autoimmune underpinnings. In the future, scaled PhIP-seq cohort studies could be used across additional syndromic and sporadic autoimmune diseases to develop an atlas of linear B cell autoantigens. Results and discussion Design and implementation of a scaled PhIP-seq protocol The ability to process large numbers of patient samples for PhIP-seq in a highly uniform manner has several important benefits, including reduction of batch effects between samples from the same cohorts as well as between disease and control cohorts, detection of lower-frequency autoantigens, and the ability to simultaneously include large numbers of control samples. In creating a scaled protocol, we searched for a method that would allow 600–800 samples to be run fully in parallel to reduce any batch or plate-to-plate variability. Thus, each wash or transfer step needed to be performed in rapid succession for all plates. We also prioritized reduction of any well-to-well contamination, particularly given that small, early contamination can amplify across subsequent rounds of immunoprecipitation. Finally, we required our protocol to minimize consumable waste and maximize benchtop accessibility without robotics or other specialized equipment. A benchtop vacuum-based protocol (rapid, consistent wash times) in deep-well 96-well filter plates with single-use foil seals (no well-to-well contamination) met our requirements (see schematic in Figure 1A). The data for APS1 samples on our moderate-throughput manual multichannel protocols were closely correlated with vacuum-based output, including identification of previously validated antigens within each sample (Figure 1B). Additional protocol improvements included: 3-D printing template for vacuum plate adaptors (for easy centrifugation and incubation steps); direct addition of protein A/G beads to Escherichia coli lysis without a preceding elution step; shortened lysis step by using square-well plates with semi-permeable membrane for aeration; and options for smaller volume and decreased reagent library preparation in both 96-well and 384-well formats. A detailed protocol, including custom part designs, for both high-throughput vacuum-based and moderate throughput multichannel-based protocols is available at protocols.io. Figure 1 Download asset Open asset Advantages of and considerations motivating scaled phage-immunoprecipitation sequencing (PhIP-seq). (A) Schematic of vacuum-based scaled PhIP-seq protocol, allowing for parallelized batches of 600–800 samples. (B) Comparison of moderate-throughput multichannel protocol data to high-throughput vacuum-based protocol data, with axes showing normalized read percentages. Controls include a commercial polyclonal anti-GFAP antibody (left), APS1 patient A with known and validated autoantibodies RFX6, SOX10, ACPT, and LCN1 (center), and APS1 patient B with the same known and validated autoantibodies as well as NKX6-3. Large control cohorts are critical for identifying disease-specific autoantibodies Some assays for autoantigen discovery, such as protein arrays, are often used as a quantitative measure of how autoreactive an individual serum sample may be. In contrast, PhIP-seq is an enrichment-based assay in which binders are serially enriched and amplified. A practical limitation of this technique is that non-specific phage may also be amplified, in addition to a wide array of autoreactive, but non-disease-related peptides. We tested whether we could detect global differences between case and control cohorts as a measure of autoreactivity. As each APS1 patient is known to have multiple, high-affinity antibodies to self-proteins (Fishman et al., 2017; Landegren et al., 2016; Meyer et al., 2016; Vazquez et al., 2020) we reasoned that this would be an ideal cohort to determine whether a global autoreactive state was discernible. As expected, each individual sample exhibits a spectrum of enriched genes, regardless of disease status (Figure 2A), indicating that measures of simple enrichment are inadequate for discrimination of cases from controls. Figure 2 Download asset Open asset Application of scaled phage-immunoprecipitation sequencing to expanded APS1 and healthy control cohorts. (A) Number of hits per sample reaching 5, 10, 25, 50, and 100-fold enrichment relative to mock-IP samples. Each dot represents a single APS1 patient (green) or non-APS1 control (gray). (B) When looking for disease-specific hits, increasing the number of healthy controls results in fewer apparent hits and is therefore critical. Shared hits are defined as gene-level signal (>10-fold change over mock-IP) which is shared among 10% of APS1 samples (n=128), present in fewer than 2% of healthy controls, and with at least one APS1 sample with a high signal (FC of 50<). Random downsampling was performed 10 times for each healthy control bin. (C) Nine gene-level hits are present in 10%< of a combined three-group APS1 cohort. North-America-1, n=62; Sweden, n=40; North-America-2, n=26. Anti-GFAP control antibody (n=5) indicates that results are consistent across plates and exhibit no well-to-well contamination. We and others have shown that PhIP-seq can robustly detect disease-associated antigens by comparing antigen-specific signal between disease and control cohorts (Larman et al., 2011; Mandel-Brehm et al., 2019; O’Donovan et al., 2020; Vazquez et al., 2020). In this dataset, encompassing 186 control samples and 128 APS1 samples, we further evaluated the importance of control cohort size. We iteratively downsampled the number of healthy control samples in our dataset to 5, 10, 25, 50, 100, or 150 (out of n=186 total control samples). The number of apparent hits was determined in each condition, where a gene-level hit was called when the following criteria were met: (1) at least 10% of APS1 samples and less than two control samples with a Z-score >10, (2) no control sample exhibiting higher signal than the highest patient signal, and (3) a minimum of one strong positive patient sample (50-fold enrichment over mock-IP). Genes that failed to meet these conditions were considered non-specific. Using these conservative criteria, a control set of 10 samples resulted in (on average) 404 apparent hits, while increasing the control set to 50 samples removed an additional 388 non-specific hits, leaving 16 apparent hits (Figure 2B). Further increasing the number of control samples to 150 samples had diminishing returns, although 4–5 more autoantigen candidates were removed as being non-specific, reducing the frequency of false positives, and ultimately leaving only ~1% of the original candidate list for further investigation. In sum, to improve downstream analysis aimed at detecting disease-associated hits, PhIP-seq experimental design should include a large and appropriate number of non-disease controls. Scaled PhIP-seq replicates and expands autoantigen repertoires across multiple independent cohorts of APS1 We previously identified and validated PhIP-seq hits based on shared positivity of a given hit among 3 (out of 39) or more patients with APS1 (Vazquez et al., 2020). While this enabled us to robustly detect frequently shared antigens within a small disease cohort, antigens with lower frequencies – or with low detection sensitivity – would likely fall below this conservative detection threshold. To improve both sensitivity and specificity, we utilized scaled PhIP-seq to explore expanded cohorts of disease, including a much larger (n=128) APS1 cohort spanning two North American cohorts and one Swedish cohort. All hits present in at least 10% of APS1 patients also spanned all three cohorts, thus further validating broad prevalence of antigens that were previously described by us (RFX6, ACPT, TRIM50, CROCC2, GIP, NKX6-3, KHDC3L) and others (SOX10, LCN1) (Figure 2C; Fishman et al., 2017; Hedstrand et al., 2001; Vazquez et al., 2020). At the gene level, we detected 39 candidate hits that were present in 6/128 (>4%) of APS1 and in 2 or fewer controls (2/186, 1%) (Figure 3A and B). As expected, the larger cohort yielded new candidate antigens that had not been detected previously. For example, PDYN is a secreted opioid peptide thought to be involved in regulation of addiction-related behaviors (reviewed in Fricker et al., 2020). PDYN was enriched by 6/128 (4.7%) of patient samples and subsequently was validated by Radioligand Binding Assay (RLBA) (Figure 3C). Indeed, this validated antigen was present in our previous investigations (Vazquez et al., 2020); however, it was enriched in too few samples to qualify for follow-up. Figure 3 Download asset Open asset Replication and expansion of APS1 autoantigens across multiple cohorts using scaled phage-immunoprecipitation sequencing (PhIP-seq). (A) Increasing the number of healthy controls results in fewer apparent hits and is therefore critical. Shared hits are defined as gene-level signal (>10-fold change over mock-IP) which is shared among 4%< of APS1 samples (n=128), present in fewer than 2% of healthy controls, and with at least one APS1 sample with a high signal (FC of 50<). Random downsampling was performed 10 times for each healthy control bin. (B) 39 candidate hits present in 4%< of the APS1 cohort. (C) Rare, novel anti-PDYN autoantibodies validate at whole-protein level, with PhIP-seq and whole-protein RLBA data showing good concordance. Other notable hits with relevant tissue-restricted expression were also observed. For example, SPAG17 is closely related to the known APS1 autoantigen SPAG16 and is expressed primarily in male germ cells and in the lung, with murine genetic mutations resulting in ciliary dyskinesis with pulmonary phenotypes (Fishman et al., 2017; Teves et al., 2013). Also potentially related to ciliary and/or pulmonary autoimmunity is CROCC2/Rootletin, a protein expressed in ciliated cells, which we previously observed, and now recognize at a high frequency across multiple cohorts (Uhlén et al., 2015; Yang et al., 2002). Similarly, GAS2L2 is a ciliary protein expressed in the airway, with genetic loss of function in mice resulting in impaired mucociliary clearance, and clustered closely with CROCC2 (Bustamante-Marin et al., 2019; Uhlén et al., 2015) in this dataset. These novel putative antigens together hint at potential mucociliary airway autoreactivity. CT45A10 and GPR64 are both proteins with expression restricted primarily to male gonadal tissues (Uhlén et al., 2015). GABRR1 is a GABA receptor expressed primarily in the central nervous system as well as on platelets (Ge et al., 2006; Zhu et al., 2019), and TRIM2 is implicated in genetic disorders of demyelination within the peripheral nervous system, and therefore may be of interest to the chronic inflammatory demyelinating polyneuropathy phenotype that can be seen in some patients with APS1 (Li et al., 2020; Valenzise et al., 2017). In addition to our previously described intestinally expressed autoantigens RFX6, NKX6-3, and GIP, we also identify CDHR5, a transmembrane cadherin-family protein expressed on the enterocyte cell surface, as a putative autoantigen in APS1 (Crawley et al., 2014; Uhlén et al., 2015). APS1 disease prediction by machine learning APS1 is a clinically heterogeneous disease, and it is also heterogeneous with respect to autoantibodies (Ferre et al., 2016; Fishman et al., 2017; Landegren et al., 2016; Meyer et al., 2016; Vazquez et al., 2020). Because PhIP-seq simultaneously interrogates autoreactivity to hundreds of thousands of peptides, we hypothesized that unsupervised machine learning techniques could be used to create a classifier that would distinguish APS1 cases from healthy controls. We applied a simple logistic regression classifier to our full gene-level APS1 (n=128) and control (n=186) datasets, resulting in excellent prediction of disease status (AUC = 0.95, Figure 4A) using fivefold cross-validation. Moreover, we found that the classification model was driven strongly by many of the previously identified autoantigens, including RFX6, KHDC3L, and others (Figure 4A), in addition to some targets that had not been previously examined (Vazquez et al., 2020). These results demonstrate that PhIP-seq autoreactive antigen enrichment profiles are amenable to machine learning techniques, and further suggest that such data could be used to derive diagnostic signatures with strong clinical predictive value. Figure 4 Download asset Open asset Logistic regression of phage-immunoprecipitation sequencing data enables APS1 disease prediction. (A) Receiver operating characteristic (ROC) curve for prediction of APS1 versus control disease status. (B) The highest logistic regression (LR) coefficients include known antigens RFX6, KHDC3L, and others. Antoantibody discovery in IPEX IPEX syndrome is characterized by defective peripheral immune tolerance secondary to impaired T regulatory cell (Treg) function. In IPEX, peripheral tolerance rather than central tolerance is impaired, resulting in a phenotypic constellation of autoimmunity that partially overlaps with APS1 (Bacchetta et al., 2006; Powell et al., 1982). Notably, the majority of IPEX patients exhibit severe enteropathy, with early-onset severe diarrhea and failure to thrive, with many of these children harboring anti-enterocyte antibodies detected by indirect immunofluorescence (Bacchetta et al., 2006; Gambineri et al., 2018; Powell et al., 1982). We hypothesized that the same PhIP-seq approach that was successful for APS1 would also yield informative hits for IPEX. A total of 27 patient samples were analyzed using scaled PhIP-seq, and the data processed in the same manner as for APS1. A handful of IPEX autoantibodies targeting antigens expressed in the intestine have been described, including harmonin (USH1C) and ANKS4B (Eriksson et al., 2019; et al., In our data, enrichment of was in two IPEX patients, and this signal was fully correlated with as previously described (Figure et al., novel putative autoantigens were that were shared among three or more IPEX patients (Figure these were several with expression restricted to the including a protein expressed by a specific of a highly expressed in the and (Figure et al., 2019; et al., 2018; Uhlén et al., 2015). BEST4 and BTNL8 were for by protein immunoprecipitation. A total of of IPEX patients were positive for autoantibodies (Figure In the case of identified of IPEX patients were positive for antibodies (Figure all patients with and/or BEST4 antibodies also had clinical of In to the of our we tested a independent cohort of IPEX patients for antibodies and found patients to be positive (Figure these results suggest that and are with IPEX Figure with 1 all Download asset Open asset sequencing (PhIP-seq) screening in IPEX and RAG1/2 deficiency novel, intestinally expressed autoantigens BEST4 and (A) PhIP-seq of frequent shared antigens among IPEX, with indicating relative to a cohort of controls. (B) Radioligand assay for BTNL8 additional positive IPEX patients Radioligand assay for BEST4 autoantibodies well with PhIP-seq data (C) of additional positive in an independent IPEX cohort by indicates of healthy controls PhIP-seq of patients with mutations in RAG1/2 two patients with assay of antibodies in both PhIP-seq positive patients. of autoantigen BEST4 in the of RAG1/2 mutations RAG1/2 mutations an additional and heterogeneous of immune dysregulation. to of peripheral T and B cells, therefore severe combined However, patients with RAG1/2 have to T and B on the of the these patients can present with or combined immune deficiency with and autoimmunity et al., 2018; et al., 2020). are particularly common in patients with and are the frequent of autoimmunity in patients with deficiency, but and have been also While antibodies have been described, other disease-associated autoantibody targets to be identified et al., 2020). patients with RAG1/2 mutations were by PhIP-seq to for overlap with APS1 and IPEX antigens, as well as for novel autoantigen overlap was between RAG1/2 deficiency and APS1 and IPEX. However, two samples from RAG1/2 patients the presence of antibodies (Figure which were through using protein (Figure positive patients had indicating the presence of autoimmune – given that in the of RAG1/2 deficiency is rare – one of the two harboring antibodies also had very early-onset inflammatory disease. other putative antigens among the larger RAG1/2 deficiency cohort were of the cohort were enriched for from (Figure 1B). is a protein with multiple autoimmune diseases, as by studies that variants in are with disease and autoimmune et al., 2011; et al., 2020). patients also had of autoantibodies targeting a receptor and a receptor that and is important for of et al., 2019; et al., 2017). Autoantibodies targeting these antigens could potentially a in the seen in cases of RAG1/2 and/or increased to and will additional follow-up. PhIP-seq rare, shared candidate autoantigens in MIS-C MIS-C to critical in of children and some common clinical with the common of disease in the Despite for a of and autoantibodies in the of KD and the etiologies of both diseases et al., 2020; et al., 2016). Recently, PhIP-seq has been deployed to explore COVID-19-associated MIS-C (Gruber et al., 2020). However, this only healthy controls and MIS-C patients, and as our results have of PhIP-seq hits the of numbers of controls (Figure these previously hits Therefore, we to an MIS-C cohort in of these as well as to explore for possible autoantibody overlap between KD and MIS-C controls, and COVID-19 controls were examined by PhIP-seq, each of which was to a cohort of healthy controls for specific enrichment was for any of the previously candidate antigens that with our PhIP-seq library (Figure and sample differences in for differences in our however, these results suggest that PhIP-seq hits should be to and/or MIS-C patients are with which has be in the of sample et al., et al., 2021). While it whether novel autoantibodies are likely to be present in the majority of our MIS-C samples are to be several of our samples are of status, each of the candidate autoantibodies below is present in at least one sample known to be Figure Download asset Open asset sequencing (PhIP-seq) screening of multisystem inflammatory syndrome in children (MIS-C) and Kawasaki disease cohorts. (A) of signal for putative hits from et al., among COVID-19 controls, and controls (B) rare, shared PhIP-seq were found among MIS-C patients. (C) of putative antigens in a cohort of KD patients. that are specific to KD and are not found among controls, are in A small number of rare putative antigens are shared between KD and MIS-C (left), with assay of antibody reactivity to protein of in three KD patients and one MIS-C patient of our MIS-C cohort for shared candidate hits yielded only three candidate hits, each in patient These were and (Figure While these targets may be of These results suggest that a much larger MIS-C cohort, by an large set of healthy controls, will be required to detect rare, shared antigens with confidence by PhIP-seq of a cohort of KD patients To for possible we analyzed a large cohort of KD by KD patients are also often with was to that each of these samples was to Using the same hit criteria as we detected shared hits among 3 or more of the KD samples, which were specific to KD relative to healthy controls (Figure these shared KD hits, were from additional control including the patients. Each of these hits was present in only a small of KD samples, significant among samples. Some of the candidate antigens have possible to the seen in including and is a protein known to the receptor with murine of resulting in increased of and and