Systematic and proactive evaluation of AIRE missense variant effects

Abstract

Pathogenic variants in the Autoimmune Regulator (AIRE) cause Autoimmune Polyendocrine Syndrome Type 1 (APS-1), a rare primary immunodeficiency disease with symptoms including hypoparathyroidism, adrenal insufficiency, and chronic mucocutaneous candidiasis. AIRE increases the expression and presentation of tissue-specific genes expressing ‘self’ antigens in the developing T cell niche, thus triggering the elimination of self-reactive T cells and preventing autoimmunity. Earlier diagnoses can benefit those with APS-1, and APS-1 diagnosis by AIRE sequencing is increasingly common. However, two thirds of reported clinical variants are missense, and more than half of these are “variants of uncertain significance” (VUS). Cell-based variant functional assays can provide strong evidence towards more informative variant classification, but these are carried out reactively, often years after clinical presentation. By contrast, proactively assessing all possible missense variants could provide immediate evidence to guide genetic diagnosis, even for never-before-seen variants. Here we used an insulin promoter-driven reporter to proactively assess the function of 9,790 AIRE missense variants. The resulting AIRE variant effect map both validates and extends current biochemical knowledge, concords with pathogenicity annotations, and provides proactive evidence for 70% of previously-reported VUS. Placing our map in the context of both an international APS-1 cohort and the UK BioBank revealed quantitative genotype-phenotype correlations. Moreover, evidence from our variant effect map resolved 32% of current VUS. Together, our proactive resource of AIRE variant impacts offers the potential to improve outcomes via more rapid and definitive APS-1 diagnosis.

Introduction

Autoimmune Polyendocrine Syndrome Type 1 (APS-1 [MIM:240300]), also called Autoimmune Polyendocrinopathy with Candidiasis and Ectodermal Dystrophy (APECED), is a rare, monogenic, predominantly autosomal-recessive primary immunodeficiency disease caused by loss-of-function (LoF) variants in the autoimmune regulator (AIRE [MIM: 607358])¹. AIRE prevents autoimmunity by increasing the expression of tissue-specific genes in the niche of developing T cells, allowing for the elimination of emerging self-reactive T-cells before they can mature². Dominant negative AIRE variants have been linked to milder non-classical phenotypes with variable clinical presentation^3,4. APS-1 has a prevalence less than 1:100,000 in most populations, although higher prevalence in certain populations such as Finns (1:25,000⁵) and Sardinians (1:60,000⁶). Usually, APS-1 diagnosis requires the presence of two symptoms amongst hypoparathyroidism, adrenal insufficiency and chronic mucocutaneous candidiasis but APS-1 can cause a broad range of other autoimmune symptoms, some of which may be life threatening⁷. Those with APS-1 can have normal life expectancies with proper management of symptoms, but the prevalence of disease components increases over the years making treatment and follow-up problematic. Hence, earlier diagnosis improves outcomes by mitigating the effects of long-term autoimmunity⁸. Sequencing is increasingly used to diagnose APS-1 and proposes to decrease diagnostic delay^1,8, but this in turn requires rapid and accurate interpretation of AIRE variants.

Of the AIRE variants reported to ClinVar⁹, a database of genetic variants and their impact on human health, two-thirds are missense variants, which are usually more challenging than other variants (for example, nonsense, frameshift, or large deletion variants) to interpret. Indeed, 345 (81%) of 426 AIRE missense variants in ClinVar are annotated as a “variant of uncertain significance” (VUS). One of the strongest forms of evidence for variant annotation under current guidelines, and one which is often unavailable, comes from functional assays of variant impact¹⁰. Currently, one-at-a-time ‘reactive’ variant functional assays are the standard, with results often lagging months or years after a rare variant has first been detected clinically¹¹. A more recent ‘proactive’ approach, made possible by advances in cellular engineering and sequencing, is to obtain functional evidence for nearly all possible missense variants (variant effect maps) that even include variants not yet seen in the clinic^11,12.

Transcriptional reporter assays have been described for AIRE, but none were immediately suited to evaluate AIRE variant functions at scale. For example, one previously reported assay used the insulin promoter sequence upstream of luciferase (which is not amenable to fluorescence-activated cell sorting; FACS) as a reporter^13–18, while another evaluated endogenous insulin upregulation by AIRE^16,19. Here we describe a variant effect map for AIRE, developed using an insulin-driven GFP reporter in HEK293 cells. The map recapitulates all but one region known to be important for function, provides evidence to guide clinical variant interpretation for APS-1, supports a genotype-phenotype correlation in an international clinical cohort, associates apparently-damaging variants with hypoparathyroidism and vitamin B12 deficiency anemia in the UK BioBank, and allows for confident classification of 32% (109/345) of VUS. The map reveals that many AIRE variants exhibit gain-of-function in terms of increased expression of the insulin-driven GFP reporter. The map also opens a new line of inquiry by identifying AIRE residues that are highly intolerant to substitution but for which the function remains unknown.

Material and Methods

Plasmids

The Tet promoter of the dAAVS1-TetBxb1-BFP plasmid²⁰ was replaced with the EF1α promoter from PL-sin-EF1α-eGFP (Addgene #21320) using Gibson Assembly²¹ to create the dAAVS1-EF1α-BFP plasmid. A Gateway-compatible attB integration construct (pDEST-HC-REC-v2) was made by replacing the Kozak sequence and PTEN open reading frame of attB-PTEN-IRES-mCherry with the 1873bp Gateway attR1-CcdB-attR2 cassette from pDEST-AD-CYH2.

A codon-optimized version of the AIRE cDNA sequence was designed (see Table S1 for sequence; Twist Biosciences) because of the high GC-content within the endogenous AIRE sequence (NCBI Reference Sequence: NM_000383.4; NP_000374.1). To clone the AIRE DNA fragment into a plasmid backbone, the fragment was PCR-amplified with primers ATTB_U2site_kozak_F and ATTB_D1site_kozak_R (see Table S2 for Primers), with additional “universal addition” site needed for later variant library preparation, a Kozak site, and attB1 and attB2 sites on either end, compatible with integration into entry vector pDONR223. The PCR product was gel extracted, BP-cloned with Gateway^™ BP Clonase^™ II Enzyme mix (Thermo Fisher Scientific) into pDONR223 and transformed into NEB^® 5-alpha Competent E. coli (High Efficiency) and selected on Spectinomycin LB plates. DNA was extracted with the QIAprep Spin Miniprep Kit and Sanger sequencing confirmed the plasmid. The AIRE WT plasmid compatible with integration into the Bxb1 site was prepared by cloning AIRE WT in pDONR223 using Gateway^™ LR Clonase^™ into the Bxb1-compatible pDEST-HC-REC-v2, the plasmid enabling integration into the Bxb1 site of reporter cell lines (which also contains an downstream internal ribosomal entry site (IRES) driving expression of the mCherry integration marker). The product was transformed into 5-alpha E.coli, selected on LB plates with 1X Ampicillin, DNA extracted and sequence confirmed with Sanger.

To prepare for integrating the insulin promoter-driven GFP reporter construct into the Flp-In site of HEK293 cells, we adapted the pcDNA5-FRT-eGFP plasmid²² (kind gift from Anne-Claude Gingras) by replacing the existing CMV enhancer and promoter with the insulin core promoter. This was accomplished using PCR with phosphorylated primers delCMV_pcDNA5_FRT_eGFP_F and delCMV_pcDNA5_FRT_eGFP_R (see Table S2 for Primers) to linearize the plasmid. Next, the product was run on 1% agarose gel and band extracted with QIAQuick Gel Extraction Kit. The resulting product was ligated and transformed into NEB^® 5-alpha Competent (High Efficiency) E.coli, colonies picked and grown overnight at 37°C in LB with 1X Ampicillin selection, DNA extracted with QIAprep Spin Miniprep Kit, and Sanger sequenced to confirm. Next, to insert the insulin core promoter sequence, flanking Gateway-compatible attB sites were added to the core insulin sequence of −384 bp to 24 bp synthesized by Twist Bioscience (sequence available in Table S1) in a PCR reaction using primers INSpromATTB_F and INSpromATTB_R (see Table S2 for Primers). The resulting product was Gateway cloned into pDONR223 using Gateway^™ BP Clonase^™ II Enzyme mix, transformed, miniprepped, and subsequently LR gateway cloned using Gateway^™ LR Clonase^™ II Enzyme mix into ΔCMV-pcDNA5-FRT-eGFP to generate the final reporter construct plasmid InsProm-pcDNA5-FRT-eGFP. This final plasmid was transformed, DNA extracted and sequence confirmed with Sanger sequencing.

Reporter cell line generation

To introduce the Bxb1 landing pad for later introduction of AIRE variants, Flp-In^™ T-REx^™ HEK293 cells were transfected with equal proportions of dAAVS1-EF1α-BFP, AAVS1-TALEN-L and AAVS1-TALEN-R (Addgene #59025, #59026) using JetPRIME reagent (Polyplus Sartorius) as previously described^20,23. Transfected cells were cultured for ~2 weeks and enriched for BFP-positive cells via FACS, and a single-cell-derived clone was validated as previously described^20,23.

To insert the insulin reporter construct into the above Flp-In^™ T-REx^™ HEK293 EF1α-Bxb1 cell line, 300K cells were seeded and co-transfected with 2 ug pOG44 (Flp-recombinase expression vector) and 200 ng InsProm-pcDNA5-FRT-eGFP using 4 uL JetPrime reagent and 200 uL of JetPrime buffer. To select for successful integrants, 200 ug/mL of Hygromycin B was added 48 hours after transfection. Surviving cells were seeded at 0.5 cells/well and clonally expanded. One clone was randomly selected as the insulin-driven GFP reporter cell line.

Cells were cultured in Serum (FBS), 90% Dulbecco’s Modified Eagle Medium (DMEM), and 1X Penicillin-Streptomycin (PS). Cells were tested for mycoplasma and grown at 37°C at 5% CO2.

Small-scale reporter assay validation

As an initial test of the ability of the insulin-driven GFP reporter cell line to separate AIRE variants by function, we selected a small set of clinically-reported AIRE variants from across the length of the gene: p.Trp78Arg (pathogenic), p.Arg92Trp (pathogenic), p.Ser278Arg (benign), p.Pro400Leu (previously classified as benign, now with conflicting classifications), p.Ala403Thr (reported as both benign and VUS), and p.Asp503Asn (benign). Site-directed mutagenesis PCR was performed using the AIRE open reading frame in pDONR223 as a template to introduce for each variant (See Table S3 for list of site-directed mutagenesis primers, designed according to the codon-optimized AIRE sequence). Products were ligated, transformed into E.coli, and LR cloned into pDEST-HC-REC-v2 as described for the AIRE WT. Variants were confirmed using whole plasmid sequencing. To evaluate variants in the insulin-driven GFP reporter cell line, the reporter cell line was plated at 40K cells/well in DMEM + 10% FBS (antibiotic-free media) and variants transfected in triplicate using FuGENE^® 6 at a 10:1 ratio by mass of variant DNA:Bxb1 recombinase 30 minutes after plating the cells. Media was replaced with DMEM + 10% FBS + 1X PS 24 hours after transfection. Cells were expanded, gated for live, single cells which express the mCherry integration marker, and next evaluated for GFP expression one week after the transfection using a Beckman-Coulter flow cytometer to generate GFP distributions for the control variants.

Library generation

We developed a modified version of the POPCode method¹² called Dual-Tag Precision Oligo-Pool based Code Alteration (DT-POPCode). Relative to POPCode, DT-POPCode eliminates the need for a uracilated template and decreases variant bottlenecking during downstream plasmid library cloning (see Figure S1 and Supplemental Methods). First, mutagenic oligos were designed such that a “NNN” degenerate codon was centered on each AIRE target codon (see Table S1 for the codon-optimized AIRE sequence. Primers ATTB_U2site_kozak_F and ATTB_D1site_kozak_R were used to amplify synthesized AIRE to ensure compatibility with downstream steps, see Table S2 for Primers) in turn, based on previous work²⁴. Primers were also designed that ‘tiled’ across the AIRE coding region, splitting it into fifteen ~150-200 base pair tiles with inclusion of overhang indexing adaptors to allow for short-read Illumina sequencing. The adaptors used for overhangs were the standard Illumina-compatible TACACGACGCTCTTCCGATCT overhang for the F primer and AGACGTGTGCTCTTCCGATCT for the R primer. We also inserted 1-3 “N” nucleotides between the adaptors and the coding sequence for each tiling primer, to increase sequencing library complexity by offsetting the nucleotide sequenced for each read. See Table S3 for a list of all mutagenic and tiling primers used. Generation of the mutagenic library via DT-POPCode was accomplished by hybridizing mutagenic oligos to each one of three WT AIRE mutagenic regions (Region 1: amino acids 1-181; Region 2: amino acids 182-362; Region 3: amino acids 363-545) while simultaneously hybridizing two oligos (the “dual tags” that give DT-POPCode its name) that allow primer extension from the 5’ end and (after ligation) addition of a 3’ end that enables us to later specifically amplify the mutagenized product using PCR. We performed quality control of the linear amplicon library using Illumina sequencing of tiles, and counted variants with the TileSeq_MutCount pipeline (https://github.com/rothlab/tileseq_mutcount). A suitable coverage of >85% of substitutions present at a frequency above 10 counts per million reads and an acceptable average number of amino acid substitutions per clone (λ) between 0.4 and 1, was achieved for each mutagenic region. Next, for each AIRE mutagenic region, we first carried out large-scale BP cloning of the linear amplicon library into pDONR223 and transformed the product using NEB^® 5-alpha Competent (High Efficiency) E.coli and selected using LB plates with 1X Spectinomycin. DNA was extracted using a Midiprep, and afterwards, the product was large-scale LR Gateway cloned into pDEST-HC-REC-v2. Large-scale LR product transformations were performed using NEB^® 10-beta electrocompetent E. coli (High Efficiency) and clones selected with 1X Ampicillin. To prevent variant coverage bottlenecks during cloning steps, we estimated the number of E.coli colonies expected to achieve 50 independent clones per codon substitutions as: 63 (number of possible codon substitutions based on NNN degenerate code) × 182 (number of amino acids per mutagenic region) × 50 (independent clones), divided by λ. The final mutagenic libraries for each region of AIRE in pDEST-HC-REC-v2 were prepared using an endotoxin-free DNA extraction kit, and quality checked similarly to the linear amplicon library above.

Large-scale reporter assay

The final LR library plasmid pool for each AIRE region was transfected into the reporter cell line using a ratio of library:Bxb1 recombinase of 10:1 by mass using FuGENE^® 6 (using Opti-MEM^® reduced-serum medium for the transfection mix), 30 minutes after plating 15 million insulin promoter-GFP reporter cells in 10% FBS and DMEM without the presence of PS. Two biological transfection replicates were performed per AIRE region. The transfection was stopped after 24 hours by replacing media with 10% FBS, DMEM and 1X PS. Cells were passaged every 2-3 days at 80% confluence. At day 7 post-transfection, cells were enriched for those with integrated AIRE by gating on single, viable cells expressing the integration marker mCherry, and lacking the expression of BFP from the landing pad, using the SONY MA900 cell sorter. 2 million cells with integration events were sorted per regional biological replicate. Integrated, sorted cells were expanded for 7 days and the top 10% of GFP-expressing cells were sorted to enrich for functional AIRE variants. A control population of ‘non-select’ cells that were not enriched based on reporter activity were obtained separately by selecting for single viable cells that contained the integration mCherry marker (and lack of BFP) on the same day as the selection sort. Cells were expanded, pellets of 15M cells were collected for select and non-select conditions, and DNA extracted. To sequence either select or non-select cell populations, the Bxb1 site was first amplified from the cells with primers Ef1a-Bxb1Amp_F1 and BxbAmp_R1 (see Table S2 for Primers) using 1000 ng of DNA used per one of 15 PCR reactions, and a total number of 25 PCR cycles. Final PRCs were pooled, gel extracted, and tiling of large-scale screens was performed using 10ng of Bxb1 PCR product for each tiling reaction to prevent bottlenecking of variants. Each condition had three tiling PCR technical replicates. The WT reference sequence used as a control for sequencing error was the tiled AIRE WT plasmid backbone sequence. Illumina sequencing for the large-scale assay was performed to a minimal depth of 1M reads/tile.

Functional score calculation

Sequencing reads were processed as previously described^12,25,26, by counting variant instances and translation to amino acids, using the TileSeq_MutCount pipeline (https://github.com/rothlab/tileseq_mutcount). Counts for each amino acid change were next processed using the TileSeq_Mave pipeline (https://github.com/rothlab/tileseqMave), which calculates the enrichment or depletion of variants between the non-select and select conditions. Variants were included if they reached a minimum threshold of 10 counts per million sequencing reads in the non-select condition. WT control variant instances were subtracted from the enrichment rations. Scores for each codon-level amino acid change were averaged, and standard error calculated based on the number of measurements. This analysis was performed separately for each biological replicate, and each replicate was separately re-scaled based on the calibration of nonsense and synonymous variants set to medians of 0 and 1, respectively. Biological replicate scores were averaged, and standard error propagated using the delta method. For downstream estimation of log likelihood ratios of pathogenicity (LLR_p) and evidence levels (described below), only variants that were present in both biological replicates were included. Otherwise, for calculation of mean scores by position and all other analysis, variants were included if only detected in one of the biological replicates and existing score and standard error measurement used. Functional scores are available in Table S4, noting the presence of the variant in one or both replicates, and whether the substitution measured is single nucleotide variant (SNV)-accessible, based on the original nucleotide sequence of AIRE (non-codon optimized). Mean scores by position are listed in Table S5.

Structural analyses

The protein structure for AIRE was predicted via AlphaFold 3²⁷ using the AIRE amino acid sequence and the addition of 4 Zn²⁺ ligands. Structural alpha helix and beta strand positions were identified via the Dictionary of Secondary Structure in Proteins (DSSP)²⁸ using the AlphaFold 3 predicted monomer structure²⁷. Structures were colored according to median AIRE positional functional scores using the tileseqMave (https://github.com/rothlab/tileseqMave) colorize script¹². The PDB structure 2KE1 was used to show the PHD1 domain in complex with H3K4me0²⁹. Accessible surface area (ASA) measurements for AIRE residues were obtained from FreeSASA³⁰, based on the monomeric AlphaFold 3 model. Buried residues were defined by having accessible surface area (ASA)<50 Ų, while surface residues were defined by ASA>110 Ų (Positions summarized in Table S6). We predicted interacting residues with PRODIGY^31,32using the AlphaFold 3 predictions for monomeric, dimeric and trimeric AIRE (positions provided in Table S7). Interacting residues were defined as those within 5.0 Å of each other.

Variant reference sets and benchmarking

ClinVar reference set variants and gnomAD v4.1.0 allele frequency data were accessed on Aug 25th, 2025. ClinVar reference sets were filtered on one star clinical classifications. To evaluate the performance of the map on ClinVar reference set variants compared to computational variant predictors, we assessed scores using balanced precision-recall analysis as described previously³³. Recall was calculated by evaluating the proportion of pathogenic variants correctly identified as damaging. Precision assessed the proportion of variants known to be pathogenic below a given threshold. “Balanced” refers to the use of a Bayesian method for adjusting precision to reflect what would be expected given a balanced (50%/50%) reference set of negative (benign or likely benign) and positive (pathogenic or likely pathogenic) variants³⁴. To evaluate performance, we used a reference set of variants that were scored in the map and by all of the assessed computational predictors and which received at least one ClinVar review star (16 negative and 30 positive variants including the PHD1 domain, and 15 negative and 29 positive variants excluding the PHD1 domain, summarized in Table S8). Each reference variant contributed equally to performance measures regardless of amino acid position.

Log likelihood ratios of pathogenicity calculations

To determine the log likelihood ratio of pathogenicity (LLR_p) for each variant, we used the reference set of variants outside of the PHD1 domain (described above and summarized in Table S8). The score distributions for both negative and positive variant sets were estimated using probability kernel density functions based on the Epanechnikov kernel³⁵, with bandwidths determined using Scott’s rule of thumb³⁶. The LLR_p values corresponding to each variant’s score were each calculated as ratio of the probability densities at that score from the positive and negative distributions, respectively. Next, we used a previous calibration approach^25,37,38 to convert LLR_p scores to evidence strength levels used in current variant classification guidelines¹⁰. Variants with scores falling outside the observed range of scores for reference variants (i.e, those with scores < −0.73 or > 1.70) were assigned undetermined evidence strengths. See Table S4 for the estimated evidence strength levels for each measured substitution. We did not assert evidence towards benignity for variants located in the PHD1 domain, due to our assay’s blind spot in this domain. However, evidence towards pathogenicity was retained for some variants in the PHD1 domain, as these may have a broader structural impact on the protein.

Cohort analyses

We assembled an APS-1 cohort from literature reports of well-characterized APS-1 cases involving AIRE missense variants on at least one allele (Table S9). Altogether, 98 individuals were identified, representing 20 different nationalities across four continents^4,39–66. Twenty-one individuals were homozygous for AIRE missense variants while 45 were compound heterozygous for a missense variant and a truncating variant (i.e nonsense, frameshift, splice site). For 32 individuals, monoallelic missense variants had been reported suggesting autosomal dominant inheritance. A total of 50 unique missense variants were collected. Among these, 16 had not been reported to ClinVar at the time of writing, and these were formally classified in light of the AIRE map. All variants were classified according to the American College of Medical Genetics and Genomics/Association for Molecular Pathology (ACMG/AMP) sequence variant guidance¹⁰, supplemented with recommendations from ClinGen (https://www.clinicalgenome.org/working-groups/sequence-variant-interpretation) and the UK Association for Clinical Genomic Science (ACGS) (https://www.acgs.uk.com/media/12533/uk-practice-guidelines-for-variant-classification-v12-2024.pdf). The evidence from functional data was assigned strength according to recommendations by Brnich et al⁶⁷, but was capped at a maximum of strong in both pathogenic and benign directions.

UK Biobank analysis was in accordance with the associated data use agreements, with consent to participate having been obtained via the UK BioBank. Our use of de-identified data was approved and overseen by The UK BioBank Ethics Advisory Committee (Project ID: 51135). ICD10 codes used for UK BioBank analyses are as follows: insulin-dependent diabetes mellitus E10, adrenal insufficiency E27.1, hypoparathyroidism E20, hypoparathyroidism and other E21, candidal stomatosis B37.0, chronic atrophic gastritis K29.4, vitamin B12 deficiency anemia D51, coeliac disease K90.0, keratitis H16, gastro-enteritis and colitis K52.9, interstitial pulmonary diseases J84, seropositive rheumatoid arthritis M05, systematic sclerosis M34, thyroiditis E06, vitiligo L80, and chronic hepatitis K73. We defined rare variation as a minor allele frequency of 1×10⁻⁴. We defined “damaging” variants for each of the two computational predictors (REVEL⁶⁸ and AlphaMissense⁶⁹) using thresholds provided by the authors (REVEL: scores > 0.5, AlphaMissense: “pathogenic” annotations).

Results

The insulin promoter-GFP reporter recapitulates known AIRE variant effects

Here we undertook missense variant effect mapping for AIRE (outlined in Figure 1A). Because insulin is one of the thousands of genes upregulated by AIRE, because dysregulation of insulin is known to lead to either insulin tolerance or autoimmunity¹⁵, and because insulin autoantibodies are implicated in the development of T1D—a symptom of APS-1^70–72—expression from the insulin promoter was chosen as the basis for our assay to evaluate AIRE function.

Figure 1: Generation of variant effect map for AIRE.

A) Overview of variant effect mapping process for AIRE.

B) AIRE variant effect map, with scores generated via an integrated insulin promoter-driven GFP reporter in HEK293 cells, and with AIRE variant constructs stably integrated and expressed at the Bxb1 landing pad. WT residues at each position are in yellow, functionally deleterious variants are in blue, synonymous-like variants are in white and variants exhibiting increased reporter expression (hyper-activation) are in red. Standard errors (stderr) for each variant are represented by diagonal bars centered on each missense variant’s square. A consensus track above each position summarizes scores at that position. Domains are indicated in teal, regions of disorder in grey, and secondary structure is indicated via helices for alpha helices and arrows for beta strands as determined through the monomeric AlphaFold 3 structural prediction for AIRE (see Material and Methods).

Although AIRE’s role in removing or redirecting effector T cells that recognize self-antigens is thought to primarily be in mTECs, HEK293 cells have become accepted as a reliable model for the study of AIRE function, and indeed many of AIRE’s interactions in mTECs are recapitulated in HEK293 cells^19,73–78. Therefore, we established the reporter system in a HEK293 cell line. More specifically, we used a derivative of the HEK293 cell line with two ‘landing pads’ that are useful for this study: The first contains a Bxb1 recombination site that allows for efficient large-scale genomic integration of a target construct^12,20,79,80, enabling the generation of heterogeneous cell pools in which each cell expresses a single specific variant of the target gene. The second landing pad is a Flp-In site enabling integration of our reporter construct consisting of an insulin core promoter (previously shown to be sufficient for insulin expression⁸¹) fused to a GFP fluorophore.

To evaluate the reporter system, site-directed mutagenesis was used to generate two test sets of AIRE variants, with the first containing variants with pathogenic or likely pathogenic (P/LP) annotations from ClinVar and the second containing variants with benign or likely benign (B/LB) annotations. These variants were stably and clonally integrated into the Bxb1 site of the insulin-promoter GFP reporter cell line. Variant function was evaluated using flow cytometry: functional AIRE was expected to upregulate the insulin promoter-GFP reporter more than functionally abnormal variants of AIRE. In cells bearing P/LP variants, there was a clear downward shift in GFP levels (Figure S2). Although the GFP distributions of cell populations carrying P/LP and B/LB variants show some overlap, Figure S2 shows that use of a specific GFP gating threshold yields clear differences between P/LP and B/LB variants in terms of the fraction of cells exhibiting expression above that threshold. Therefore, the reporter system demonstrated concordance with expected variant effects and was amenable to high-throughput fluorescence-activated cell sorting for evaluation of AIRE variants at large scale.

Large-scale library generation for AIRE

To generate a large-scale clone library, ideally covering all possible single-amino acid substitutions in AIRE, we developed Dual-Tag Precision Oligo-Pool based Code Alteration (DT-POPCode) which eliminates a requirement of the previous POPCode approach¹² for a uracilated template (see Methods). DT-POPCode was used to generate three amplicon libraries of the full-length AIRE cDNA, each focused on a different region of the protein (Region 1: amino acids 1-181; Region 2: amino acids 182-362; Region 3: amino acids 363-545). Regional mutagenesis enables each variant to have a frequency within its library that is well above the background rate at which a variant appears simply due to base-calling error. For each library, all possible missense variants of AIRE within the corresponding region were generated using NNN degenerate oligos centered on each codon. Variant library coverage was confirmed with Illumina sequencing after en masse Gateway cloning into the integration vector, generating pools that contained 1.2, 0.96, and 0.62 mutations per clone for Regions 1-3 respectively (Figure S3A). Experience has shown that, for variants appearing less than 10 times per million, it is challenging to evaluate relative variant representation before and after a selection procedure. Even considering the variant above this threshold, our libraries collectively contained nearly all amino acid changes (98, 97 and 98% for regions 1-3; Figure S3B).

Systematic functional evaluation of AIRE missense variants

We next conducted large-scale transfections to introduce the corresponding integration vector library (with mutagenesis focused in that region) into cells bearing the insulin promoter-GFP reporter. These were carried out separately focusing mutagenesis in each of the three regions, with two biological transfection replicates for each of the regions. To select for cells with an integrated AIRE construct, cells were first sorted to obtain viable single cells exhibiting mCherry but not BFP expression (Figure S4). To enrich for cells bearing functioning AIRE variants, these cells were first expanded and then sorted again to retain the top 10% of GFP-expressing cells (‘select’ cells). We also obtained a control population of ‘non-select’ cells sorted to ensure at least some detectable GFP expression (indicating presence of a functioning reporter construct). After expansion, genomic DNA was extracted from both select and non-select cell pools, and PCR was used to amplify the integrated AIRE cDNA locus from each pool. To evaluate the frequency of each variant within each cell pool, we used a set of 200bp tiles that collectively covered the entire AIRE coding region in each library. These tiles were amplified using primers that introduced flanking sequences which enabled duplex Illumina sequencing. To derive a functional score for each AIRE variant, we used the TileSeq pipeline¹². Briefly, this involved first calculating the log-ratio of each variant’s frequency in select relative to non-select cells. Log ratio scores were rescaled to functional scores (see Material and Methods for details) such that the median functional score of nonsense variants was 0, and the median functional score of synonymous variants was 1. Further steps (see Material and Methods) included estimating experimental error based on trends in replicate agreement as a function of variant abundance in the non-select library, filtering to keep scores only for those variants that were well-represented within the non-select library and–for each variant–taking the mean of biological replicate scores and propagating estimated experimental error to each final score (see Material and Methods). Thus, we generated an AIRE missense variant effect map (Figure 1B) with scores for 9790 (90%) of the 10900 possible amino acid substitutions in AIRE, including 3127 (98%) of the 3189 substitutions that are ‘SNV-accessible’ (in the sense that the new amino acid can be encoded by changing a single nucleotide in the reference codon) and therefore more likely than other amino acid changes to be observed clinically.

AIRE variant effect map quality evaluation

One initial measure of multiplexed assay quality is correlation between biological replicates for missense variants, which was moderate but significant (Pearson’s R=0.49, p<2×10⁻¹⁶; Figure S5A). Another measure of map quality is the extent to which scores for synonymous variants can be distinguished from those of nonsense variants: the score distributions of nonsense and synonymous variants were well separated (p<2×10⁻¹⁶ by Wilcoxon test; Figure S5B). Proteins can often tolerate nonsense variants at the C-terminus and indeed the last 8 amino acids appeared not to be essential for AIRE to activate the insulin promoter in our assay. Not surprisingly, separation between nonsense and synonymous variants remained significant after removal of these C-terminal nonsense variants (p<2×10⁻¹⁶ by Wilcoxon test; Figure S5C). Another test was whether scores followed the expectation that alleles that are more common in a human population would be less damaging. Indeed, the median scores of singletons (defined here by minor allele frequency (MAF) of < 1×10⁻⁶ in gnomAD v4.1.0) were more intolerant to variation than others (median score of singletons was 0.88, median scores of above-singleton variants was 0.95, p=4×10⁻³ by Wilcoxon test).

The quality of a variant effect map can additionally be assessed by how well it corresponds to general biochemical trends. We demonstrated that missense variant changes that were SNV-accessible are less damaging than multi-nucleotide substitutions (p=2×10⁻² by Wilcoxon test, Figure S5D) as expected⁸². Usually, buried amino acid residues are less tolerant of variation than surface residues, especially when buried hydrophobic positions are disrupted^83,84. To compare the scores between surface and buried residues of AIRE, solvent accessible surface areas were predicted (using FreeSASA applied to the monomeric AlphaFold 3 AIRE structure^27,30,85), and annotations of residues as either buried or surface (Figure S6; ASA<50Ų, ASA>110Ų, respectively) were compared to map scores. Although the predicted AIRE AlphaFold structure is of low confidence, we observed significantly more deleterious scores for predicted-buried residues than predicted-surface residues (Figure 2A, median map score at buried positions=0.43, median map score at the surface=0.89, p=3×10⁻¹⁴ (Wilcoxon), delta medians=0.46).

Figure 2: AIRE variant effect map scores in structural context.

A) Distributions of AIRE functional scores corresponding to buried and surface positions. Buried and surface predictions were determined by FreeSASA on the monomeric AlphaFold 3²⁷ structure prediction of AIRE (accessible surface area (ASA)<50Ų for buried, ASA>110Ų for surface). Boxes represent interquartile range, and bold bars represent medians. Annotation with “****” indicates p<0.0001 by Wilcoxon test.

B) Predicted monomeric AIRE structure (AlphaFold 3) ‘painted’ according to median map scores at each amino acid position. Blue indicates positions intolerant to variation and red indicates those with apparently-increased insulin promoter-GFP reporter activity. Start and stop codons are indicated with Met1 and Ser545, respectively.

C) Median map scores painted onto a predicted structure (AlphaFold 3) of AIRE’s SAND domain.

D) Median AIRE variant effect scores were painted onto the PHD1 interaction with H3K4me0 NMR crystal structure PDB ID:2KE1²⁹. H3K4me0 is painted in green. Zn²⁺ ions are in blue, with Zn²⁺-coordinating positions indicated.

E) Median variant effect map scores painted onto the structure (AlphaFold 3) of the AIRE PHD2 domain, with Zn²⁺-coordinating positions indicated.

We next evaluated trends relating our map scores to the reference (“from”) and destination (“to”) amino acids (Figure S7). The lowest variant map scores were found for variants that replaced large hydrophobic amino acids (score medians for replacement of Phe, Trp, Tyr and Ile were 0.28, 0.53, 0.33, 0.28 as compared to replacement of other amino acids where medians ranged from 0.55 to 0.99). Not surprisingly, large hydrophobic residues were enriched in buried residues (representing 65% of buried positions, but only 24% of surface positions) and also in known AIRE structured domains. Other than one Tyr residue found at amino acid position 394, none appeared within predicted disordered regions.

We assessed the effects of changing within or between different categories of amino acid such as hydrophobic, polar, and positively or negatively charged. Here we expected within-category substitutions to have a higher (less damaging) score than those between categories. Indeed, median scores for within-category substitutions ranged from 0.73-0.95, with the most damaging changes being from one positively-charged residue to another and the least damaging being from one negatively-charged residue to another (Figure S8). Median scores for between-category substitutions ranged from 0.59-0.9, with the most damaging (0.59-0.6) being substitutions of one charged residue for another with opposite-sign charge. We also found that introduction of hydrophobic residues tended to increase reporter activation (p=2×10⁻¹³; Figure S9A) — discussed more below — and that introduction of negatively-charged residues tended to decrease reporter expression (Figure S9B, p=7×10⁻⁸), while introduction of positively charged residues did not (Figure S9C, p>0.05).

We expected our map to capture the trend that introducing proline disrupts secondary structure⁸⁶. Indeed, median scores of initially hydrophobic, polar, positively or negatively charged positions substituted to proline ranged from 0.67-0.73 (Figure S8). In general, substitutions to proline had the lowest (most damaging) median scores compared with substitutions to all other amino acids (Figure S9D, p=3×10⁻⁴). These trends are in keeping with conclusions from larger-scale analyses of other variant effect maps that introducing proline is generally more damaging than introducing other amino acid changes^87,88. Interestingly, replacement of prolines by other amino acids in other categories was often tolerated, with median scores ranging from 0.77-1.07 (Figure S8).

Correspondence of AIRE map scores to known functional elements

Features known to be important for AIRE function were largely recapitulated by our map (Figure 2B). Substitutions were generally deleterious as expected within the CARD domain (positions 1-104), a domain which is essential for AIRE multimerization and its ability to assemble transcriptional condensates. Within the CARD domain, hydrophobic residues within the alpha-helical bundle pointing into the core of the CARD domain (according to the predicted monomer structure) were found intolerant to substitution (Figures S10A and S10B; median buried = 0.03, median surface = 0.22, p>0.05) while the outer alpha helices of the CARD alpha-helical bundle were more tolerant to variation.

AIRE dimer, trimer and tetramer structural models using AlphaFold 3²⁷ (Figure S11), identified residues that were buried by contacts between dimeric, trimeric or tetrameric subunits (see Table S7 for subunit positions). That residues involved in intersubunit interaction showed less tolerance to substitution than non-interacting residues (Figure S11) supports previous reports⁵² that multimerization is important for AIRE function. Dimerization residues demonstrated a greater difference in interacting vs non-interacting residue scores than did trimerization or tetramerization residues (Δmedian scores were 0.26, 0.24, 0.20 for dimer, trimer and tetramer respectively), implying that AIRE dimerization is more important to AIRE function than other multimerization states.

AIRE’s function requires nuclear localization. Our map (Figure 1B) showed intolerance to variation at positions 131-133, the previously reported nuclear localization signal (NLS)⁸⁹. The AIRE SAND domain (positions 181-280), implicated in DNA and protein-protein interaction, does not contain the DNA-binding residues KDWK found in other many other SAND domain proteins that are known to bind DNA such as NUDR, Sp100b, and GMEB1⁹⁰. The corresponding AIRE sequence (NKAR, positions 244-247) exhibited mixed tolerance to variation (mean scores between 0.61 and 0.85). The predominant model is that AIRE upregulates genes which have poised transcriptional complexes⁹¹, but it is not clear that this requires direct DNA binding by AIRE^19,77,92. Our map provides modest support for the proposition that the (potentially) DNA-binding residues NKAR are dispensable for AIRE function.

Within the SAND domain, the map found positions 181-223 intolerant to hydrophobic substitutions, suggesting that these residues are solvent-accessible (Figure 1B). Together with previous evidence that SAND mediates protein interaction⁹³, these positions are candidate protein interaction-mediating residues. The SAND domain’s beta strands (positions 221-231), the only structured SAND residues in the predicted AlphaFold structure, were highly intolerant to variation in our map (Figure 2C). Interestingly, an intrinsically disordered segment just C-terminal to SAND’s beta strands (positions 234-290) had additional intolerant positions (Phe236, Asn244, and Arg257, with mean scores of 0.63, 0.60 and 0.65, respectively). The C-terminal half of the SAND domain is disordered and, while generally tolerant to substitution, was largely intolerant to the introduction of charged residues (Figure 1B). Overall, our map supports the importance of both ordered and disordered segments of the SAND domain.

The PHD1 domain (residues 296-343), which interacts directly with H3K4me0¹⁹, harbors most of the pathogenic AIRE variants reported in individuals who have the autosomal dominant form of APS-1⁴. The damaging nature of these variants is supported by previous small-scale studies using reporter assays^{16,18,19,94–98} and by computational variant predictors REVEL, ESM1b, AlphaMissense and VARITY for the likely pathogenic p.Arg303Gln and p.Pro326Leu variants^34,69,99. The median score at PHD1 positions was 0.98, indicating a high tolerance for substitution, and our map clearly did not capture pathogenic variant effects known for this domain. More specifically, scores for residues known to be directly involved in the interaction of PHD1 to H3K4me0 appeared to tolerate variation in our map (Figure S12). That our map failed to capture the functional importance of the PHD1 domain can also be seen by painting the average score by position onto the existing NMR structure for the PHD1 domain interaction with H3K4me0, demonstrating that the H3K4me0-proximal residues appeared to tolerate modification in our map (Figure 2D).

Although our map did not generally capture PHD1 function, exceptions include residues Cys299 (mean positional score = 0.56) and Cys302 (mean positional score = 0.55), confirming importance of the first Zn²-binding site of PHD1 (Figure 2D), potentially for overall protein stability. Although the pathogenic variant p.Cys311Tyr was previously hypothesized to impact disease through decreased zinc-chelation^4,100, our map showed Cys311 as tolerating substitutions, which is evidence against its involvement in zinc binding. Interestingly, the PHD1 Cys337 and Cys340 residues showed increased reporter activity when modified (mean positional scores = 1.30 and 1.39, respectively; Figure 2D), although the reasons for this are unclear.

The PHD2 domain encodes another zinc-finger, coordinating two Zn²⁺ ions, which is known to be important for protein-protein interactions⁸⁹. The map reveals that within the PHD2 zinc-binding domain, all zinc-adjacent cysteines are intolerant to variation, (Cys434, Cys437, Cys446, Cys449, Cys457, Cys472 and Cys475) in addition to zinc-proximal His454 (Figure 2E). Thus, the map captures the known importance of Zn²⁺ binding of the PHD2 domain to transactivation of the reporter.

AIRE’s C-terminal domain (positions 514-545) has been demonstrated to be important for its transactivation activity, probably through protein-protein interactions with transcriptional coactivators CBP/p300 localized to enhancers^97,101. Our map showed general agreement with this finding, in that variation in this region was not tolerated (Figure 1B). However, our map showed that nonsense variation as described above was tolerated within the last 8 amino acids. An interesting exception to this trend was that the map found substitutions introducing a positively charged residue within the final 8 positions to be damaging (Figure 1B).

Insights from the AIRE map

Two residues (Pro126 and Pro129 with mean positional scores = 0.31 and 0.46, respectively) displayed intolerance to variation but did not correspond to known functional roles. These fell within a disordered region from positions 101 to 178 with very low confidence of being structured according to AlphaFold^85,102. These prolines fit with a larger pattern of many prolines in positions 115-129 (Pro115, Pro116, Pro119, Pro124, Pro125, Pro126 and Pro129). More generally, prolines in intrinsically disordered regions have been described as both disorder promoters and involved in structural compaction¹⁰³, but only Pro126 and Pro129 were intolerant to variation, with Pro126 being especially intolerant. Notably, both of these prolines were close to the NLS (positions 131 to 133¹⁰⁴). Thus, our map allows us to suggest that Pro126 and Pro129 function to maintain proper AIRE conformations that are necessary for NLS function.

Many (3484) missense variants exhibited reporter activation that was higher than that of wild type (functional scores greater than 1). Interestingly, visual examination of the map showed that, within disordered regions in general (excluding the NLS) and especially at SAND domain positions 145-175, greater-than-wild-type reporter activity was frequent for substitutions that newly introduced a hydrophobic residue (Figure 1B). Indeed, substitution to a hydrophobic residue within a disordered region significantly increased reporter activation (median functional score difference between substitutions to hydrophobic and non-hydrophobic amino acids was 0.14, p=1×10⁻²²; Figure S13A). However, within non-disordered regions, no difference was found between substitution to hydrophobic substitutions and non-hydrophobic substitutions (median difference = 0.03, p>0.05; Figure S13B). This phenomenon could be explained by the fact that hydrophobic residues on surface-exposed positions typically provide much of the energy for protein-protein binding¹⁰⁵, with newly introduced hydrophobic residues in disordered regions potentially providing increased binding energy.

In an attempt to distinguish true ‘gain-of-function’ variants, we further assessed one-tailed empirical P-values for each variant using synonymous variants as the reference score distribution. In total, 8 variants were found to significantly upregulate the reporter (p<0.05 after adjusting for multiple testing; Table S10). Amongst variants that we observed to significantly upregulate the reporter, the three highest-scoring were p.Ala138Met, p.Pro162His, and p.Pro162Tyr. To confirm that map findings can be replicated, we applied a small-scale assay to these, to the two low-scoring variants p.Pro126Val and p.Pro129Arg that were potentially related to nuclear localization, as well as additional controls. Flow cytometry demonstrated that all variants and controls behaved in accordance with their map scores (Figure S14), with p.Pro126Val, p.Pro129Arg, and negative controls upregulating the reporter significantly less than WT. The three high-scoring variants showed reporter expression significantly above WT, representing gain of transcriptional function. Of the missense variants which numerically increased reporter expression, 35% (1101/3189) are SNV-accessible and so are likely to already exist in the population. However, only three (p.Thr46Met, p.Pro162His and p.Lys395Met) of the 8 variants (Table S10) exhibiting significantly increased reporter expression are SNV-accessible. UK BioBank analysis did not support an association between variants exhibiting increased reporter expression and altered frequency of AIRE-related symptoms. However, this analysis was based on only 45 participants who all carried the SNV-accessible VUS variant p.Thr46Met (the only variant significantly above a score of 1) that was observed in the UK Biobank, and was therefore not well powered to detect subtle effects.

Using the AIRE map to classify clinical variant pathogenicity

To evaluate the potential of our AIRE map in classifying variant pathogenicity, we first analyzed the tradeoffs between precision (the proportion of variants below a given score that are annotated P/LP) versus recall (the proportion of variants that are annotated as P/LP and scored below a given threshold). More specifically, we reported “balanced precision”, defined as the precision that one would have expected to observe with a balanced test set with equal numbers of pathogenic and benign reference variants, which was derived from nominal precision as previously described³⁴. Here, our variant reference set was composed of 17 variants annotated as B/LB and 36 variants annotated as P/LP with at least one star in ClinVar. Of these, map scores seen in both biological replicates were obtained for 16 B/LB and 30 P/LP. We next evaluated computational predictors including SIFT¹⁰⁶, ESM1b⁹⁹, VARITY³⁴, REVEL⁶⁸, and AlphaMissense⁶⁹. A balanced precision recall curve generated using this reference set (Figure 3A) showed that the map, and most predictors, performed well. REVEL and VARITY were the best predictors in terms of recall at a fixed stringent (90%) precision (R90BP: map=0.87, ESM1b=0.90, AlphaMissense=0.93, VARITY=1.00, SIFT=0.80, REVEL=1.00; Figure 3A). Eliminating variants of the PHD1 domain (one pathogenic and one benign) only marginally improved the performance of the map (R90BP: map=0.90, ESM1b=0.90, AlphaMissense=0.93, VARITY=1.00, SIFT=0.79, REVEL=1.00). Overall, our map performs on par with the computational predictors REVEL, AlphaMissense and VARITY in separating pathogenic from benign missense variation.

Figure 3: Benchmarking of AIRE map to assess performance.

A) Performance of AIRE variant effect map and computational predictors of variant effect (ESM1b⁹⁹, AlphaMissense⁶⁹, VARITY³⁴, SIFT¹⁰⁶ and REVEL⁶⁸) by balanced precision vs recall analysis based on a variant reference set scored by the map and all predictors tested (30 pathogenic/likely pathogenic (P/LP) and 16 benign/likely benign (B/LB) variants).

B) Proportion of map-derived evidence labels for missense variants currently classified on ClinVar to B/LB, Conflicting, P/LP and VUS, and for single nucleotide-accessible variants with no current clinical annotation (None).

We next sought to calibrate map scores to yield a quantitative log-likelihood ratio of pathogenicity (LLR_p) corresponding to each score. LLR_p scores are not only compatible with Bayesian methods for variant interpretation, but can also be converted into the categorical evidence strength labels used within the ACMG/AMP framework^25,37,38 . Briefly, LLR_p scores were calculated in two steps: First, probability density distributions for B/LB and P/LP variants were separately estimated using observed map scores for each set, using kernel density estimation (Figure S15A). These distributions were estimated using reference variants in each set that had been measured in both biological replicates. Given the above-mentioned limitations of our map in detecting damaging variants in the PHD1 domain, we performed this conversion of scores to evidence strength labels using a reference set of 16 B/LB and 30 P/LP variants that excluded one P/LP and one B/LB variant in the PHD1 domain. Second, at each possible map score, LLR_p was calculated as the log₁₀ of the ratio of P/LP probability density to B/LB probability density. As previously described^25,37,38, LLR_p scores were then converted to ACMG/AMP evidence strength labels (Figure S15B). This procedure yielded non-undetermined functional evidence for 70% (242/345) of variants currently annotated as VUS (Figure 3B). Although we retained evidence provided for PHD1 variants with evidence towards pathogenicity (Figure 3B), we declined to provide evidence towards benignity for variants in the PHD1 domain. As a further element of conservatism in providing clinical evidence, we added a “caution” flag to evidence where the confidence interval of score overlapped 0.5 (the midpoint between the medians of nonsense and synonymous variants; caution flags are summarized in Table S4).

We next evaluated concordance of our functional evidence with current annotations of our reference set variants. Our new evidence was concordant with classifications for 39 (89%) of 44 variants if PHD1 variants are excluded and 40 (87%) of 46 if PHD1 variants are included (Figure S16A). Of the 6 variants for which the LLR_p pointed in the opposite direction from current annotation (Table S11), four were not truly discordant in the sense that our map provided no evidence strength label (in one case because we declined to provide evidence towards benignity for PHD1 domain variants). Of the two discordant variants, one was p.Arg9Pro, a known pathogenic variant to which we assigned supporting evidence towards benignity, but this variant received a caution flag based on its confidence interval. The other was p.Arg471Cys, a common variant (MAF of 1-2% in Northern Europeans) which is annotated as benign but which our map assigned pathogenic supporting evidence. That p.Arg471Cys is a known contributor to polygenic risk for Addison’s disease, pernicious anemia and type 1 diabetes^107–111, supports the validity of our map. The AIRE VUS which were given non-undetermined evidence were more often provided evidence towards pathogenicity in known protein domains, as expected, with most variants provided evidence towards benignity in regions of disorder (Figure S16B). Indeed, 47 VUS received pathogenic evidence and of these, 20 (43%) were in the CARD domain, while 27 (57%) were not. Overall, our assay largely recapitulated expected reference set variant effects and generated evidence for 6103 of AIRE variants (of which 1019 had a caution flag) that do not currently have a clinical classification (Figure 3B).

We next re-evaluated the pathogenicity of all AIRE missense variants in ClinVar in light of new evidence from our map, using ACMG/AMP criteria as described in Material and Methods. Variants were subjected to re-evaluation if our map offered a calibrated evidence strength label. Of the 13 conflicting variants which were VUS/B/LB conflicting, 10 were re-classified to B/LB, and 3 remained as VUS. Of the 2 conflicting variants (for which both a VUS and an LP classification had been reported on ClinVar), one was assigned P both with and without our evidence, while the new evidence led us to reclassify the other from VUS to LP. Among the 242 currently-VUS variants evaluated, 133 (55%) were re-classified: 14 (6%) to P/LP (13 to LP and 1 to P) and 49% (119) to LB. Of the 12 B/LB and 23 P/LP we re-evaluated, all remained in the same category except the current P/LP p.Arg9Pro which was downgraded to a VUS albeit with low confidence (caution flag). Thus, evidence from our largely recapitulated previous pathogenicity annotations, while enabling re-classification for a total of 133 variants which had previously been classified as VUS (109 of which were of high confidence, see Table S4), providing confident evidence for 32% (109/345) current VUS.

Dominant negative variant impacts

Some AIRE variants associate more often with dominant inheritance of APS-1, and are thought to be dominant negative (DN) as a result of their ability to sequester WT AIRE^4,112–114. Our assay does not convey any information about the ability of AIRE variants to sequester WT AIRE (HEK293 cells do not express WT AIRE¹⁹), but does allow us to assess the intrinsic activity of AIRE variants. Therefore, we evaluated map scores of a set of 18 known DN variants that have been confirmed in small-scale assays¹¹⁵, comparing them both to B/LB variants and to all other (presumed AR) P/LP variants reported in ClinVar (Figure S17A). While AR P/LP variants showed more damaging scores than B/LB variants as expected (Δmedian score 0.74, p_adj: 8×10⁻⁶), non-PHD1 DN variants had scores not significantly different than those of AR variants (Δmedian score 0.47, p_adj>0.05) or B/LB variants (Δmedian score 0.27, p_adj>0.05). However, most known DN variants are located in the PHD1 domain, for which our assay showed reduced sensitivity (Figure S17B). Four known DN variants fall outside of the PHD1 domain. Two of these, p.Pro368Ala and p.Gly467Arg, did not exhibit damaging scores. Interestingly, the two remaining, well-characterized DN variants that fell outside of the PHD1 domain (p.Gly228Trp and p.Cys446Gly) were both found deleterious in our map, suggesting that they have a functional impact on transcription beyond sequestration impacts.

Usually, DN variants cause less-severe autoimmune symptoms and are associated with organ-specific autoimmunity⁴. However, some DN variants, such as p.Cys302Tyr and p.Cys311Tyr (which are both within the PHD1 domain at positions involved in coordinating Zn²⁺-binding), are known to cause symptoms closely resembling those in individuals with classic APS-1⁴. Of these, p.Cys302Tyr showed moderate functional impact in our reporter assay (score = 0.47), while p.Cys311Tyr was not found to be as damaging in our assay (score = 0.79). Interestingly, a range of phenotypes (from asymptomatic to severe) have been reported for p.Cys311Tyr⁴. Overall, while the map has reduced sensitivity to detect DN variant impacts, we do detect the impacts of some DN AIRE variants on insulin promoter activity.

Comparing map scores to an international APS-1 cohort supports a genotype-phenotype relationship for APS-1

To evaluate the potential predictive value of our AIRE map in the context of human phenotypes, we examined a well-phenotyped cohort of 98 individuals with APS-1 collectively described in 29 studies^4,39–66, including 21 with homozygous missense variants, 45 having a missense variant in trans with a nonsense or splice site variant, and 32 carrying a single heterozygous missense variant. Within this cohort, 45 unique symptoms were collectively described (Table S12), with only 26% of these seen in more than 10 participants. Each study reported on the phenotypes present that are potentially relevant to APS-1, without mentioning removal of phenotypes a priori. A total of 16 AIRE variants that had not yet been reported to ClinVar were present in this cohort.

An initial barrier to quantitatively evaluating genotype-phenotype associations in the context of our maps was that our map provides functional information for single-variant AIRE alleles, rather than diploid AIRE genotypes. We therefore derived an combined-allele map (CAM) score by combining map scores for AIRE variants from both alleles (using simple addition of variant effect map scores for variants on different alleles, as was previously shown to yield significant genotype-phenotype correlation for the protein CBS²⁴). For the purpose of calculating the CAM score, splice site and nonsense variants were modeled as null (i.e., given scores of 0) and WT alleles were assigned a score of 1.

A second barrier to this analysis was the lack of widely accepted quantitative measures of APS-1 disease severity. To address this, we derived a simple proxy for severity: the number of symptoms reported for each affected individual. The number of symptoms correlated negatively with CAM scores, such that individuals with a lower (more damaging) CAM score tended to exhibit more symptoms (R=−0.4; p=4.5×10⁻⁵; n=97; Figure S18A). We worried that this correlation might have resulted from the fact that PHD1 variants may both yield less severe disease and also be less readily detected in our assay. However, after eliminating any participants with at least one PHD1 variant, a significant negative correlation remained (R=−0.29; p=0.01; n=71, Figure 4A). These results provide statistically significant evidence of quantitative genotype-phenotype relationship for APS-1.

Figure 4: Genotype-phenotype assessment of AIRE map against individuals with APS-1 or related symptoms.

A) For a cohort of individuals with APS-1 (n=71; excluding those with variants in the PHD1 domain), the relationship between the combined-allele map (CAM) score (defined as the sum of functional scores for both alleles, where nonsense/frameshift and WT alleles are assigned scores of 0 and 1, respectively) and the number of APS-1 symptoms (R=−0.29, p=0.01).

B) Distributions of CAM scores for members of the cohort who either do (“dyad”, n=53) or do not (“else”, n=18) exhibit the diagnostic dyad of APS-1 symptoms (“**” indicates p=6×10⁻³, by Wilcoxon). We excluded individuals with PHD1 variants (see Figure S18 for a PHD1-inclusive version of panels A and B with similar conclusion).

C) Confidence intervals for the log-odds ratio of having a hypoparathyroidism diagnosis for those UKBB participants carrying at least one rare (<1×10⁻⁴ gnomAD minor allele frequency) AIRE variant deemed intolerant to variation (these included classical predicted loss-of-function (cLoF) variants including stop, frameshift or canonical splice site variants, or missense variants that were damaging according to our variant effect map or computational predictors AlphaMissense (AM) or REVEL). Here log-odds ratio = 7.4[3.1,18.1], p_adj=7×10⁻⁴ for the variant effect map and cLOF (N <= 5); OR = 5.0[1.9,13.5], p_adj=9×10⁻³ for REVEL and cLOF (N <= 5); log-odds ratio = 5.4[2.2,13.1], p_adj=3×10⁻³ for all predictors and map combined (N <= 5).

D) As in C), but considering the log-odds ratio of having a Vitamin B12 deficiency anemia diagnosis. Here, log-odds ratio = 1.8[1.1,3.1], p_adj=4×10⁻² for the variant effect map and cLOF (N observed participants with symptom = 14). Annotation with “*” =p_adj<0.05.

As another proxy for disease severity, individuals with APS-1 were grouped into those with and without the presence of the APS-1 diagnostic dyad (two out of hypoparathyroidism, adrenal insufficiency or chronic mucocutaneous candidiasis). Those presenting with the diagnostic dyad had CAMs that scored significantly lower in the map than those in the non-dyad group (median dyad CAM=0.22, median non-dyad CAM=1.68; p=2×10⁻⁷; Figure S18B). Similarly, excluding individuals with PHD1 variants from analysis demonstrated significantly lower CAMs for those with the diagnostic dyad (median dyad CAM=0.20, median non-dyad CAM=0.64; p=6×10⁻³; Figure 4B). Overall, our results provide evidence for a genotype-phenotype correlation for APS-1 in this clinical cohort.

UK BioBank participants with damaging AIRE variants associate with hypoparathyroidism

We next evaluated whether variants identified in the AIRE map were associated with APS-1 and its related symptoms in whole exome sequencing data from 469,765 UK BioBank participants¹¹⁶. APS-1 is a very rare disease, and indeed only nineteen participants had an APS-1 diagnosis in UK BioBank at the time of our analysis, of which only four participants each had one of three AIRE missense variants. The lack of sequenced variants in this group could be attributed to AIRE’s high GC content, as GC-rich regions can limit coverage by exome sequencing.

AIRE genotypes have also been associated with autoimmune disease in the heterozygous state⁴. This heterozygous set identified 22 classical predicted loss of function (cLOF; stop, frameshift or canonical splice site) variants, and 382 rare missense variants in AIRE. To evaluate whether having at least one damaging AIRE allele is associated with APS-1-related symptoms we examined symptoms commonly associated with AIRE variants, including both endocrine symptoms (insulin-dependent diabetes mellitus, adrenal insufficiency, and hypoparathyroidism) and non-endocrine symptoms (candidal stomatosis, chronic atrophic gastritis, vitamin B12 deficiency anemia, and others (see Material and Methods)).

We assessed the association between these APS-related symptoms in UK BioBank and participant sets defined by presence of at least one rare (minor allele frequency of <1×10⁻⁴), putatively damaging AIRE variant from five non-exclusive definitions of damaging: 1) classical predicted loss of function (cLoF; stop or frameshift or canonical splice site variants) only, 2) either cLoF or a damaging score in our variant effect map, 3) cLoF or a damaging AlphaMissense score, 4) cLoF or a damaging REVEL score and 5) Any, i.e., cLoF or AlphaMissense or REVEL or our map (see Material and Methods for the thresholds that define damaging variants from each score source).

Hypoparathyroidism showed consistent association with AIRE scores across the sets with the strongest associations coming from the variant effect map (OR = 7.4[3.1,18.1], p_adj=7×10⁻⁴; Figure 4C) and REVEL (OR = 5.0[1.9,13.5], p_adj=9×10⁻³). A combination of scoring methods (generating participant set 5, above) also showed a strong hypoparathyroidism association (OR = 5.4[2.2,13.1], p_adj=3×10⁻³). The most conservative model (scoring only cLoF variants) captured fewer variants and did not meet nominal significance (OR = 4.3[1.1,17.2], p_adj=0.08). We also observed increased risk for vitamin B12 deficiency anemia using a combination of cLoFs and our map (participant set 2; OR=1.8[1.1-3.0], p_adj=0.04; Figure 4D). These results provide additional support for the role of AIRE in autoimmune disease susceptibility, and for the value of our variant effect map in identifying variants that contribute to disease susceptibility.

Discussion

Prior to this study, experimental functional impact data had been reported for fewer than 7% of clinically-reported AIRE missense variants. Where they had been studied, the data often became available years after individuals present with symptoms. This lagging ‘reactive’ pace of functional evaluation is unfortunate, especially given that essentially all possible single-nucleotide variants already exist in someone alive today¹¹⁷. Here we provide a systematic missense variant effect map of AIRE, evaluating 9,790 AIRE amino acid substitutions for their ability to enhance expression at the insulin promoter. Amongst these, we assayed 3,127 (98%) of the 3,189 missense variants that can occur via a single-nucleotide change from the reference sequence (the amino acid substitutions most likely to be clinically observed). Benchmarking of the map against a set of clinical reference variants allowed us to provide confident functional evidence for 32% (109/345) current VUS. Additionally, we identify a genotype-phenotype correlation for scores and APS-1 severity for an international clinical cohort, and associate decreased-function variants with hypoparathyroidism and vitamin B12 deficiency anemia in a dominant model of inheritance in the UK BioBank. Importantly, this AIRE missense variant map proactively provides scores for yet-to-be-observed clinical variants.

Our AIRE variant effect map was well aligned with expectation in the sense that residue positions known to be important for structure and function were less tolerant to variation. Such positions included buried residues, NLS residues, and residues of known domains critical for AIRE function such as the CARD, SAND, PHD2 and C-terminal domain. One exception was substitutions in PHD1 domain positions which, except for p.Cys299 and p.Cys302 residues involved in zinc-binding, appeared to be generally tolerated. Given that this domain is known to be important for AIRE function where promoters fall within closed chromatin, we consider this a limitation of our map. We attribute this limitation to the fact that our insulin promoter reporter construct was located within an open-chromatin safe harbor site, so that AIRE’s PHD1 function was not necessary to enhance reporter expression. A previously-reported functional assay of variant effects demonstrated that variants of the PHD1 domain could be captured (such as the pathogenic p.Cys311Tyr variant) using a non-endogenous synthetic basal insulin promoter additionally contained either class I or class III variable number of tandem repeat (VNTR) of the IDDM2 locus upstream of the insulin promoter, known to contribute to Type 1 Diabetes susceptibility¹⁸. We speculate that this longer promoter region better captured PHD1 variant impact through more extensive chromatinization of the transient plasmid reporter compared to the insulin promoter used in this work. Despite our assay’s PHD1 blind spot, our use of an open-chromatin reporter may have had a compensating advantage for non-PHD1 functions, in that AIRE is known to provide expression at closed chromatin only stochastically¹¹⁸ with presumably less-robust signal output.

We used our map to identify AIRE positions, which had not been previously reported as functionally important, as having either loss- or gain-of-function impacts according to our reporter activation assay. For example, the two residues p.Pro126 and p.Pro129 were intolerant to most substitutions. Although we do not know the mechanism for the importance of these residues (and no clinical variants at these positions have been reported), they are located in a disordered region proximal to the NLS. Given that prolines have often been implicated in maintaining structural compaction of disordered regions¹⁰³, we hypothesize that these prolines serve to maintain proper NLS conformation. Follow-up subcellular localization studies of substitutions could help elucidate their role. We also noted an interesting trend that variants increasing reporter activation were often hydrophobic residues introduced to disordered regions. Indeed, in small-scale follow-up testing, we were able to experimentally confirm increased reporter activation for three of these substitutions: p.Ala138Met, p.Pro162His, and p.Pro162Tyr. It will be interesting, as larger genotyped and phenotyped cohorts become available, to identify whether variants increasing reporter activation are associated with any clinical outcomes.

The map achieved excellent performance in the sense that the score distributions for known pathogenic and benign were markedly different, especially after excluding the PHD1 domain. One potential caveat is that some pathogenic variants had been classified as such using functional evidence. This was the case for at least 8 (24%) of 33 ClinVar pathogenic variants (and possibly more, as only 22 ClinVar-reported variants had an explanatory paragraph which would have noted functional evidence where it exists). Thus, reclassification based on our map could be ‘double-counting’ functional evidence in some cases. However, when we examined ClinVar for a random sample of 33 VUS, only one had reported functional evidence, thus supporting the accuracy of our reclassification of 55% (133/242) of VUS.

The ability of our map to distinguish pathogenic and likely-pathogenic variants from benign and likely benign variants was on par with the current standard of computational effect predictors including REVEL, AlphaMissense, and ESM1b. Current clinical guidelines¹⁰ treat computational and experimental evidence as independent, so that these approaches are complementary rather than being in competition. In converting functional scores to LLR_p scores that reflect ACMG/AMP categorical evidence strengths, we provided functional assay evidence for 71% (2270/3189) of all possible SNV-accessible missense variants, including 70% (242/345) of variants currently annotated as VUS, including SNV-accessible variants within the PHD1 domain that showed evidence towards pathogenicity. In addition, we provide confident, formal re-classification for 109 VUS according to current ACMG/AMP guidelines.

The AIRE map enabled us to characterize genotype-phenotype relationships within cohorts. For those variants from individuals diagnosed with APS-1 which also had scores from the map (49/50), 67% (33/49) had a score below 0.5 and 59.2% (29/49) had a score below 0.371 (the threshold below which we provided evidence towards pathogenicity). Surprisingly, given the typically low penetrance of APS-1 (even siblings in the international clinical cohort did not have the same number or types of symptoms), individual-level CAM scores derived from our map correlated with two different proxy measures of APS-1 severity, and also with the presence/absence of the APS-1 diagnostic dyad. Using the UK BioBank, additional associations were found between the presence of a rare and damaging AIRE variant (using a dominant inheritance model) with both hypoparathyroidism and vitamin B12 deficiency anemia. These results complement previous reports describing mono-allelic effects for AIRE mutations within the PHD1 domain⁴, in that both of the associations detected here using our map-derived CAM scores were entirely driven by non-PHD1 variants.

Given that the HEK293 cell line we used does not recapitulate all physiological conditions of the thymus in vivo, and given that AIRE may function differently at different target gene loci, perhaps we should have expected our assay to be uninformative about pathogenicity. If that had been the case, we might have pursued a potentially-more-faithful thymic cell line and sought to optimize high-throughput transfection and evaluation via flow cytometry. Fortunately, as indicated in our small-scale evaluation of the HEK293-based insulin reporter assay, and as borne out in the final map, the more facile HEK293 assay appears highly predictive of pathogenicity. Success with HEK293 cells is consistent with their previous use in a range of biochemical experiments fundamental to the characterization of AIRE function^92,119.

One limitation is that our assay was based on variant AIRE cDNAs, such that impacts of SNVs (either coding or non-coding) on splicing will not have been captured. In future work, by introducing AIRE variants at the endogenous locus and comparing the results with our cDNA-based map, one might hope to identify variant impacts on splicing, mRNA stability, enhancer and other non-coding functions.

Another limitation is that, while our assay can detect coding variant impacts on AIRE’s transcriptional function, it cannot inform us as to the mechanism of functional impact, such as loss of stability or weakened protein interaction. Variant effect maps examining additional ‘subfunctions’ could address this, for example, by systematically testing AIRE variant impacts on protein abundance⁷⁹ or nuclear localization¹²⁰.

We acknowledge the limitation that this assay was only performed with one reporter gene. To address this, it would be ideal to perform follow-up studies to potentially fill in “blind spots” and identify context-dependent variant effects using many reporter genes. It would be ideal to expand this work to assess other AIRE-induced genes to evaluate whether there are promoter target-specific differences in AIRE variant impacts. Interestingly, AIRE-downregulated genes have also been reported such as DUSP4^78,121, which would be interesting to compare against our findings. To accomplish this (and also address our map’s ‘blind spot’ in the PHD1 domain) one might generate a variant effect map using single-cell RNA sequencing (scRNA-seq)¹²². This approach would benefit in that variant impacts on many different target genes could be assessed in parallel, rather than using insulin alone, and include more genes located to regions of closed chromatin. This approach would be limited by the stochastic nature of AIRE, such that different cells upregulate different genes¹¹⁸, and compounded by the limited number of independent reads that can be assayed in each cell using scRNA-seq. However, these limitations could be overcome by aggregating the results from ‘like’ sets of genes (for example, genes for which there is a priori evidence for the assayed cell line that the promoter lies within closed chromatin).

Our APS-1 cohort-related analyses did not consider age, which is a limitation because individuals with APS-1 develop more severe symptoms over time. Our CAM scores also did not account for the effects of dominant negative variation, and ideally a future assay would be performed in the presence of an expressed WT AIRE, at least for variants with known dominant negative impacts. Also, where individuals in the APS-1 cohort had greater than one AIRE variant, we were unable to disentangle the effect of one variant from another. Another limitation is that we may not have detected all AIRE variants that are present, either for the APS-1 clinical cohort or for UK BioBank participants: While this is a formal possibility for any gene, AIRE is GC-rich and has reduced sequence coverage relative to other genes. However, any such ascertainment biases would have led to underestimation rather than overestimation of phenotypic associations.

Perhaps the most important use of our AIRE variant effect map will be as proactive (and therefore immediately available) functional evidence for classifying AIRE variants in individuals suspected to have APS-1, even for variants that have never before been clinically observed. This evidence promises to improve prognosis through earlier identification of individuals who are at risk for APS-1 and may therefore benefit from earlier therapeutic intervention. Detection of pathogenic variants can also enable ‘cascade screening’ to identify first-degree relatives with the same variant who may also be at risk. Our map also provides evidence of benignity, with only one apparent blind spot in the PHD1 domain. As additional clinical variant data becomes available, we may be able to stratify evidence calibration within specific protein domains. We also note that procedures to calibrate map scores are currently evolving given different definitions of what constitutes a benign ‘truthset’¹²³, and the use of these scores in the clinic should ultimately be calibrated according to guidelines set by clinical panels such as the ClinGen AIRE Variant Curation Expert Panel.

Additionally, we have identified features of AIRE function that can lead to further studies and characterization, such as the impact of variants that can increase reporter expression, and positions with deleterious impact that had not previously been reported. We have also identified genotype-phenotype correlations for APS-1-related symptoms using our map scores. Thus, we provide a variant effect map for AIRE, contributing to AIRE clinical variant interpretation, to the knowledge of sequence-structure-function relationships, and also to the growing atlas of experimentally measured human variant effects¹²⁴.

We measure the impact of every possible amino acid change in the Autoimmune Regulator protein AIRE via fluorescent reporter assay, providing a variant effect map and proactive evidence which enables more accurate pathogenicity classification of missense variants. We also report AIRE genotype-phenotype correlations within each of two cohorts.

Data and Code Availability

All files and scripts used for both analysis and figure generation are available at https://github.com/axakova/AIRE. AIRE functional scores, LLR_p’s, evidence strengths and formal classifications are available in Table S4, and are also hosted on MaveDB¹²⁵ (accession number: urn:mavedb:00001255-a-2).