Distinct HLA associations of LGI1 and CASPR2-antibody diseases

The mechanisms responsible for the generation of LGI1- or CASPR2-antibodies are unknown. Binks et al. describe two strikingly dichotomous HLA-associations, most significantly in HLA-DRB1*07:01 for LGI1, and HLA-DRB1*11:01 for CASPR2. Prediction of antigen-specific HLA-binding peptides generates testable hypotheses for T cell identification.


Introduction
The discovery of autoantibodies against leucine-rich, glioma-inactivated 1 (LGI1), contactin-associated protein 2 (CASPR2) (Irani et al., 2010;Lai et al., 2010) and, more recently, intracellular epitopes of voltage-gated potassium channels (VGKCs) (Lang et al., 2017), have redefined the immunology of the VGKC-complex (Thieben et al., 2004;Vincent et al., 2004). Patient stratification by these antigenic targets has shown that the 'double-negative' VGKC-complex antibodies, those without LGI1 or CASPR2 reactivities, are observed across all ages, in healthy controls and in a variety of syndromes, many of which are not immune-mediated (Graus and Gorman, 2016;van Sonderen et al., 2016;Lang et al., 2017). In contrast, patients with LGI1 or CASPR2 antibodies often have clinically-indistinguishable late-onset forms of limbic encephalitis and neuromyotonia with associated dysautonomia, sleep disturbances, pain and seizures (Irani et al., 2010;Lai et al., 2010;Klein et al., 2013;Gadoth et al., 2017). While these features occur at different rates in LGI1versus CASPR2-antibody cohorts, only faciobrachial dystonic seizures (FBDS) robustly predict LGI1 reactivity (Irani et al., 2011;Gadoth et al., 2017;Thompson et al., 2018). Furthermore, these two autoantibodies are both often of the IgG4 subclass and frequently co-exist in patients with the ultra-rare Morvan's syndrome (Irani et al., 2012;Ariñ o et al., 2016). The striking overlaps of these rare neurological features and autoantibodies, and the frequent co-expression of their antigenic targets within mammalian CNS-membrane complexes (Irani et al., 2010;Binks et al., 2018), suggest they are involved in autoimmunization. Indeed, this has been reported in abattoir workers with autoantibodies against VGKC-complexes and, less so, CASPR2 (Meeusen et al., 2012). The nature of the available complexes, antigen presentation mechanisms and the available T cell repertoires are likely to determine which antigen dominates the ensuing T-B cell response. If so, human leucocyte antigen (HLA) variants, intimately related to antigen presentation, may play critical roles in distinguishing the aetiology of these syndromes.
Previously, high rates of adverse drug reactions were observed in patients with LGI1 antibodies, typically secondary to antiepileptic drugs (AEDs) and, less so, corticosteroids (Irani et al., 2011(Irani et al., , 2013Thompson et al., 2018). As HLA variants have been implicated in several adverse drug reactions, including those associated with AEDs and immunosuppressants (McCormack et al., 2011;Yip et al., 2015), and have essential antigen-presenting functions (Trowsdale and Knight, 2013), we hypothesized that HLA associations existed in patients with LGI1 antibodies. Indeed, recently, HLA-DRB1*07:01, HLA-DQB1*02:02 and HLA-DRB4 were found to be present in varying proportions of patients with LGI1 antibodies in two cohorts totalling 40 patients, from Korea and the Netherlands (Kim et al., 2017;van Sonderen et al., 2017).
To extend these early observations, and given the hypothesis that the VGKC complex may be the initiating immunizing agent, we sought to compare and contrast HLA-associations in a sizeable cohort of clinically wellcharacterized patients with antibodies against LGI1, CASPR2, both LGI1 and CASPR2, and VGKCs, and in silico to identify peptides that may be presented by these HLA molecules.

Patients
One hundred and eleven Caucasian patients were identified from previous studies (n = 51) (Irani et al., 2011(Irani et al., , 2013Lang et al., 2017), referrals to the Oxford Autoimmune Neurology Group (n = 46) or from the Autoimmune Encephalopathy Clinic, University of California San Francisco (n = 14). These patients had serum antibodies against LGI1 only (n = 68), CASPR2 only (n = 31), both LGI1 and CASPR2 (n = 3) or intracellular aspects of VGKCs (n = 9), as determined by previously described antigen-specific cell-based assays (Irani et al., 2010;Lang et al., 2017). Clinical phenotypes, including information relating to past medical history and adverse drug reactions (Table 1), were evaluated via direct patient and relative interviews and case-note reviews. All patients provided written informed consent (REC16/YH/0013 or the IRB 10-04905 approvals).
To complement this, DRB1, DRB4 and DQ alleles underwent intermediate-resolution HLA-typing using PCR-sequencespecific primers (SSP), updated from Bunce et al. (1995). PCR-SSP defined the first-field plus a string of second-field possibilities: the highest frequency allele in Caucasians was considered most likely. For all discordant data, the PCR-SSP first-field was accepted as the final result. HLA alleles from 5553 Caucasian healthy controls (from Oxford Biobank) were available from imputation using the same platforms, and confirmed in 70 individuals within the same laboratory by PCR-SSP (Neville et al., 2017). Probable haplotype blocks were calculated on the basis of a Bayesian algorithm using PHASE V2 software with 10 000 iterations for three haplotype blocks: HLA-C-B, HLA-DRB1-DQA1-DQB1, and HLA-DPA1-DPB1 (Stephens et al., 2001;Stephens and Donnelly, 2003). Further details on genotyping and imputation are provided in the Supplementary material.

Statistical analyses
For each antibody group, Fisher's exact test (two-tailed) was used to compare the HLA allele and haplotype carrier frequencies between patients and the healthy control dataset. Hochberg's method was used to correct for multiple comparisons. Corrected P-values 5 0.05 were considered significant, and are presented. Odds ratios (ORs) were calculated using the median-unbiased estimation method.

HLA binding predictions
The NetMHCIIpan 3.1 server model based on artificial neural networks (Andreatta et al., 2015) evaluated HLA haplotype binding affinities for 15-amino acid-long consecutive overlapping peptides from full-length LGI1 and CASPR2 sequences (UniProt accession numbers O95970 and Q9UHC6, respectively). Predicted peptide affinities (nM) were compared to 200 000 random peptides of the same length to generate rank values: this measure is less susceptible to the intrinsic capacity of some HLA alleles to generate high-affinity predictions, and rank values (%) 53 were considered strong binders. As expected, consecutive 15mer peptides with high rank values often shared a core sequence.
Given these significant and distinct allelic and haplotypic HLA associations, for each serologically-defined group, we evaluated their value in explaining sub-phenotypes (limbic encephalitis or epilepsy; peripheral or CNS), long-term outcomes or adverse drug reactions (Supplementary Table 1), and found no significant HLA allele or haplotype associations. However, within LGI1-antibody patients, five of six patients with antibiotic-induced rashes carried HLA-B*57:01 known to associate with risk of rash to abacavir and flucloxacillin (Yip et al., 2015), and four of six patients with psoriasis harboured the psoriasis risk allele C*06:02 (Arakawa et al., 2015), suggesting the extended haplotypes may explain these specific co-morbidities. Finally, from the nine LGI1-and four CASPR2-antibody patients with a tumour, there were no significant HLA differences compared to non-tumour patients (Supplementary Table 2).

Predictions of HLA-binding peptides
These robust HLA class II associations strongly implicate CD4 + T cells in the pathogenesis of both LGI1-and CASPR2-antibody-associated diseases. To locate potentially high-affinity peptides that complex with HLA class II heterodimers, and may interact with patient T cells, in silico haplotype (B) associations and their frequency in patients with antibodies to LGI1 (n = 68, red denotes significant associations) and CASPR2 (n = 31, blue denotes significant associations), together with the frequency of these alleles or haplotypes in 5553 healthy controls (black bars). HC = healthy controls.
modelling was used and focused on all the class II haplotypes identified above (Fig. 2).
As expected for the shared core sequences between consecutive 15-mers, many highly ranked peptides were from tightly clustered locations within the full-length protein (Fig. 2B, D and Supplementary Table 4). Most peptides within these clusters showed potential to bind the HLA variants observed in both the LGI1-and CASPR2-antibody cohorts ( Fig. 2B and D, black circles). This included one previously identified peptide (Kim et al., 2017) and argues against its role in disease specificity. However, 9/13 LGI1derived peptides and 7/13 from the CASPR2 sequence showed binding potential that was more restricted to the variants associated with the corresponding antibody cohort ( Fig. 2B and D, pink circles). From LGI1, 4/9 core peptides were predicted to bind with high affinity (540 nM), typically to HLA-DRB1*07:01, although interestingly the highest affinity peptide was predicted to bind HLA-DPA1*02:01-DPB1*11:01 (Supplementary Table 4). From CASPR2-derived peptides, 7/7 were predicted to bind with high affinity, distributed across the variants within the Figure 2 Peptides derived from full-length LGI1 and CASPR2 predicted to bind MHC-dimers encoded by over-represented HLA haplotypes. Rankings and position of peptides derived from full-length sequences of LGI1 (A and B) and CASPR2 (C and D). The haplotypes correspond to Fig. 1B and when in bold they relate to those observed in patients with antibodies to the corresponding protein. Red circles denote the LGI1-antibody cohort and blue the CASPR2-antibody cohort. Grey circles and italicized haplotypes relate to peptides from the other antigenic protein (i.e. CASPR2 in A and B; and LGI1 in C and D). Rank describes the predicted peptide affinities (IC 50 , nM) by comparison to 200 000 random peptides of the same length. Dotted lines represent the 3% cut-off for peptide rank. Within B and D, circles represent the highly-ranked peptides across the full-length sequences of LGI1 or CASPR2: black circles represent peptides with some predicted promiscuity across LGI1-and CASPR2-antibody HLA variants, whereas pink circles highlight peptides that are not predicted to cross-react. Table 4).

Discussion
This study is the first comparative HLA analysis of LGI1and CASPR2-autoantibody mediated diseases, and shows marked and strikingly different HLA associations for these patients, at both allelic and haplotypic levels. Given the frequently overlapping clinical features in patients with LGI1 and CASPR2 antibodies, and their co-expression in VGKC complexes, these findings indicate that dichotomous predisposing HLA variants govern the generation of LGI1 versus CASPR2 antibodies. Furthermore, they strongly implicate T cells in disease initiation and the candidate HLAbinding peptide partners generated by our in silico data may help identify these interacting T cells.
While HLA-DRB1*07:01 and linked class II alleles, including the haplotype HLA-DRB1*07:01-DQA1*02:01-DQB1*02:02, showed very strong associations with LGI1-antibody patients, this was not observed among CASPR2-antibody patients in whom we found clear associations with HLA-DRB1*11:01 only. Among LGI1-antibody patients, DRB1*11:01 was observed at around healthy control rates, DRB4 was less frequently detected than DRB1*07:01, homozygosity for HLA-DRB1*07:01 was recognized, and other independent associations involved HLA class I alleles HLA-B*57:01 and HLA-C*06:02. Albeit limited by their intrinsic rarity, intriguingly, the three patients with both LGI1 and CASPR2 antibodies had yet another complement of HLA alleles. Perhaps this implicates further divergence in molecular mechanisms responsible for the generation of both autoantibody specificities within an individual. However, the HLA associations do not appear to distinguish between sub-phenotypes, outcomes or, in contrast to a previous observation, the presence of associated tumours (van Sonderen et al., 2017). Also, the 9-27% frequencies of these HLA variants in healthy Caucasians are far higher than disease prevalence, implicating additional loci, environmental or stochastic influences in disease manifestation.
Furthermore, our data also provide several intriguing insights into the immunopathogenesis of these diseases. First, they extend the frequent HLA associations in IgG4-related diseases (Huijbers et al., 2015), but here, exceptionally, with no DQ5 association. Second, the presence of dominant HLA class II associations implicates extracellular antigen processing and CD4 T cells in disease initiation (Trowsdale and Knight, 2013), but the LGI1 antibody class I associations found here, and HLA-B*44:03 and HLA-C*07:06 reported in seven patients previously (Kim et al., 2017), are compatible with a role for intracellular antigen processing, including viruses and drugs. These class I differences between studies may be explained by ethnicity, sample size and relatively weak associations (Kim et al., 2017;van Sonderen et al., 2017). Indeed, this extended haplotype and the related complex linkage disequilibrium in this region of the genome warrant further analysis. Furthermore, our original hypothesis of adverse drug reaction-related HLA variants may relate to the linked adverse drug reaction-related class I and II HLA variants (HLA-DRB1*07:01, HLA-DQA1*02:01, HLA-B*57:01) (Yip et al., 2015). These and future observations in patients with LGI1 antibodies may inform the genetic basis of more common adverse drug reactions. Third, the HLA similarities between tumour and non-tumour LGI1-antibody cases suggests the absence of a unique paraneoplastic signature, in contrast to Lambert-Eaton myasthenic syndrome (Wirtz et al., 2005). Perhaps this implies tumours in patients with LGI1 antibodies largely reflect the agematched background rate, rather than a distinct immune mechanism, although a paucity of tumours classically associated with paraneoplastic neurological syndromes may limit our interpretation. Finally, our in silico predictions suggest that HLA-DQA1*02:01-DQB1*03:03 is unlikely to mediate presentation of LGI1-derived peptides, whereas the HLA-DQA1*05:01-DQB1*03:01 heterodimer may be implicated in the CASPR2-antibody phenotype. Also, the promiscuity of both CASPR2 and LGI1 peptides for some HLA variants, including HLA-DRB1*07:01 (Kim et al., 2017), may explain why immunization with the same VGKC complexes may generate two distinct disease entities, and underlie the observed co-existence of both antibodies at rates far higher than expected by chance (Irani et al., 2012). However, this in silico approach is inherently limited by the possibility that high affinity peptides are more effectively deleted through central tolerance. Nevertheless, taken together, the range of antigen-restricted peptides derived herein, and the relative HLA variant frequencies in disease versus control populations, generate hypothesisdriven approaches to expand disease-specific T cells in vitro and complement recent clinical and laboratory observations which strongly implicate T cell dependence of antibodymediated diseases Wilson et al., 2018a, b).
In summary, the distinct HLA associations in patients with LGI1 and CASPR2 autoantibodies, together with differing clinical features relating to autoimmunity, support an immunological dissociation in generation of these clinicallyoverlapping autoantibody-mediated syndromes. The dominant class II HLA involvement combined with in silico predictions, offers potential to better understand the likely initiating T-B cell interactions. Further work should focus on the environmental factors that influence the presentation of peptides in genetically predisposed individuals.