Polyclonal lymphoid expansion drives paraneoplastic autoimmunity in neuroblastoma

SUMMARY

Neuroblastoma is a lethal childhood solid tumor of developing peripheral nerves. Two percent of children with neuroblastoma develop opsoclonus myoclonus ataxia syndrome (OMAS), a paraneoplastic disease characterized by cerebellar and brainstem-directed autoimmunity but typically with outstanding cancer-related outcomes. We compared tumor transcriptomes and tumor-infiltrating T and B cell repertoires from 38 OMAS subjects with neuroblastoma to 26 non-OMAS-associated neuroblastomas. We found greater B and T cell infiltration in OMAS-associated tumors compared to controls and showed that both were polyclonal expansions. Tertiary lymphoid structures (TLSs) were enriched in OMAS-associated tumors. We identified significant enrichment of the major histocompatibility complex (MHC) class II allele HLA-DOB*01:01 in OMAS patients. OMAS severity scores were associated with the expression of several candidate autoimmune genes. We propose a model in which polyclonal auto-reactive B lymphocytes act as antigen-presenting cells and drive TLS formation, thereby supporting both sustained polyclonal T cell-mediated anti-tumor immunity and paraneoplastic OMAS neuropathology.

In brief

Rosenberg et al. sought features of the systemic immune response underlying improved tumor outcomes and neurological dysfunction in patients with opsoclonus myoclonus ataxia syndrome (OMAS), an autoimmune disease caused by neuroblastoma. Diverse B and T cell lymphocytic infiltration organized in enriched tertiary lymphoid structures predominate in OMAS-associated tumors.

Graphical Abstract

INTRODUCTION

Immune surveillance, the idea that the immune system plays an important role in eliminating tumor cells, was first introduced over a hundred years ago by Paul Ehrlich.¹ The complex process of immune modulation (“immune editing”) of tumor growth is robustly supported by multiple mouse studies that demonstrate spontaneous tumor generation and metastasis (reviewed in, e.g., Swann and Smyth²). This editing process, involving early elimination of tumor cells, an equilibrium of evolving tumor and immune restriction, and eventual tumor escape, finds abundant support in human disease as well. Complete spontaneous regression of certain types of neural crest cancers, such as neuroblastoma and melanoma,³ likely represent extreme successes of effective immune surveillance to eliminate cancer in humans. Careful investigation of rare patient populations that exhibit particularly effective deployment of immune surveillance is therefore warranted.

In rare instances in a naturally occurring setting, individuals with solid tumors develop autoimmunity triggered by the tumor, a condition termed paraneoplastic autoimmune disease. Many of these paraneoplastic diseases involve self-antigens that are expressed in endogenous tissue of the central nervous system (CNS), causing severe neurological symptoms that range from psychosis (e.g., NMDA-receptor encephalitis, driven by teratoma; reviewed in Dalmau et al.⁴) to motor deficits, mood and behavioral changes, paralysis, and other symptoms (e.g., limbic encephalitis associated with non-small cell lung cancer [NSCLC]⁵). The autoimmunity is presumed to be driven by a shared epitope between tumor and brain.⁶ Consistent with an important role for immune mechanisms in controlling tumor growth, patients with paraneoplastic diseases often have better tumor-related outcomes than patients with the same cancer but no autoimmune component.^7–10 Improved tumor outcomes may arise in the context of complete or partial tumor elimination; paraneoplastic disease may persist even in the absence of remaining tumor cells.

A hallmark of adaptive immunity is the remarkable combinatorial potential of lymphocytes, which are able to generate diverse antigen receptors, permitting broad and potent immunity. However, the same diversity that protects from a broad array of foreign antigens carries greater risk for autoimmunity. Therefore, in mammals, negative selection in the thymus and the bone marrow is needed to cull self-reactive immune receptors to prevent targeting of self, causing autoimmunity. The paradox of paraneoplastic disease, then, is that patients with autoimmunity possess a broader repertoire of immune reactivity with which to restrict or eradicate solid tumors than patients with proper immune selection, even as it leads to pathology of native tissue. Further evidence of this tenuous relationship is the observation that cancer patients treated with checkpoint inhibitors often develop autoimmunity.^11,12 Understanding how the delicate balance between powerful anti-tumor immunity and deleterious anti-self pathology is achieved is of critical importance in improving immunotherapy strategies for treatment of a wide range of cancers. The molecular analysis of anti-tumor immunity in rare patients with paraneoplastic disease is therefore of great interest.

Both antigen reactivity in neuroimmunity (e.g., NMDA-receptor encephalitis¹³) and immune repertoires in solid tumors (e.g., metastatic breast cancer¹⁴) have been separately investigated. However, to date, few studies have linked molecular characterization of tumor and its immune infiltrate with the paraneoplastic autoimmune phenotypes of the same patients to permit elucidation of the immune process underlying paraneoplastic disease. Integrated analysis of paraneoplastic-disease-associated tumors offers a unique setting for the evaluation of systemic immune features driving both powerful anti-tumor immunity and often severe native tissue pathology.

Pediatric opsoclonus myoclonus ataxia syndrome (OMAS) is a rare but devastating autoimmune disorder characterized by sudden onset of uncontrollable, irregular, multivectorial eye movements, myoclonic jerking of the limbs, ataxia, and disordered mood/behavior in a previously well child.¹⁵ These prominent neurological symptoms, which frequently result in lifelong sequelae, often precipitate diagnosis of the underlying tumor. Pediatric OMAS is often associated with neuroblastoma, a solid tumor of the peripheral sympathetic nervous system that arises from the neural crest during development, but can also occur when no tumor is detectable. Most OMAS patients have localized, low-risk (LR) neuroblastoma disease, infrequent MYCN amplification (a strong negative prognostic determinant for neuroblastoma associated with low major histocompatibility complex [MHC] expression¹⁶), and may harbor other genomic copy number profiles that ordinarily accompany higher-risk tumors but are nevertheless favorably resolved.¹⁷ Importantly, as with other paraneoplastic diseases, patients with OMAS and neuroblastoma have better tumor outcomes than even LR neuroblastoma patients without OMAS.^18,19

Elegant single-cell longitudinal studies of adaptive immune response to tumor have been carried out in mouse models of high-risk (HR) neuroblastoma,²⁰ a system that offers outstanding opportunities for detailed mechanistic studies of a subset of neuroblastoma patients with poor outcomes and urgent, unmet therapeutic needs. There is currently no animal model of OMAS-associated neuroblastoma. However, a unique study of this paraneoplastic disease was initiated at multiple Children’s Oncology Group institutions starting over two decades ago, archiving material from these rare patients. In the current study leveraging these archival samples, single-cell/single-nucleus analyses were not possible to perform. Given the precious samples and their limitations, we have instead utilized a suite of highly complementary “classic” methodologies to analyze these rare paraneoplastic-disease-associated tumors. Here, we present a systematic study of OMAS-associated neuroblastoma tumors accrued on prospective Children’s Oncology Group (COG) clinical trial ANBL00P3²¹ to define the mechanisms of improved anti-tumor immunity as well as molecular correlates of the neuroimmune disease phenotype in neuroblastoma patients with OMAS.

RESULTS

Tumor gene expression profiling shows highly diverse tumor lymphoid infiltrate

To identify gene expression differences underlying differential anti-tumor immunity and neuroreactivity, we performed RNA sequencing (RNA-seq) on the 38 archival primary neuroblastoma samples from patients with OMAS treated on COG clinical trial ANBL00P3,²¹ with 13 LR and 13 HR (seven with MYCN amplification) neuroblastomas from age-matched patients without OMAS, obtained through the COG neuroblastoma biology study ANBL00B1 as comparators. Owing to the poor RNA quality of some of these archival samples, we used an exome capture RNA-seq protocol for this study.²² Differential expression analysis was consistent with significant lymphoid infiltrate in the OMAS tumors, as expected, but showed enrichment of memory B and T cells (Figures 1A and 1B; Tables S1 and S2). Among the most differentially expressed genes between OMAS neuroblastomas and LR non-OMAS neuroblastomas were CD22 and BANK1, both of which modulate B cell activity, and CCRL1, a regulator of immune and cancer cell migration. Notably, OMAS-associated neuroblastoma showed significant increased differential expression of TCF7, a marker of stem-cell-like CD8⁺ T cells and regulator of autoimmunity, in both comparisons (Table S2 and Figure S1; reviewed in Escobar et al.²³). In contrast, GLUD2, which has been reported to be an OMAS autoantigen,²⁴ was not significantly differentially expressed in our OMAS-associated neuroblastoma dataset (Figure 1A). Highly expressed outlier genes in OMAS compared to non-OMAS also included CR2, a complement receptor that is expressed on non-stromally derived follicular dendritic cells and on B cells, where it enhances binding of immune complexes by B cells and B cell receptor (BCR) signaling in autoimmunity.²⁵ In line with prominent B cell infiltration, we observed significant differential expression of B cell chemokine, CXCL13, and its receptor, CXCR5 (Figure 1A).

Figure 1. RNA-seq analysis highlights enhanced lymphocytic infiltration and activation in OMAS-associated neuroblastoma compared to control neuroblastoma.

(A) A volcano plot comparing expression (log2 fold change) of transcripts (as dots) in OMAS-associated neuroblastoma compared to non-OMAS neuroblastoma (OMAS, n = 38; non-OMAS, n = 26). The x axis indicates enrichment of expression in OMAS. Significance of differential expression (−log P value) is given on the y axis. Adjusted p < 0.05 indicated in red. Gene names in black are given for genes with expression differences of greater than log₂(2.25-fold) between groups. Gene names referred to in subsequent analyses labeled in light and dark blue and purple.

(B–D) CIBERSORT analysis of gene expression values from tumor RNA-seq permit estimates of immune cell fractions in OMAS vs. control neuroblastoma tumor infiltrate, including (B) memory B cells, (C) CD8⁺ T cells, and (D) follicular T helper cell fractions.

We used the R package coseq^26,27 to identify sets of coregulated genes from our differential expression analysis comparing OMAS to non-OMAS controls (Figure 1; Table S1). For each network, or “cluster,” we used GENIE3 to examine network connectivity and highlight the most connected hub genes, and we used ENRICHR²⁸ to identify the putative functions of each network as well as potential transcriptional regulators, based on reference chromatin immunoprecipitation sequencing (ChIP-seq) and transcription factor co-expression datasets. Networks of coregulated genes with higher expression in OMAS compared to control neuroblastoma included B and T cell signaling and differentiation networks (e.g., T helper 17 [Th17] cells, memory B cells; Figures 2A–2C). Two networks exhibited significantly lower expression in OMAS-associated neuroblastomas compared to controls (Figures 2D–2F). These include one network regulating extracellular matrix composition important for neural crest cell migration with putative immune modulatory properties (Dermatan sulfate, chondroitin sulfate) and one regulating mitosis, G2/M checkpoint, and early developmental cell identity.

Figure 2. Clustering and network analysis of differentially expressed genes between OMAS-associated neuroblastoma and control neuroblastoma.

Differentially expressed genes meeting fold-change and significance thresholds were used to identify sets of coregulated genes (clusters) using coseq, and further analyzed for connectivity using GENIE3. Network diagrams were plotted using Cytoscape. Line weight indicates number of connections. Genes accounting for 20% of all connections were designated as putative hub genes (red ovals).

(A–C) Gene clusters upregulated in OMAS-associated neuroblastoma relative to control neuroblastoma. (A and B) Network diagrams of two gene clusters upregulated in OMAS compared to non-OMAS samples.(C) Boxplot showing average expression profile for each patient subgroup in each upregulated cluster. Dots indicate mean values for one gene in the signature in that sample group. LR and HR comprise the non-OMAS subgroup from (A) and (B).

(D–F) Gene clusters downregulated in OMAS-associated neuroblastoma relative to control neuroblastoma. (D and E) Network diagrams of two gene clusters downregulated in OMAS compared to non-OMAS samples. (F) Boxplot showing average expression profile for each patient subgroup in each downregulated cluster. Dots indicate mean values for one gene in the signature in that sample group. LR and HR comprise the non-OMAS subgroup from (D) and (E).

OMAS-associated tumor gene expression reveals increased inflammation

We next explored immune landscape signatures derived from OMAS transcriptomes compared to non-OMAS tumor samples (Figures 3 and S2). On a granular level, a previously published tool²⁹ highlighted significantly higher mean expression of CD8, B cell score, cytotoxic lymphocyte immune signature (CLIS), T cell co-stimulatory molecules (CD28), markers of activation (CTLA4) and exhaustion (PD1) in OMAS neuroblastomas (Figure S2). These data are consistent with previously published reports of increased lymphocytic infiltration in OMAS tumors.^30–32 However, OMAS samples did not segregate into a single cluster among all neuroblastoma samples in the study but rather into several clusters bearing differing levels of expression of immune gene features, including several OMAS samples clustered with HR non-OMAS neuroblastoma samples exhibiting low immune-marker scores (Figure S2A). The lone HR, MYCN-amplified OMAS-associated neuroblastoma in the present cohort, PARSCY, did not appear in this cluster.

Figure 3. Cancer immune subtype classification identifies dominant immune signaling pathways in neuroblastoma patient cohort.

Immune subtype classifications were applied using normalized RNA-seq (log) expression levels for each patient, as previously described.³³

(A) Features of immune subtypes. Distinctive features of immune response correlated with each subtype C1–C6 based on meta-analysis of The Cancer Genome Atlas (TCGA) cancer dataset are indicated.

(B) Distribution of immune subtypes in OMAS and control neuroblastomas in this cohort.

(C) Enrichment of immune subtypes in OMAS relative to control neuroblastoma patient groups are indicated. Significant values are shaded in gray.

To more deeply probe the differences in the tumor microenvironment in OMAS patient samples compared to controls, we adapted a recently developed approach for classification of immune responses to cancer using RNA expression³³ (Figure 3A). This analysis builds on a classifier that predicts phenotypes, such as immune repertoire diversity and long-term survival, based on RNA-seq signatures that define immune subtypes. We used the classifier to bin samples in our cohort into immune subtypes C1–C6 (Figure 3A). We found that 50% of OMAS-associated neuroblastomas were classified as interferon gamma (IFNγ) dominant, a phenotype that is associated with strong CD8⁺ T cell signal and greater T cell receptor (TCR) diversity, while only 10%–15% of non-OMAS tumors had this classification (Figures 3B and 3C). Indeed, we observed an increased fraction of CD8⁺ T cells in OMAS tumors as estimated from RNA-seq data using CIBERSORT^34,35 (Figure 1B). In contrast, 50% of HR neuroblastomas were classified as “wound-healing dominant,” a phenotype that predicts high proliferative index and angiogenic gene expression, as well as Th2 cell bias and, importantly, a poor overall prognosis³³ (Figures 3B and 3C). We observed a small but significant increase in the proportion of OMAS tumors over LR non-OMAS tumors classified as C3, or “inflammatory” subtype, a classification that predicts lower levels of cell proliferation, aneuploidy, somatic copy number variation (sCNV), and superior outcomes (OMAS vs. non-OMAS; FDRq = 0.046). Both classifiers (Figures 3 and S2) highlight some heterogeneity within OMAS-associated neuroblastomas in their immune response to tumor in spite of their convergent clinical neuroimmune phenotypes.

Global gene expression profiling and clustering of OMAS vs. non-OMAS neuroblastomas appear to be driven by the degree and type of their immune infiltrate. Therefore, we also used a machine learning classifier, XGBoost,³⁶ to determine whether a distinguishing gene expression profile of OMAS-associated neuroblastoma could be identified. The algorithm was able to clearly distinguish OMAS from non-OMAS (auROC = 0.94; Figures S3A and S3D) and to distinguish OMAS from either HR (auROC = 0.69; Figures S3B and S3E) or LR neuroblastoma (auROC = 0.69; Figures S3C and S3F) to a lesser degree. It is noteworthy that the classification was driven by very few genes, as opposed to a broader gene expression signature. The top 10 features that were, on average, most important for the correct prediction of patient population (Figure S3G) included MRPS2, RMB15B, and MRPS17, encoding mitochondrial proteins. Lower expression of each of these genes drives the prediction toward OMAS (Figure 1A, dark blue; Figures S3A and S3G).

Although OMAS-associated neuroblastoma is expected to have similarly modest mutational load to typical non-OMAS neuroblastoma,^37,38 we investigated potential sources of neoepitope variation that could contribute to increased anti-tumor immunity by analysis of SNV burden using RNA-seq data. We identified 94 genes enriched for SNV variation that were significantly different between OMAS and non-OMAS samples, of which 47 genes are significant compared to HR alone (FDRq < 0.20; Table S3). However, we did not identify any single source of epitope variation in all patients that obviously underlies the observed immune response to OMAS-associated neuroblastomas.

Expression of genes encoding several CNS cell-surface proteins are correlated with OMAS disease severity

OMAS can present with neurological symptoms ranging from mild to severe and debilitating, and a semi-quantitative grading system has been devised.³⁹ We examined whether gene expression or immune features in the tumor correlated with neurological disease severity scores of OMAS collected at the time of diagnosis. Expression of two genes encoding neuronal cell surface receptors, the serotonin receptor, HTR6, and an alpha 2 adrenergic receptor, ADRA2C, correlated significantly with severity of OMAS neuroimmune symptoms (Figure S4). The gene NCAN, a CNS-specific extracellular matrix protein whose expression has been linked to malignant behavior of neuroblastoma,⁴⁰ and additional genes encoding cell surface proteins relevant for adhesion of neurons and leukocytes (DSCAML1, MAD-CAM1) also correlated significantly with OMAS neuroimmune symptoms.

MHC class II alleles distribution in OMAS-associated neuroblastoma

Susceptibility to many autoimmune diseases has been linked to genes encoded by the MHC (reviewed in Dendrou et al.⁴¹). We inferred HLA types from tumor-derived transcriptome data, using HLAprofiler, a published computational tool for HLA calling from RNA-seq data with >99% concordance with direct DNA sequencing.⁴² To establish background HLA allele frequencies in neuroblastoma, we inferred HLA types from a large set of neuroblastoma transcriptomes from the NCI-TARGET neuroblastoma dataset (n = 167)³⁸ using HLAprofiler and compared allele frequencies from our OMAS cohort (n = 38) to non-OMAS controls from this study (n = 26) and the non-redundant set of TARGET transcriptomes. The non-classical class II allele, HLA-DOB*01:01, was significantly enriched in OMAS (FDRq = 0.002; Tables 1 and S4). HLA-DO regulates MHC class II peptide loading and is almost exclusively expressed in B cells and in thymic epithelial medullary cells but not other professional antigen-presenting cells (APCs).⁴³ HLA-DOB enrichment and differential expression in OMAS tumors compared to non-OMAS tumors (Table S1) may reflect greater B cell infiltration of OMAS-associated tumors. Use of a less stringent FDR threshold of 0.2 to allow for discovery of additional alleles from our relatively small cohort of cases (with false discovery rate of 20%) allowed detection of HLA-DRB1*01:01 as being enriched in our OMAS cohort, consistent with a previous report⁴⁴ (HLA DRB1*01; FDRq = 0.18), as well as HLA-DRB*13:02 (FDRq = 0.16) and one MHC class I allele, HLA-C*04:01 (FDRq = 0.16). The most skewed HLA alleles we identified were two different alleles of the MHC class I pseudogene HLA-L. HLA-L is highly expressed in EBV transformed B cells; however, its functional significance is unknown.

Table 1.

Enrichment of HLA alleles in OMAS compared to control neuroblastoma patient groups

Allele model
HLA allele	OMAS frequency	LR frequency	HR frequency	Non-OMAS frequency	OMAS vs. LR FDR	OMAS vs. HR FDR	OMAS vs. non-OMAS FDR	LR vs. HR FDR
L*01:01	0.67	0.43	0.27	0.28	0.38	8.21E-08	5.90E-08	1
L*01:02	0.28	0.09	0.09	0.00	0.49	0.0056	0.0016	1
DOB*01:01	0.63	0.26	0.38	0.36	0.01	0.0118	0.0016	1
DRB1*13:02	0.12	0.04	0.04	0.03	1	0.28	0.16	1
C*04:01	0.21	0.11	0.09	0.09	1	0.19	0.16	1
DRB1*01:01	0.2	0.09	0.09	0.09	1	0.28	0.18	1
Population model
HLA allele	OMAS frequency	LR frequency	HR frequency	Non-OMAS frequency	OMAS vs. LR FDR	OMAS vs. HR FDR	OMAS vs. non-OMAS FDR	LR vs. HR FDR
L*01:01	0.89	0.44	0.35	0.35	0.04	1.79E-07	5.69E-08	1
L*01:02	0.50	0.11	0.16	0.16	0.09	0.0064	0.0015	1
DOB*01:01	0.71	0.33	0.41	0.40	0.21	0.10	0.04	1
DRB1*01:01	0.39	0.19	0.16	0.16	1	0.14	0.07	1
DRB1*13:02	0.24	0.07	0.07	0.06	1	0.18	0.10	1
DQB1*05:01	0.50	0.33	0.25	0.25	1	0.18	0.10	1
DQB1*06:04	0.16	0.04	0.03	0.03	1	0.18	0.10	1
DQA1*01:01	0.42	0.22	0.20	0.19	1	0.22	0.13	1
DQA1*05:01	0.42	0.26	0.21	0.21	1	0.22	0.15	1

Two different models were used to test for enrichment of HLA alleles that may contribute to OMAS autoimmunity. (A) Allele model. This model assesses occurrence of each HLA allele in the pool of total alleles found in patients of one subtype compared to another subtype. Allele frequency calculated as # of observed alleles/total number of alleles in that population pool (2 × # samples). (B) Population model. This model for enrichment tests for each HLA allele in patients from each population compared to another. Here, the number of patients containing the allele, regardless of copy number, is compared to the total number of patients in the pool. The total sample size for each population = the number of patients; patients homozygous for the allele are counted only once. Note that frequencies were computed based on tumor gene expression levels, using an algorithm previously shown to have high concordance to DNA-based measurements.

Tumor-infiltrating T cells exhibit greater antigen receptor diversity in OMAS-associated neuroblastoma

A link between the OMAS autoimmune response and improved anti-tumor immunity would predict that the repertoires of tumor-infiltrating T cells and B cells would be strongly shaped by OMAS causative antigen(s). We hypothesized that the OMAS tumor lymphocytic infiltrate would be predominantly oligoclonal. We used genomic DNA from tumors to sequence their TCRβ and immunoglobulin heavy chain (IgH) repertoires,^45,46 and we analyzed lymphocyte repertoires from 31 OMAS samples and 13 LR and 13 HR control samples. We analyzed in-frame sequences corresponding to the TCRβ and IgH CDR3 regions, which provide most of the antigen-binding specificity to the receptor and therefore are used as a proxy for antigen specificity of each receptor type in this analysis. OMAS-associated neuroblastoma TCR repertoires were significantly larger than those recovered from HR neuroblastoma samples (Figure 4A; FDRq = 0.001) and 2-fold larger than LR neuroblastoma samples (Figure 4A; FDRq = 0.075). These T cell number estimates based on genomic DNA sequencing of TCRβ repertoires are consistent with transcriptome-based estimates of elevated T cell numbers in OMAS samples (Figures S5A–S5D).

Figure 4. OMAS tumor-infiltrating TCR analysis reveals significant diversity and small clones with limited similarity.

For (A)–(E), sample numbers are OMAS, n = 29; LR, n = 11; HR, n = 9.

(A) Repertoire size. Number of TCRβ sequences in each patient repertoire, normalized for input DNA amount.

(B) Shannon diversity index of OMAS-associated and non-OMAS-associated neuroblastoma TCRβ repertoires. Repertoires were subsampled to 1,382 sequences and Shannon index computed. Average over 100 iterations plotted for each patient. Median value indicated in red.

(C) Gini index of inequality of OMAS-associated and non-OMAS neuroblastoma TCRβ repertoires. Average over 100 iterations plotted for each patient. Median value indicated in red.

(D) Sums of clonal frequencies for top 100 clones of TCRβ repertoires. Cumulative individual frequencies of top 100 clones in each patient repertoire were summed and plotted as a single point. Median value in each patient subgroup indicated in red.

(E) PCA plot of Hill values for TCR repertoires. Hill values were used to describe the diversity of the TCR repertoires. The PCA plot of all patient samples depicts two clusters according to their top two principal components. One cluster composed of OMAS samples is roughly indicated with a purple oval (cluster 1), while another cluster, composed of samples from all groups, is indicated by an orange oval (cluster 2).

(F) Sequence sharing in OMAS and non-OMAS patient TCRβ repertoires. Each dot represents the fraction of sequences in the given sharing level, normalized by the number of samples in each group. The figure is in log10-log10 scale. Sample numbers for sharing level calculations: OMAS, 31; LR, 13; HR, 13.

We next evaluated the diversity and clonality of the TCRβ repertoires. To minimize the effect of sample size on diversity estimates, we subsampled all repertoires to a common size (reducing the analysis to 49 samples out of 57 total). We then computed Shannon diversity index (a measure for diversity) and Gini index (a measure for clonal inequality) for each sample, averaging over 100 iterations of subsampling. We found that OMAS repertoires are significantly more diverse than HR (Figure 4B; FDRq = 0.014) and more diverse than LR (FDRq = 0.053), while the latter non-OMAS cohorts were similarly diverse (FDRq = 0.456). The higher diversity of OMAS tumor-infiltrating lymphocytes (TILs) is in line with the prediction of increased TCR diversity for tumors of immune classifier subtype C2,³³ which is dominant among our OMAS samples (Figures 3B and 3C).

TCR repertoires within OMAS samples had significantly lower Gini indices, a measure of clonal inequality, than non-OMAS neuroblastoma samples (Figure 4C), indicating more even distribution of clone sizes without considerable expansion. In accordance with their Gini indices, we found that the summed frequencies of the top clones were also significantly lower in OMAS compared to either LR or HR non-OMAS samples (Figures 4D, and S5E). Diversity may be measured in a variety of ways, depending on the weight given to the abundance of any individual sequence, using measurements called “Renyi values.” Therefore, we examined repertoire diversity over a range of Renyi (Hill) values and found that OMAS repertoires are robustly and consistently more diverse than non-OMAS repertoires (Figure S5F). Together, these results invalidated our original prediction of oligoclonality in TIL repertoires and instead support the notion that OMAS-associated neuroblastomas harbor diverse, polyclonal repertoires of T cells.

TCRβ repertoires from OMAS patients exhibit limited sharing of TCR CDR3γ sequences

We then compared similarity of tumor-infiltrating TCR repertoires from patients with and without OMAS, using TCRdist,⁴⁷ an algorithm that scores occurrence of a TCR in different repertoires within a specified distance threshold of permitted substitutions or gaps, with concomitant scoring penalties, and assesses overlap of clusters of similar TCRs with a specified cohort. TCRdist did not return any significant similarity of shared, cohort-specific sequences (Table S5).

We then took a different approach to assess the characteristics of the TCR repertoires by using Hill numbers, a generalization of the classic diversity measures (as in Greiff et al.⁴⁸), for each repertoire as a basis for principal-component analysis (PCA; Figure 4E). To again minimize the bias of the larger repertoire sizes of OMAS samples, Hill numbers were calculated after subsampling repertoires to a common size (1,382 sequences, which reduced the total cohort to 49 total samples). Average Hill numbers for each repertoire were used to plot samples according to their two principal components, which explained 98% of the variance, and resulted in the segregation of two clusters. One of these was composed exclusively of OMAS samples (Figure 4E, purple oval), while the other, larger cluster included a mix of OMAS, LR, and HR samples (orange oval). This indicated that at least a subset of OMAS TCR repertoires are more similar to each other than to the other (mostly) non-OMAS repertoires based on their diversity. Nevertheless, plotting the histogram of sharing for the two groups supports only nominally greater similarity between OMAS samples (Figure 4F) than non-OMAS samples. The OMAS distribution lies uniformly above the control sharing distribution, suggesting that any limited sharing between OMAS repertoires is likely driven by a small number of TCR sequences.

We noted that most of the highly shared TCRβ sequences in OMAS repertoires, as well as in non-OMAS neuroblastoma repertoires, are also highly shared in peripheral blood mononuclear cells (PBMCs) of healthy donors (found in >75% of 786 repertoires reported in Emerson et al.⁴⁹; Figures S5G and S5H; Table S6, public). A subset of OMAS-associated shared TCRs that are less shared among non-OMAS neuroblastoma patients in our cohort (OMAS overshared) and another subset that are enriched in non-OMAS neuroblastoma (control overshared) are summarized in Table S6. While their specificity may still be unknown, some shared enriched TCRs in different patient subgroups have been previously reported in other disease contexts, which may yield additional insights from the literature. Altogether, our results suggest that, as a class, OMAS-associated neuroblastomas are characterized by limited sequence sharing, smaller clone size, and significantly greater diversity, a feature that seems to define a further, distinct subset of OMAS patients whose clinical disease correlates are still not known.

Diversity of B cell IgH repertoire is associated with improved tumor-related outcomes in OMAS

B cell infiltration of solid cancers generally has positive prognostic value, and yet the role of B cell infiltration of solid tumors is far less well understood than that of CD8⁺ T cells (reviewed in Nelson⁵⁰). In contrast, the central role for B cells in OMAS neuropathology is underscored by the efficacy of the anti-CD20 antibody rituximab in mitigating neurological symptoms in OMAS.^51,52 Given the significant B cell infiltrate evident from tumor RNA-seq, we predicted an oligoclonal response that would be evident in analysis of IgH repertoires from OMAS-associated tumors.

As with TCRs, OMAS-associated neuroblastomas had larger BCR repertoires than either HR (FDRq = 0.01) or LR (FDRq = 0.12); non-OMAS neuroblastoma repertoire sizes were not significantly different in size (HR-LR, FDRq = 0.46) (Figures 5A and S6A). As for TCRβ, we calculated the Shannon diversity index for all IgH repertoires after subsampling to a common size. We found that OMAS BCR repertoires were significantly more diverse than HR control neuroblastomas (Figure 5B). Shared clinical features of OMAS may be associated with dominance of a few large clones responding to the OMAS antigen(s) in the CNS compartment, which we predicted would also be represented in OMAS tumors. We therefore investigated the clonal structure of OMAS tumor repertoires. LR and HR tumors both possessed larger clones than patients with OMAS (Figure 5C; OMAS-HR, FDRq = 0.031; OMAS-LR, FDRq = 0.171; LR-HR, FDRq = 0.171). We also looked for differences in VH or JH gene or gene family usage or for differences in CDR3 length in OMAS. However, only very low-frequency events were detected as significant (Figure S6).

Figure 5. IgH repertoire analysis of TILs reveals greater diversity and reduced clonality of OMAS-associated neuroblastoma BCR repertoires.

For this analysis, sample numbers passing minimum repertoire size cutoff were OMAS, n = 20; HR, n = 5; LR, n = 8.

(A) Repertoire size. Number of IgH sequences in each patient repertoire, normalized for input DNA amount.

(B) Shannon diversity index of OMAS- and non-OMAS-associated neuroblastoma IgH repertoires. Mean index value after 100 iterations of subsampling and index calculation is plotted as one point for each patient. Red line indicates median for each patient group.

(C) Clone size of OMAS- and non-OMAS-associated neuroblastoma TIL repertoires. Summed frequency of top 100 clones in each patient is given as a point. Red line indicates median value for each patient group.

(D) IgH clusters enriched in OMAS. Clusters of IgH sequences with at least 85% sequence similarity and comprising at least seven OMAS patients and not more than two LR or HR patients are shown, with V family, J family junction length, and cluster index indicated.

OMAS-enriched clones exhibit similar sequence features

Owing to the uneven sizes of the OMAS and control repertoires and to the small repertoire sizes for all samples, we were unable to test whether clones observed only in OMAS repertoires are truly OMAS specific. Figure 5D highlights clusters of sequences possessing 85% amino acid sequence similarity and shared by at least seven OMAS patients, grouped by VH and JH gene usage and junction length. Several sequences were not observed at all in HR patients in this study; many were also only shared by a single LR patient. We also characterized numbers of somatic mutations in IgH V genes as a marker of somatic hypermutation in B cell clones. Increased numbers of mutations would be acquired in mature germinal center (GC) B cells and are used as a proxy for B cell clonal selection. We detected a few significant increases in somatic mutation frequency in the IGHV genes in OMAS compared to LR or HR (Figure S6F, stars). However, we cannot infer any biological relevance of these mutation rates from the limited current cohort.

Taken together, the significantly greater B cell infiltration in OMAS tumors was characterized by a paucity of large clonal expansions. The B cell infiltrates were significantly more polyclonal in OMAS compared to control neuroblastoma samples. This diversity, as well as our limited number of control samples and their small repertoire sizes, precluded nomination of any specific BCR clone or sequence as specifically correlated with OMAS or anti-tumor immunity.

OMAS tumors contain tertiary lymphoid structures and exhibit apparent neuronal localization of TILs

Histological examination revealed numerous tertiary lymphoid structures (TLSs) in 10 of 14 (71%) OMAS tumors available for evaluation and scoring⁵³ (Figures 6A–6B′ and S7; Table S7) and were usually accompanied by widespread interstitial lymphocyte infiltration (Figure 6C). In contrast, two of six (33%) non-OMAS LR neuroblastomas (Figures 6C–6D′) exhibited similar structures with significant but somewhat lesser interstitial infiltration (Figure 6E); only one of five (20%) non-OMAS HR neuroblastomas contained TLSs (Figures 6G–6H′), and these had the least interstitial infiltration among groups (Figure 6I). The TLSs contained dense cores of CD20⁺ B cells surrounded by CD3⁺ T cells (Figures 6J–6J″) and were easily distinguished from neighboring tissue by morphology using differential interference contrast (DIC) or bright-field microscopy. Using an antibody against Ki67, a marker of cell proliferation, we observed relatively few Ki67-positive cells within TLSs in OMAS tumors (Figure S7). We also noted localization of B cells and T cells to putative neuronal processes within small patches of what appear morphologically to be differentiating ganglia in OMAS tumors (Figure S7C; Table S7). This often included B cells at the center with T cells enriched nearby.

Figure 6. Lymphocyte enrichment and localization to TLSs in OMAS-associated neuroblastomas.

(A–I) H&E staining of TLS and GC-like (GC-L) TLS, highlighting morphologically different features of TLS and GC-L. For (A)–(I), all images shown are at 200× magnification (scale bar, unit measure of 50 μm). (A–C) TLS in OMAS-associated neuroblastoma. (A–B′) GC-L structures in OMAS-associated neuroblastoma characterized by a core of larger lymphocytes resembling centrocytes in GCs of lymph nodes. (C) TLSs in OMAS-associated neuroblastoma are easily distinguishable from tumor cell nests (tumor). Note the overall marked density of small lymphocytes throughout the tissue in (A)–(B′), and in the lower-magnification image in (C) highlighting the extensive lymphocytic infiltrate both in the context of lymphoid structures and throughout the tumoral stroma. (D) GC-L TLS, and (D′) TLS in an LR neuroblastoma. (E and E′) TLS in another LR neuroblastoma mass, with uniformly small lymphocytes clustered within the tumor tissue. (F) Lower-magnification image of section in (D) and (D′) highlighting the context of significant lymphocytic infiltration (both as TLSs and diffuse) but less dense than in OMAS-associated neuroblastomas (compare to C). (G) GC-L TLS in HR neuroblastoma sample, characterized by a core of lymphocytes resembling centrocytes in GCs of lymph nodes, surrounded by tightly packed small lymphocytes with scarce cytoplasm and hyperchromatic (strongly hematoxylin-positive) nuclei. (G′) A TLS from the same tumor, characterized only by tightly packed small lymphocytes with scarce cytoplasm and hyperchromatic nuclei. (H and H′) Two TLSs from a different HR neuroblastoma mass at high magnification. (I) TLS from the HR neuroblastoma in (H) and (H′) at lower magnification, highlighting tumor tissue with only scattered infiltrating lymphocytes outside the TLS.

(J–J″) Representative TLS in OMAS-associated neuroblastoma, containing (J) B cells (anti-CD20⁺; red), (J′) T cells (anti-CD3⁺; green). (J″) Merge of green and red channels. Images are taken at 20× magnification; scale bar, 32.2 μm.

DISCUSSION

Here, we sought to understand the underlying mechanisms of neuroblastoma-associated autoimmunity with a characterization of tumors from patients enrolled in the only prospective OMAS clinical trial reported to date.²¹ While the study involved very limited cohort size owing to the rarity of the disease and the scarcity of samples, we were able to include a variety of experimental methodologies in our sample analysis, leading to well-supported insights. To our surprise, we found that the robust immune cell infiltrate is dominated by polyclonal B and T cells, absent the identification of a unifying single antigenic stimulus, as has been seen in other paraneoplastic diseases (e.g., NMDA receptor encephalitis^4,54). We confirmed a major role for auto-reactive B cells in neuroblastoma-associated OMAS, and here highlight a major role for T cells in anti-tumor reactivity and likely neuropathology, importantly, in the context of TLSs. We also identify an MHC class II allele, HLA-DOB*01:01, as significantly enriched in OMAS tumors compared to neuroblastoma controls.

In this work, we compared OMAS to non-OMAS neuroblastoma, with additional contrast of OMAS vs. LR neuroblastoma, to highlight the influence of paraneoplastic autoimmunity on superior anti-tumor reactivity and to pinpoint foci of OMAS neuroimmune targeting. While no clear, single molecular target of neuroimmunity emerged, we identified several pathways enriched in OMAS and several genes whose expression correlates with neurological symptom severity. Validation of candidate molecular antigens and pathways will be essential to understanding the clinical consequences of OMAS neuroimmunity.

We identify four conspicuous differences between OMAS- and non-OMAS-associated neuroblastomas, also remarkable in OMAS vs. LR neuroblastoma, which align with reported signatures from solid tumor literature as having positive prognostic value. These same features accompany tissue infiltrates in human autoimmune disease, supporting their relevance for CNS tissue pathology in OMAS and supporting their centrality in a systemic OMAS disease process. These are (1) increased numbers and activation of B cells in tumor infiltrate, rich in memory B cells; (2) localization of B cell infiltrate to TLSs rich in T cells; (3) polyclonality of lymphocytic infiltrate; and (4) differential expression of TCF7, CXCR5, and CXCL13. These features accompany significant TCR and BCR diversity in OMAS tumors compared to controls, which is a defining feature of OMAS-associated neuroblastoma, but one whose relevance to disease outcomes is less clear. Combining these observed differences with insights from both cancer and autoimmunity, we propose a framework to explain how systemic autoimmunity drives superior tumor outcomes and neurological damage in OMAS.

For anti-tumor immunity, it is striking that the same defining features of OMAS mirror the immune characteristics of tumors from other cancers with positive response to immune checkpoint blockade, including another neural crest-derived cancer, melanoma.^55,56 While CD8⁺ T cells are considered the workhorses of tumor destruction, OMAS tumors exhibit the greatest differences in B cell numbers, exceeding even LR neuroblastomas, which also have excellent outcomes; OMAS tumor transcriptional profiles suggest enrichment of memory B cells, and histopathological evaluation finds that OMAS neuroblastomas contain more tertiary lymphoid structures (Table S7^,31,32). The presence of TLSs in tumors has been identified as a strongly predictive prognostic factor for positive tumor outcomes across cancer types^57,58 and has been noted after successful cancer immunotherapies (reviewed in Sautès-Fridman et al. and Trüb and Zippelius^59,60). While it is not known what drives TLS formation, we observe differential expression of B cell chemokine CXCL13 and its receptor CXCR5 in OMAS tumors compared to non-OMAS (Table S1), two features correlated strongly with ectopic lymphoid structure formation in a variety of settings in both cancer and autoimmunity (reviewed in Kazanietz et al.⁶¹). A TCF7+ T cell subset has independently been identified as enriched in TLSs of an oral solid tumor and predictive of positive tumor outcomes.⁶² Consistent with both TLS enrichment and superior outcomes in OMAS-associated tumors, TCF7 is strongly differentially expressed in OMAS tumors. The signature of TCF7, CXCR5/CXCL13 in B cell-rich TLSs, was found as a predictor of survival in melanoma independently of all other variables,⁵⁵ underscoring their importance for outstanding tumor outcomes across cancer types. The extreme diversity and significant polyclonality of OMAS lymphocytic infiltrate is not easily or universally aligned with solid tumor outcomes in other cancers, where diversity and clonality may accompany either positive or negative outcomes. However, increased diversity of TIL TCR repertoires has been suggested as an indication of durable pre-treatment tumor control in a variety of cancers, whereas pre-treatment clonality predicted improved response to PD1 blockade therapy.⁶³ This diverse infiltrate in OMAS-associated tumor may contain a high proportion of “bystanders,” whose specificities include T cells cross-reactive for non-tumor-specific antigens, such as cytomegalovirus (CMV) or Epstein-Barr virus antigens, that are recruited to the tumor for tumor control (e.g., Chiou et al., Simoni et al.^64,65). This view is supported by the observation of highly shared “public” TCRs in OMAS-associated tumor infiltrate with known specificity for viral antigens from CMV and influenza (Table S6, “OMAS highly shared public”) and others observed in neuroimmune diseases with radically different phenotypes (e.g., Rasmussen encephalitis^66–68). Taken together, we therefore identify TLS with diverse polyclonal lymphocytic infiltrate, and strong expression of B cell chemokines and TCF7, as the signature of paraneoplastic autoimmunity most prominently associated with the superior tumor outcomes of OMAS.

These same features of OMAS-associated neuroblastoma have also been noted in pathological tissue infiltrates in human autoimmune disease (reviewed in Jones and Jones⁶⁹). B cells and their trafficking to sites of inflammatory cytotoxicity are emerging as central to disease severity in autoimmunity as well. In autoimmune encephalitis caused by multiple sclerosis (MS), B cell follicles and B cell chemokine CXCL13 expression are enriched at brain lesions associated with severe, progressive disease presentation,⁷⁰ while loss of CXCL13 in a mouse MS model mitigates severe disease phenotypes.⁷¹ Similarly, high levels of CXCL13 have been found in inflamed synovia of patients with severe rheumatoid arthritis (RA⁷²), while loss of CXCR5 in mouse models of RA reduced joint damage and impaired TLS formation.^73,74 In previous studies of OMAS, high levels of CXCR5 and CXCL13 were noted in cerebrospinal fluid (CSF) of patients with OMAS, correlated with increased disease severity.⁷⁵ This chemokine/receptor pair mediates migration of B cells, which we now link to trafficking both to tumor and CNS in OMAS. The presence of tertiary lymphoid structures accompanies disease severity and target tissue damage in a range of autoimmune diseases (reviewed in Pipi et al.⁷⁶) and predicts similar pathology in the CNS of OMAS patients, although TLSs in the brains of living OMAS patients cannot be investigated.

Tertiary lymphoid follicles are sites of antigen presentation that arise in peripheral tissues upon chronic inflammatory stimulation that often accompanies autoimmunity or infection (reviewed in Sautès-Fridman et al. and Trüb and Zippelius^59,60). They support memory B cell formation and auto-reactive T and B cell activation and can also lead to production of high-affinity antibodies, via plasma cell differentiation. GCs are those TLSs with mature plasmablasts that have undergone somatic hypermutation to produce high-affinity, presumably cytotoxic, antibodies.⁷⁷ In OMAS-associated neuroblastomas, we identify TLSs and memory B cell enrichment as well as B cell follicles rich in T cells. However, Ki67, a histological marker of proliferation often associated with clonal expansions of antibody-rearranged B cells, was largely absent from these structures in our cohort (Figures 6 and S7). Furthermore, in our data, we observe an absence of dominant species of expanded B cell clones in IgH repertoire analysis. We also note the absence of strong BLIMP1 expression, a marker of GCs, along-side strong differential expression of CD22, a B cell marker that is not expressed in mature plasma cells. Together, these findings could suggest either that OMAS-associated neuroblastomas contain many TLSs that are not GCs, or that we have observed a snapshot of TLS maturation that precedes a complete GC reaction and, perhaps, that the antibody function of OMAS B cells may not be its essential one.

We propose that the critical function of B cells in OMAS tumor and CNS immunity is not only the production of pathogenic antibodies but as potent APCs in long-lived TLSs. In the context of neuroimmunity, B cells function crucially as APCs in a lupus-prone mouse model⁷⁸ and in the experimental autoimmune encephalomyelitis (EAE) murine model of MS.⁷⁹ EAE model mice expressing the MOG-specific BCR but unable to secrete antibodies are fully susceptible to EAE induction by myelin oligodendrocyte glycoprotein (MOG) in an MHC class II-dependent manner.⁷⁹ Antigen-experienced B cells of animals with autoimmunity function as APCs and may spontaneously drive TLS formation. These interactions result in CNS targeting and T cell-mediated cytotoxicity in both neuroimmune disease models and human patients, resulting in neuropathology. Further support for B cell function as APCs in OMAS comes from the increased frequency in OMAS of HLA-DOB*01, an HLA allele expressed predominantly in B cells that modulates presentation of immunodominant epitopes (reviewed in Jiang et al. and Welsh and Sadegh-Nasseri^80,81). Finally, the observation of B cell-trafficking, TLS-promoting chemokines in OMAS supports the central role of B cells in TLS prevalence and B cell-T cell interactions accompanying both positive tumor outcomes and neuropathology.

If indeed a single mechanism underlies both CNS pathology and anti-tumor immunity in OMAS neuroblastoma patients, then OMAS tumors (and, indeed, tumors of other paraneoplastic diseases associated with neuroimmunity) may offer a system in which to study the cellular basis of neuronal damage in the CNS, which cannot be addressed in living patients. It is still unclear whether the observed diversity and polyclonality of tumor infiltrate in OMAS arises because of lymphocytic influx from the periphery, which would be consistent with the dominance of public TCRs in tumor with similar representation to their frequencies in peripheral blood. Specific predictions made in the current study, such as the properties of OMAS-associated TLS B cells and selected T cells in in tumor control and the putative role of auto-reactive T cells in brain neuropathology in OMAS, should be addressed in future work using freshly isolated and cell-sorted CSF and tumor samples and in humanized mouse models. Our work supports renewed focus on antigen-presenting B cells as potentiators of cancer immunotherapy through generation of long-lasting TLSs to promote tumor destruction. Modulation of accompanying autoimmunity will be a critical bottleneck for clinical applications.

Limitations of the study

This study has several important limitations that should be considered in evaluating our observations and conclusions. First, OMAS is a very rare disease for which patient samples are quite limited. Therefore, cohort sizes presented here are small compared to cohort studies for more common cancers. Second, this study was carried out exclusively on archival samples collected as part of a clinical trial that enrolled patients over a decade, so single-cell methodologies and techniques requiring fresh tissue sections, including both transcriptomics and immune repertoire analyses with paired receptor information, were not possible to employ. Third, we were able to analyze somewhat smaller cohorts of each control cohort (LR, HR) than the OMAS cohort, leading to some mismatch in statistical comparisons among three groups. This was only partially mitigated in OMAS vs. non-OMAS comparisons, which have closer but still unequal group sizes (OMAS = 38 vs. non-OMAS = 26).

STAR★METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Miriam I. Rosenberg (miriamirosenberg@gmail.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• DATA. RNA-seq, TCR-seq and BCR-seq data have been deposited at GEO (https://www.ncbi.nlm.nih.gov/geo/) and are publicly available as of the date of publication. Accession numbers are listed in the key resources table.

• CODE. This paper does not report original code.

• ADDITIONAL INFORMATION. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit anti-human CD3	DAKO	A0452
Mouse anti-human CD20	DAKO	M0755
Goat anti-human CD4	R&D Systems	AF-379-NA; RRID:AB_354469
Rabbit anti-human CD8	Thermo	RB-9009-P0
Alexa 488 anti-Rabbit	Life Technologies	A21206
Alexa 594 Anti-Mouse	Life Technologies	A11032
Alexa 488 anti-Goat	Life Technologies	A11055
Alexa 594 anti-Rabbit	Life Technologies	A21207
DAPI Hydrochloride	Sigma	32670
anti-Ki-67:Alexa Fluor 647 direct conjugate	Abcam	ab196907
Biological samples
Tumor RNA	Children’s Oncology Group	N/A
Tumor genomic DNA	Children’s Oncology Group	N/A
Tumor FFPE sections	Children’s Oncology Group	N/A
Chemicals, peptides, and recombinant proteins
Antigen unmasking solution	Vector Laboratories	H-3300
Deposited data
Neuroblastoma tumor transcriptomes	TARGET-NBL Dataset, NCBI	https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000467.v17.p7
TCRbeta reference dataset, healthy and CMV adults (786 individuals)	Emerson, R et al.49	https://www.adaptivebiotech.com/immuneACCESS DOI https://doi.org/10.21417/B7001Z
Neuroblastoma tumor transcriptomes, this paper	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE189367	GSE189367
Neuroblastoma TCRbeta repertoires, this paper (Adaptive)	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE189742	GSE189742
Neuroblastoma IgH repertoires, this paper (Adaptive)	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE189741	GSE189741
Software and algorithms
R v4.1	Hornik and R Core Team (2022), “The R FAQ”. https://CRAN.R-project.org/doc/FAQ/R-FAQ.html	https://www.R-project.org/
python scikit-learn (version 0.18.2)		https://scikit-learn.org/0.18/install.html
HLAProfiler	Buchkovich, M et al.42	https://expressionanalysis.github.io/HLAProfiler/
XGboost	Chen T and Guestrin C.36	https://github.com/dmlc/xgboost
SHAP	Lundberg S and Lee S-I82	https://github.com/slundberg/shap
Alakazam software	Gupta NT, et al.83	https://github.com/cran/alakazam/blob/master/DESCRIPTION
Change-O	Gupta NT, et al.83	http://clip.med.yale.edu/changeo/
IgBlast	Ye, J et al.84	https://ncbi.github.io/igblast/
Shazam	Gupta NT, et al.83	https://shazam.readthedocs.io/en/stable/
Immune Subtype Classifier	Thorsson, V et al.,33 Gibbs DL85	https://github.com/Gibbsdavidl/ImmuneSubtypeClassifie
RAbHIT package	Peres A, et al.86	https://cran.r-project.org/web/packages/rabhit/
coseq	Rau, A. and Maugis-Rabusseau, C.26	https://www.bioconductor.org/packages/devel/bioc/vignettes/coseq/inst/doc/coseq.html
GENIE3	Huynh-Thu VA et al.87	https://www.bioconductor.org/packages/release/bioc/html/GENIE3.html
cytoscape	Shannon, P et al.88	https://cytoscape.org/
limma4	Ritchie, ME et al.89	http://bioconductor.org/packages/release/bioc/html/limma.html
DESeq2	Anders, S and Huber, W90	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
edgeR	Robinson, MD et al.91	https://bioconductor.org/packages/release/bioc/html/edgeR.html
EBseq	Leng, N et al.92	https://bioconductor.org/packages/release/bioc/html/EBSeq.html

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

All subjects were enrolled on the COG Neuroblastoma Biology Protocol ANBL00B1 and samples banked centrally. All OMAS patients were also enrolled on the ANBL00P3 clinical trial (REF). Access to samples for this study was approved by the Children’s Oncology Group Neuroblastoma Committee under the auspices of the NCI Cancer Therapy Evaluation Program (CTEP). Anonymized patient data and tumor characteristics are provided in Table S7, sheet 2.

METHOD DETAILS

Patient tumor samples

We retrospectively procured all primary tumor samples (N = 38) available from the COG ANBL00P3 clinical trial, in which the efficacy of IV immunoglobulins (Ivig) in neuroblastoma patients with OMAS was tested.²¹ All patients enrolled were <8 years old with biopsy-proven, newly diagnosed neuroblastoma and OMAS. Samples collected from each patient included tumor tissue, cerebrospinal fluid (CSF) and blood sera from time of diagnosis. We also sequenced 13 patients with low-risk (LR) and 13 patients with high-risk (HR) non-OMAS neuroblastomas as comparators. We obtained reliable data from all samples, using the Illumina RNA Access platform, an exon capture kit designed to salvage usable data from low-quality RNA samples. However, as a consequence of using this platform, our ability to harmonize our data with existing neuroblastoma RNA-seq datasets (using other platforms) was rather limited.

RNA sequencing

Patient tumor RNA was sequenced with 2 × 150 bp, paired end sequencing, using the TruSeq RNA Access kit from Illumina (now called TruSeq RNA Exome; Qiagen, Valencia CA, USA) and quantified on a NanoDrop spectrophotometer. RNA purity and integrity was assessed by Agilent 4200 Tapestation. RNA integrity (RIN) scores for the samples varied from 1 to 7.9, though all samples had DV200 values of >30%. Sequencing libraries were prepared from 100ng total RNA from each sample and were run on high output flowcells on an Illumina NextSeq 500, yielding an average of 30M reads per sample. Paired-end sequence reads were analyzed according to currently available best practices for whole-transcriptome analysis, as described below.

RNA-seq data analysis

Raw FASTQs from both the OMAS/LR/HR cohort and NCI TARGET³⁵ datasets were processed using fastq-mcf (https://github.com/ExpressionAnalysis/ea-utils/blob/wiki/FastqMcf.md):

http://expressionanalysis.github.io/ea-utils/; parameters: –max-ns 4 –qual-mean 25 -H -p 5 -q 7 -L 25). Clipping completely removed reads with large homopolymers, overall low base quality scores or less than 25 nucleotides and removes low quality bases at the end of the sequence and adapters. These clipped reads were aligned to the human reference genome hg19 using STAR v2.4⁹³ and a UCSC reference transcriptome supplemented with lincRNAs from Ensembl. RSEM v1.2.14 (https://github.com/ExpressionAnalysis/STAR-SEQR; evaluated in⁹⁴) was used for both gene and isoform quantification. RNA fusion events were detected using STAR-SEQR v0.6.5 parameters: -m 1).

Differential expression analysis was performed using Q2 Solutions’ ensemble two group comparisons suite. This method summarizes the differential expression p values and classification probabilities from five tools—t-test, limma4,⁸⁹ DESeq2,⁹⁰ edgeR⁹¹ and EBSeq⁹²—to produce a new p value for differential expression. For any given gene, the p values of each constituent model are input into a logistic regression model, which estimates the probability that the gene is differentially expressed. This probability is transformed into a p value for differential expression by comparing it against its empirical cumulative distribution as estimated by bootstrap resampling of TCGA data from various cancer types.

Clustering and network analysis of differentially expressed genes

Genes that were differentially expressed between OMAS and non-OMAS patients, with cutoff of 2-fold difference in gene expression between groups and FDRq<0.05 were used for clustering and network analyses. The matrix of the normalized expression level these genes in 64 patient tumors sequenced in this study were used as the input for gene clustering analysis. Gene clustering was performed using the R package “coseq”,^26,27 using Gaussian mixture model (“Gaussian_pk_Lk_Bk” under the “Normal” model), arcsin transformation of expression values, and an expression value cutoff of 50. Eight clusters were generated because K = 8 minimized ICL (integrated classification likelihood), with a series of trials with values of K between 2 and 25. Networks of genes from each of the 8 clusters were generated based on their normalized expression profiles, using the default settings of R package, “GENIE3”.^87,95 The top 20% of weighted connections among genes of the GENIE3 output were used to visualize the network of genes in each cluster, using Cytoscape.⁸⁸ The mean values for each gene within a cluster were obtained for samples within each patient group and depicted as a boxplot where each point represents a mean expression value for one gene in the cluster. Groups were compared using Wilcoxon test and p values were adjusted for multiple comparisons.

HLA typing

HLA types were identified in both OMAS/LR/HR and TARGET datasets using the default parameters of HLAProfiler⁴² and each allele tested for enrichment. For some genes, HLAProfiler identified alleles in less than 25% of samples. Alleles from these genes or alleles identified in only a single sample were excluded from the enrichment analysis. Significance of enrichment was determined by testing the distribution of each allele among patient subgroups compared to all alleles for the gene in the population (n = 2*number of samples). Fisher’s exact test p values were adjusted for multiple hypothesis testing using a Benjamini-Hochberg correction. Significantly enriched alleles are shown in Table 1 and complete results are presented in Table S4. For this analysis, OMAS n = 38; LR = 27 (13 this study, 14 from TARGET); HR = 146 (13 this study, 133 from TARGET), non-OMAS = 186.

Immune landscape signatures

Immune landscape signatures, including cytotoxic lymphocyte activity (esp. CD8 T cells), B cell activity, IFNγ levels, T cell trafficking, immune suppression activity from myeloid-derived cells (M2TAM cells, TGFB1 levels, PD-L1, etc), checkpoint ratios, and stromal responses, were detected in each sample as described in.²⁹ These immune signature scores represent weighted averages of (log) expression levels of genes within each signature. Each immune subtypes was tested for enrichment in OMAS patients using a t test following by multiple testing correction of p values using Benjamini-Hochberg. Features that showed statistically significant differences between OMAS and non-OMAS samples were plotted in a separate box (top) and features not showing significant differences between groups were plotted in a heatmap below. Clustering of samples was performed to maximize similarity of gene expression patterns in heatmap for significant features in upper box, using Pearson’s correlation.

Variant identifications

Raw FASTQ were processed with a pipeline optimized for variant calling in RNA-seq data. First, raw reads were processed using FASTP v0.19.4⁹⁶ and the parameters: –trim_poly_g, –trim_poly_x, –cut_by_quality3, –cut_mean_quality 20, –n_base_limit 4, –qualified_quality_phred 15, –length_required 25, –complexity_threshold 30, –low_complexity_filter, –correction, –html. Trimmed reads were aligned against the GRCh37 reference genome and GENCODE v27 transcriptome using the STAR v2.5.4b aligner (parameters: –runMode alignReads, –alignSJDBoverhangMin 2, –alignSJoverhangMin 8, –chimFilter None, –chimJunctionOverhangMin 10, –chimMainSegmentMultNmax 10, –chimOutType SeparateSAMold, –chimScoreDropMax 30, –chimScoreMin 1, –chimScoreSeparation 7, –chimSegmentMin 10, –chimSegmentReadGapMax 3, –outFilterIntronMotifs RemoveNoncanonicalUnannotated, –outFilterMultimapNmax 20, –outFilterMultimapScoreRange 1, –outFilterScoreMinOverLread 0.66, –outMultimapperOrder Random, 0-outSAMstrandField intronMotif, –outSAMunmapped Within, –quantMode TranscriptomeSAM, –readFilesCommand zcat, –twopassMode, Basic). Variants were detected using “GATK best practices for variant calling on RNA-seq”, using Sentieon’s suite of tools⁹⁷ in place of GATK. Gene expression was evaluated in-pipeline using RSEM v1.3.0. These gene counts, as well as other metrics such as coverage statistics, gene region annotations, RNA editing sites, and clinVar and dbSNP annotations, were input as features into a random forest model which further filtered variants and removed false positive variant predictions. Genes containing exonic variants in one or more samples were examined for enrichment of these variants in OMAS patients. Significance was determined using Fisher’s exact test and corrected for multiple tests using FDR. Significant genes were further examined to identify any single variants driving the significance results. For each significant gene, SNPs within the gene were tested independently for enrichment in OMAS patients, with FDR correction for multiple tests.

Immune subtype classifier

The Immune Subtype classifier, as described in³³ and updated in Gibbs DL⁸⁵ was applied to the RNA-seq data collected in the current study, as well as to previously published neuroblastoma data from TARGET³⁸ and to data from the Pan Cancer Atlas.⁹⁸

TCR repertoire analysis

Tumor genomic DNA

Tumor genomic DNA was obtained from COG, and 31 OMAS, 13 LR and 13 HR patient samples were sequenced for TCRβ locus, using the Adaptive Biotechnologies Immunoseq platform. Since input genomic DNA samples were not of uniform concentration, to compute repertoire size, total number of sequence reads obtained were normalized for the amount of input DNA loaded into the sequencing assay.

Data cleaning and normalization

For repertoire analysis, CDR3γ sequence reads that are in-frame and have no stop codon were considered; all other sequences were filtered out. For each amino acid sequence in a given sample, we summed the frequencies of all its nucleotide variants (due to convergent recombination) to obtain the frequency associated with the amino acid sequence in the given sample.

TCR data analysis

All computations were done using R (R version 3.6.3), running on a CentOS Linux 7 core. Data manipulation, plotting, and standard statistical tests were done by base R and standard packages. All computations involving, clonality and diversity were done using the same subsampling scheme. We subsampled all patient TCRβ repertoires to a common size (1,382 reads), computed the statistic and averaged the value of the statistic over 100 such iterations, samples smaller than the common size were discarded from further analysis. Sampling was done by the sample function in base R. Shannon index was computed using the vegan community ecology package (version 2.5–6). Average values over 100 subsampling iterations were plotted using ggplot2 package. Unless otherwise indicated, comparisons between groups were made using Wilcoxon rank-sum test, and FDR corrected for number of tests.

TCR sharing level

For each amino acid CDR3 sequence, we calculated its sharing level in the cohort, i.e., to how many samples it belongs. For each sharing level, we calculated the number of sequences that have this sharing level. Figure 4E describes in log-log scale the relative frequency of sequences in each sharing level. In Figure S4A, we compared the sharing level within neuroblastoma patient group to the sharing level in PBMC of healthy individuals as captured by the Emerson data set.⁴⁹ Each sequence was plotted according to its Emerson sharing level (X axis) and Patient Group sharing level (OMAS/LR/HR; Y axis). Some of the sequences highlighted in color are given in the “Overshared” sequences in Table S6.

Emerson data set

To estimate background frequencies of TCRβ receptor sequences, we used the Emerson data set,⁴⁹ a set of 786 repertoires from healthy volunteers The Emerson dataset was downloaded from the Adaptive web site (DOI https://doi.org/10.21417/B7001Z).

BCR repertoire analysis

For this analysis, we included available material from 37 OMAS-associated neuroblastomas, 13 LR, and 13 HR non-OMAS-associated neuroblastomas. IgH sequencing was performed on genomic DNA using the Adaptive Biotechnologies platform. For analyses requiring subsampling, a common size of 219 sequences was chosen, as a tradeoff between repertoire size and # of samples needed for statistical power. Using this minimum size, this left 5 HR, 8 LR, and 29 OMAS samples for comparison.

BCR data analysis

For BCR analysis, we used the immcantation portal packages to compute gene usage, clonality, clustering, mutation frequency and diversity. All computations were done using R (R version 3.6.3), running on ubuntu 16. Data manipulation, plotting, and standard statistical tests were done using base R and standard packages. Diversity and Shannon index analysis was done using alakazam and shazam R packages from immcantation.⁸³ For Shannon index, each sample was subsampled 100 times to a minimum repertoire size (219 sequences) with sequence replacement. Significance was determined using the Wilcoxon test and p values were corrected for multiple tests with FDR. Clonality was performed using Change-O from immcantation. Unless otherwise indicated, comparisons between groups were made using Wilcoxon rank-sum test, and FDR corrected for number of tests.

IgH gene assignment

IgH sequences were aligned to IGHV, IGHD, and IGHJ genes by applying IgBlast⁸⁴ using a reference germline that was downloaded from IMGT in 2017. The repertoires were sequenced using the Adaptive Biotechnologies ImmunoSeq platform, which returns only a partial V sequence assignment. This can cause mis-assignment of the V gene. Thus, for better clone inference for each patient, clones were defined as the same V family, J gene, and junction length using Change-O.⁸³ The cutoff threshold was determined with the shazam package.⁸³

IgH clusters

To define clusters of sequences, all subjects’ repertoires were pooled, and clusters were inferred by the DefineClones function from Change-O using the complete linkage method. The clusters were defined as sequences that share the same V family, J gene, and junction length. We also required a minimum of 85% amino acid identity across the junction sequence for inclusion. Clusters containing at least one sequence from at least 7 OMAS subjects were chosen for plotting.

Mutation analysis

Mutation frequency of a sequence was calculated as the number of mutations compared to the V germline sequence divided by the length of the V region sequences. For each subject the sequences for each V family were grouped and the median mutation frequency was selected. Significance was determined using the Wilcoxon test and p values were corrected for multiple tests with FDR.

IGHV gene usage

IgH sequences obtained using the ImmunoSeq platform carry only a partial V region, which hinders accurate assignment of V gene identity. To avoid mis-assignment biases, uncertain or unreliable gene assignments were filtered out using the RAbHIT package.⁸⁶ Then, relative gene usage was calculated using the alakazam package.⁸³ Significance was determined using the Wilcoxon test and p values were corrected for multiple tests with FDR.

XG Boost: Building a binary classifier out of RNA-seq data

Machine learning procedures were carried out using the python scikit-learn (version 0.18.2) and XGBoost package. We chose Gradient Boosting Decision Trees (specifically eXtreme Gradient Boosting, XGBoost³⁶ as the prediction algorithm for its ability to capture non-linear interactions between features, its efficiency and the fact that is has been successfully used in a wide range of applications.

Due to the relatively low number of samples available, we used leave-one-out as the cross-validation scheme and did not perform hyperparameter optimization to avoid reducing the sample size even further by putting aside a dedicated subset used only for model optimization. For each iteration, XGBClassifier was trained on FPKM values from all but one sample, and the resulting model was used to predict the class of the left out sample (either OMAS vs. non-OMAS, OMAS vs. HR, or OMAS vs. LR). The performance was scored using the area under the ROC curve as a metric. ROC curves for each comparison, as well as top features for each XGBoost model, are given in Figure S3. Feature importance and effect on the model was determined using SHAP analysis.⁸²

Immunohistochemistry, TLS imaging, and histological scoring

Formalin-fixed, paraffin-embedded (FFPE) sections from OMAS and non-OMAS patient tumors (5 μmmicron sections, charged slides, air dried) were obtained from primary tumor resection (with two exceptions, which were obtained from biopsies). Sections were obtained on slides from Children’s Oncology Group. Images of H&E-stained sections from the same specimens, which had been prepared, stained using standard methods, and imaged previously by COG at 40X magnification, were also obtained for scoring.

Immunohistochemistry

Unstained slides of FFPE sections were stained as follows: Slides were rinsed in 2 changes of xylene for 5 min each, then rehydrated in a series of descending concentrations of ethanol. Slides were then treated in a pressure cooker with antigen unmasking solution (Vector Laboratories H-3300) for 30 min. After cooling, slides were rinsed in 0.1M Tris Buffer, and then blocked in 0.1M Tris buffer, 0.01% tween with 2% fetal bovine serum for 15 min. For primary antigen detection, the following primary antibody combinations were used: a) Rabbit anti-CD3 (1:50, Dako A0452), incubated overnight, and mouse anti-CD20 (1:500, Dako M0755), which were incubated with the slides for 1 h at room temperature, and b) Goat anti-human CD4 (1:400, R&D Systems AF-379-NA) and Rabbit anti-human CD8 (1:400, Thermo RB-9009-P0), which were both incubated for 1 h at room temperature. After primary antibody staining, slides were again rinsed several times in 0.1M Tris Buffer with 0.01% Tween, and then incubated with the following secondary antibody combinations: For CD3/CD20 detection, Alexa 488 anti-Rabbit (Life Technologies, A21206), with Alexa 594 Anti-Mouse (Life Technologies, A11032) were used. For CD4/CD8 detection, Alexa 488 anti-Goat (Life Technologies, A11055) with Alexa 594 anti-Rabbit (Life Technologies, A21207) at a 1:400 dilution were used. All slides were incubated with secondary antibodies for 1 h at room temp. Slides were rinsed several times in 0.1M Tris/0.01% Tween, then counterstained for 5min in DAPI Hydrochloride (Sigma 32670). Slides were then rinsed, and then coverslipped with Prolong Gold (Life Technologies, P36930). Slides were digitally scanned at 20x magnification (Aperio IF, Leica Biosystems).

For Ki67 staining, coverslips from stained slides were removed by incubating the slides in 1xPBS at 37°C overnight, and then washed for 2 h in 1x PBST with several changes, before proceeding to Ki67 staining. Without removing prior staining (CD3-alexa 488/CD20 Alexa 594/DAPI), slides were further stained using Rabbit monoclonal anti-Ki-67:Alexa Fluor 647 direct conjugate (Abcam ab196907, 1:100) at 4°C overnight. Slides were then washed in 1x PBST with several changes for 2 h before mounting and coverslipping in Slow-fade Gold mounting medium (ThermoFisher).

Histological and immunohistochemical examination of tumor specimens

Formalin-fixed paraffin-embedded tissue sections stained with hematoxylin and eosin (HE) from each of the tumor samples included in the study were histologically revised to confirm the initial diagnosis of neuroblastomas or ganglioneuroblastomas applying the criteria for classification of neuroblastic tumors suggested by the International Neuroblastoma Pathology Committee.⁹⁹ Signs of differentiation tendency in the neuroblastic tumors, such as presence of neuropils, Homer-Wright rosettes, and different stages of maturation towards ganglion cells were recorded. Additionally, we assessed the possible presence of tertiary lymphoid structures containing lymphatic follicles with or without germinal centers according to previously published quantification criteria (none = 0; present in <10% of tumor tissue = 1+; present in 10%–50% of tumor tissue = 2+; present in >50% of tumor tissue = 3+; Hudlebusch et al., 2011).

Lymphocytic populations in the tumor-associated lymphoid structures and elsewhere in the tumors were assessed by immunofluorescent staining of tissue sections using primary antibodies against human CD20 and CD3, a B-cell and T cellT-cell marker, respectively, as described above. Proliferation activity characteristic of the germinal centers of lymphatic follicles was assessed using immunofluorescent staining against human Ki67, as described above.

Imaging of TIL immunohistochemistry

Images were acquired on a Leica LMD upright widefield microscope driven by the LAS X acquisition software, with a 20X objective. Raw images were identically scaled and then exported as TIFFs.

QUANTIFICATION AND STATISTICAL ANALYSIS

For all figures, statistical details of each experiment are noted within the figure and/or figure legends. Unless otherwise indicated, the median for each sample set is indicated by a red line in boxplots, and the number of patient samples (n) is given in the figure legend for each figure.

Differential expression

To assess differential gene expression between OMAS associated and control NB tumor RNA samples (Figures 1 and S1), the differential expression p values and classification probabilities from five tools– t test, limma4, DESeq2, edgeR, and EBSeq – were integrated to produce a new p value for differential expression. For any given gene, the p values of each constituent model were input into a logistic regression model, which estimates the probability that the gene is differentially expressed. This probability is transformed into a p value for differential expression by comparing it against its empirical cumulative distribution, as estimated by bootstrap resampling of TCGA data from multiple cancer types.

Differentially expressed pathways, gene clusters and networks

Genes that were differentially expressed between OMAS and non-OMAS patients, with cutoff of 2-fold difference in gene expression between groups and FDRq<0.05 were used for clustering and network analyses (Figure 2). The matrix of the normalized expression level these genes in 64 patient tumors sequenced in this study were used as input for gene clustering, using the R package “coseq”,^26,27 using a Gaussian mixture model, arcsin transformation of expression values, and an expression value cutoff of 50. Eight clusters were generated because K = 8 minimized ICL (integrated classification likelihood), with a series of trials with values of K between 2 and 25. Networks of genes from each of the 8 clusters were generated based on their normalized expression profiles, using the default settings of R package, “GENIE3”.^87,95 The top 20% of weighted connections among genes in the GENIE3 output were used to visualize the network of genes in each cluster, using Cytoscape.⁸⁸ Mean values for each gene within a cluster were obtained for samples within each patient group and depicted as a boxplot (Figures 2C and 2D) where each point represents a mean expression value for one gene in the cluster, for each group. Groups were compared using Wilcoxon test and p values were FDR corrected.

Subsampling of immune repertoires

In-frame sequences from TCRβ and IgH loci obtained from genomic DNA from patient tumors were used to generate immune repertoires for analysis, using the Adaptive Biotechnologies ImmunoSeq platform. To account for differences in repertoire sizes between patients in comparing repertoire characteristics, each repertoire was subsampled uniformly 100 times, and the mean of these computed (of the Gini, Shannon etc.). This value was used as the estimated diversity for each sample. Between samples in a patient group, the median was computed and is indicated in the plots by a red bar. In the randomization subsampling process, samples smaller than 1382 reads were discarded. This resulted in the exclusion of some patient samples compared to the RNAseq analysis. Significance was defined by p value ≤ 0.05.

HLA allele enrichment

HLAProfiler was used to identify HLAalleles based on RNA expression levels, as previously described (Table 1). For some genes, HLAProfiler identified alleles in fewer than 25% of samples. Alleles from these genes or alleles identified in only a single sample were excluded from the enrichment analysis. Significance of enrichment was determined by testing the distribution of each allele among patient subgroups compared to all alleles for the gene in the population (n = 2*number of samples). Fisher’s exact test p values were adjusted for multiple hypothesis testing using a Benjamini-Hochberg correction. For this analysis, OMAS n = 38; LR = 27 (13 this study, 14 from TARGET); HR = 146 (13 this study, 133 from TARGET), non-OMAS = 193 (includes 20 non-OMAS samples listed as IR or UK from TARGET).

SNP enrichment

Genes containing exonic variants in one or more samples were examined for enrichment of these variants in OMAS patients (Table S3). Significance was determined using Fisher’s exact test and corrected for multiple tests using FDR. Significant genes were further examined to identify any single variants driving the significance results. For each significant gene, SNPs within the gene were tested independently for enrichment in OMAS patients, with FDR correction for multiple tests. For this analysis, OMAS n = 38; LR = 27 (13 this study, 14 from TARGET); HR = 146 (13 this study, 126 from TARGET), non-OMAS = 186 (includes 20 non-OMAS samples listed as IR from TARGET).

Immune landscape signature

Immune signature scores (Figure S2) represent weighted averages of (log) expression levels of genes within each signature.²⁹ Immune subtypes were tested for enrichment in OMAS patients using a Fisher’s exact test with FDR correction.

CIBERSORT

Immune cell type subsets in tumor RNA samples were estimated based on cell type expression profiles, using CIBERSORT (Figures 1B–1D). Boxplots were generated using ggplot2, with the box covering the 1st to the 3rd quantile of the data, whiskers covering the minimum to the maximum of all non-outlier data and the bold line in the middle of the box indicating the median of the data. Each datapoint representing the cell type fraction from a single sample was overlaid on the boxplot. Patient groups were compared using Wilcoxon sum test with FDR correction.

ADDITIONAL RESOURCES

Samples used in this study were collected as part of a clinical trial, described in de Alarcon PA et al. 2018,²¹ and registered here: https://clinicaltrials.gov/ct2/show/NCT00033293.

Highlights.

• OMAS-associated neuroblastomas contain more B and T cells than control neuroblastomas

• OMAS-associated neuroblastoma B and T cell repertoires are diverse, with small clones

• Tertiary lymphoid structures are enriched in OMAS-associated neuroblastomas

• Gene expression correlated to neurological symptom severity nominates autoantigens

INCLUSION AND DIVERSITY

We support inclusive, diverse, and equitable conduct of research.