Two neuroblastoma cohorts were analyzed. The discovery cohort included patients from Therapeutically Applicable Research to Generate Effective Treatments (TARGET) program (n=149; 123 high-risk) (dbGAP accession ID phs000218.v22.p8) (online supplemental table 1). RNAseq paired-end (PE) FastQ files, whole exome sequencing (WES) alignment BAM files, somatic mutation MAF (Mutation Annotation Format) files, and clinical data were downloaded from Genomic Data Commons (GDC)30 (https://portal.gdc.cancer.gov) (accessed 07/2017). The validation cohort included patients with clinical information in the International Neuroblastoma Risk Group (INRG) Data Commons31 and tumor RNAseq data in the Gabriella Miller Kids First (GMKF) program (n=198; 48 high-risk) (online supplemental table 2). Universal system identification (USI) numbers were used to link the datasets. Access to RNAseq PE FastQ files in GMKF could not be obtained at the time of study, and therefore, preprocessed gene expression TSV files from the GMKF data portal (https://kidsfirstdrc.org/) (accessed 08/2020) were used for analysis. Of the 209 patients identified, 11 were determined by USI number to also be included in the discovery cohort, hence were excluded from the validation cohort; 198 were kept for validation (online supplemental table 2).

The quality of raw sequencing reads was assessed by FastQC32 (V.0.11.5) for the tumor samples in the discovery cohort. Read counts were quantified at transcript level using Kallisto33 (V.0.44.0) with human reference assembly GRCh38 and Gencode gene annotation (V.28), summarized into gene level using tximport34 (V.1.4.0), normalized by trimmed mean of M-values (TMM) method, and log₂-transformed.

Using a defined T cell-inflamed gene expression signature,13 21 the tumors in the discovery cohort were categorized into three groups (T cell-inflamed, non-T cell-inflamed, and intermediate) using consensus clustering methods following previous protocols.13 In brief, an expression matrix consisting of the 160 genes from the T cell-inflamed signature21 was subset from the TMM-normalized and log₂-transformed RNAseq gene expression quantification matrix and was used to cluster tumors into 12 clusters by ConsensusClusterPlus (V.1.42.0) using hierarchical clustering with Euclidean distance and Ward.D2 linkage (2000 bootstraps and 80% usage of gene features). Tumors were then assigned with each of the three immune groups based on high, low, or intermediate expression of the T cell-inflamed signature. The number of clusters was determined using the elbow method.

The assignment of tumor groups in the discovery cohort cannot be migrated directly to that of tumors in the validation cohort due to the relative nature of gene expression data without spike-in controls. To address this issue, we projected T cell-inflamed gene expression of the validation cohort to the space of the discovery cohort using 11 patients that overlap between the two cohorts. First, we normalized and log₂-transformed gene expression within each cohort. We calculated a T cell-inflamed score for each tumor, defined as the mean expression of all genes from the signature.21 Then, we fit a linear regression model on the T cell-inflamed scores of the discovery cohort and validation cohort using tumors from the 11 overlapping patients, ScoreT=−0.7798+1.2418×ScoreG (adjusted R²=0.967), where ScoreT represents T cell-inflamed score of the discovery cohort (T as TARGET), and ScoreG represents T cell-inflamed score of the validation cohort (G as GMKF). We used this model to convert all T cell-inflamed scores of tumors in the validation cohort (ScoreG) to values comparable to that of the discovery cohort (ScoreT), then sorted all tumors by T cell-inflamed scores lower to higher. Lastly, we assigned new tumor groups to the validation cohort based on existing tumor groups from the discovery cohort, employing the rules as follows: for all tumors harboring a score less than or equal to that of the last non-T cell-inflamed tumor on the sorted list, assign as non-T cell-inflamed; for all tumors harboring a score greater than or equal to that of the first T cell-inflamed tumor on the sorted list, assign as T cell-inflamed; otherwise, assign as intermediate.

For the discovery cohort analysis, we focused on 19,883 protein-coding genes defined in Gencode annotation (V.28) and followed the protocol established in our previous work.21 In brief, after removing genes with low expression (defined as CPM (counts per million of mapped reads) ≤3), 15,580 genes with CPM>3 in at least 30 tumors were TMM-normalized and log₂-transformed. Differentially expressed genes (DEGs) comparing non-T cell-inflamed with T cell-inflamed groups were identified using Linear Models for Microarray Data (limma) voom35 method with precision weights (V.3.36.2) and filtered by false discovery rate (FDR)-adjusted p<0.05, and fold change ≥1.5 or ≤−1.5. Upstream transcriptional regulators and change of direction (activation or inhibition) as a result of target molecules (encoded by DEGs) were predicted using Ingenuity Pathway Analysis (IPA) (QIAGEN, Germany) causal network analysis36 with the curated Ingenuity Knowledge Base (accessed 12/2017). Transcriptional programs activated in non-T cell-inflamed relative to T cell-inflamed tumors were filtered at overlap p<0.05 (measuring the enrichment of target molecules in the dataset) and z-score ≥2.0 (measuring the predicted activation level of the pathways). For the validation cohort analyses, preprocessed RNAseq expression data downloaded from the GMKF data portal was quantified using Kallisto33 (V.0.44.0), and the per-tumor gene expression files were downloaded and aggregated into cohort level, TMM-normalized, and log₂-transformed for further analysis.

For the discovery cohort, the somatic mutations were harmonized using four somatic variant callers (MuTect2, VarScan2, SomaticSniper, and MuSE).30 After rigorous filtering following GDC’s guidelines (https://docs.gdc.cancer.gov/Data/File_Formats/MAF_Format), somatic variants that were detected by at least two callers and passed all the filters were selected for further analysis. Total TMB was defined as the total number of non-synonymous somatic mutations (NSSMs), those that were predicted to alter protein sequence in tumor (insertions/deletions, missense/nonsense/stopgain mutations, and those that modify splicing sites). Putative neoantigens were predicted from NSSMs using netMHCpan37 (V.4.0), filtered by gene expression from the RNAseq data described as follows. Patients’ major histocompatibility complex (MHC) class I haplotypes were predicted from WES of germline DNA using Optitype (V.1.3.1). Nine-mer peptides were generated from the mutated site through a sliding window approach using in-house python scripts. Our previous work had suggested that peptides of SYFPEITHI38 mutant score >25 or delta score (mutant – wildtype)>5 bind to MHC class I molecules.13 In this study, we used netMHCpan that covers more human leukocyte antigen (HLA) genotypes than SYFPEITHI. To select neoantigens that are likely to have strong binding affinity to HLA-A molecules and expressed in tumor, we filtered for 9-mer peptides of netMHCpan mutant score >0.638 (equivalent to IC 50 nmol, strong binding) or delta score >0.070 (correlated with SYFPEITHI delta score 5) and derived from genes upregulated compared with the median of its expression across all tumors.

For each tumor, a T cell-inflamed score was computed as the mean expression of the 160 genes involved in the signature after scaling and centering across all tumor samples.21 A pathway activation score was calculated to each tumor following our published protocol,17 21 requiring at least 50% of the pathway-specific target molecules to be upregulated in a tumor sample (relative to its median expression across all tumor samples) in non-T cell-inflamed relative to T cell-inflamed group. For pathways in which less than 10 target molecules were present, 5 or more molecules were required to be upregulated to classify the pathway as activated. For pathways in which less than 5 target molecules were present, only pathways with all molecules upregulated were classified as activated. In addition, for each pathway identified in this study (MYCN, ASCL1, SOX11, and KMT2A), the expression level of a pathway was defined by the mean expression of all target molecules from this pathway, which was then used to correlate with the T cell-inflamed gene expression across all tumors by Spearman’s correlation.

Cox proportional hazards (PH) models were used to test the association between the tumor group (T cell-inflamed, non-T cell-inflamed, and intermediate) and the survival outcome (event-free survival (EFS); overall survival (OS)) in the discovery (n=118 high-risk patients diagnosed between 2000 and 2010) and validation cohorts (n=17 high-risk patients with survival data available) using R package survival (function coxph) (V.2.41.3). Univariable and multivariable Cox PH models were used to assess the significance of tumor group as a single predictor or after adjusting for covariates including age, MYCN status, and ploidy. In addition, Kaplan-Meier (KM) estimator with log-rank test was performed using R package survminer (V.0.4.2).

Immunofluorescence (IF) staining on human neuroblastoma tumors was performed by the Human Immunologic Monitoring Core Facility at The University of Chicago using tissue from 17 intermediate or high-risk neuroblastomas (5 MYCN-amplified and 12 MYCN-non-amplified). Briefly, slides were baked, cleared, and rehydrated. After heat-inducted epitope retrieval, the slides were placed in a humidity chamber, blocked by 10% donkey serum for 1 hour, incubated with anti-CD8 Ab (Dako, M7103) at 1:100 dilution for 1 hour, followed by Cy3 donkey anti-Mouse IgG (Jackson Immunological Research Lab, 715-165-150) at 1:500 dilution for 1 hour. The slides then incubated with anti-Batf3 Ab (Novus, AF7437) at 1:40 dilution for 1 hour, followed by Cy5 donkey anti-Rabbit IgG (Jackson Immunological Research Lab, 711-175-152) at 1:200 dilution for 1 hour. After thorough wash, slides were incubated in DAPI and mounted with Fluoromount (Sigma, F4680). Images of the slides were taken using a Leica SP8 laser scanning confocal microscope at Integrated Light Microscopy Core Facility. A pathologist (PP) scored tumors for intensity and distribution of CD8⁺ cells and Batf3⁺ cells in a blinded fashion.

For analysis of contingency tables including comparison of tumor sample frequency between groups, Fisher’s exact test was used. Differential gene expression analysis between groups were performed using empirical Bayes regression models in limma voom with precision weights. For multiple comparisons, p-value was adjusted using Benjamini-Hochberg FDR correction for multiple testing.39 Spearman’s correlation ρ was used for measuring statistical dependence between normalized and log₂-transformed expression level of different genes and between gene expression of the T cell-inflamed signature and pathways. p<0.05 was considered statistically significant. Statistical analysis was performed using R (V.3.5.2) and Bioconductor (release 3.8).

Using a defined T cell-inflamed gene expression signature, we categorized the 149 primary neuroblastoma tumors from the discovery cohort (TARGET) into three subsets (figure 1A). High expression of T cell signature genes (T cell-inflamed) was detected in 57 (38.3%) tumors, low or no expression (non-T cell-inflamed) was identified in 45 (30.2%) tumors, and 47 (31.5%) had intermediate levels of expression (intermediate) (table 1). In the validation cohort (n=198, GMKF), 89 (44.9%) were categorized as T cell-inflamed; 55 (27.8%) were non-T cell-inflamed; 54 (27.3%) were intermediate (figure 1B, table 2). In the discovery cohort, 123 of 149 patients were classified as high-risk, whereas 48 of 198 patients in validation cohort have high-risk neuroblastoma.40 In analyses restricted to high-risk patients, 53 (43.1%) and 23 (47.9%) were categorized as T cell-inflamed in the discovery and validation cohorts, respectively, and 33 (26.8%) and 18 (37.5%) were categorized as non-T cell-inflamed.

In the discovery cohort, MYCN amplification was significantly more prevalent in the non-T cell-inflamed tumors (17/45, 37.8%) compared with the T cell-inflamed tumors (3/56, 5.4%) (p=0.000080, odds ratio [OR]=10.5, two-sided Fisher’s exact test). Additionally, patients diagnosed at age <18 months had tumors that were enriched in the non-T cell-inflamed tumor group (12/45, 26.7% non-inflamed vs 6/57, 10.5% inflamed; p=0.040, OR=3.06). The enrichment of MYCN amplification in non-T cell-inflamed tumors was also observed in the validation cohort (14/54, 25.9% non-inflamed vs 4/89, 4.5% inflamed; p=0.00038, OR=7.33). In addition, patients <18 months of age in the validation cohort had tumors that were enriched in the non-T cell-inflamed tumor group (40/54, 74.1% non-inflamed vs 44/89, 49.4% inflamed; p=0.0049, OR=2.90).

We analyzed EFS and OS according to the level of expression of the T cell-inflamed signature in 118 high-risk patients diagnosed between 2000 and 2010 from the discovery cohort, which consists of 51 T cell-inflamed, 33 non-T cell-inflamed, and 34 intermediate tumors. In Cox PH univariable models, patients with T cell-inflamed tumors had significantly better OS compared with those with non-T cell-inflamed tumors (p=0.043, hazard ratio [HR]=0.56) (table 3). A similar trend was observed in EFS but the results did not reach statistical significance (p=0.17, HR=0.69) (table 3). Similar to other high-risk cohorts,41 42 MYCN status was not statistically significantly associated with OS or EFS (OS: p=0.38; EFS: p=0.58) (table 3). However, OS, but not EFS, was better for high-risk patients with hyperdiploid neuroblastoma compared with those with diploid tumors (OS: p=0.05; EFS: p=0.10) (table 3). In Cox PH multivariable models adjusting for age, MYCN status, and ploidy, the T cell-inflamed signature maintained independent statistical significance for OS (p=0.035, HR=0.48). Stage and histology were not included in the multivariable analysis because the high-risk patients had predominantly stage 4 disease and unfavorable histology tumors (117/118, 99.1% as stage 4; 107/110, 97.3% with unfavorable histology tumors, 8 unknown).

In the discovery cohort, the T cell-inflamed and intermediate groups showed similar probabilities of survival. Therefore, we combined these two groups and compared probability of survival to patients with non-T cell-inflamed tumors (OS: p=0.0076, EFS: p=0.10, log-rank test) (KM estimator shown in figure 2A,B). Similar associations between T cell-inflamed/intermediate tumors and improved survival were observed in the 17 high-risk patients with available survival data from the validation cohort (OS: p=0.016, EFS: p=0.0098) (figure 2C,D). No significant association with survival outcome was detected for age, MYCN status, or ploidy. Patients were not selected by diagnosis year 2010 or earlier due to small sample size. However, EFS and OS were not significantly different between the 13 patients diagnosed between 2008 and 2010 and the four patients diagnosed between 2011 and 2012 (OS: p=0.98, EFS: p=0.66), and hence these are unlikely to contribute to the significantly better survival outcome observed in T cell-inflamed group relative to non-T cell-inflamed group in the validation cohort.

Neoantigens are mutant antigens that are only expressed on tumor cells and not normal cells. Neoantigen-derived epitopes (neoepitopes) are recognized by antigen-specific CD8⁺ T cells.43 To evaluate the neoantigen load in neuroblastoma tumors, WES data from 198 matched tumor/normal pairs of the discovery cohort carrying one or more somatic single nucleotide variants (SNVs) were analyzed. After combining calls of four somatic callers and rigorous quality filtering, 4235 somatic SNVs were identified in 3369 genes. Each tumor harbors a median of 17 somatic SNVs (range, 1–168 SNVs), with 15 somatic SNVs predicted to alter protein sequences (range, 1–162), which is consistent with the somatic mutation profile previously reported in high-risk neuroblastoma.44 To investigate if neoantigen load was associated with outcome in high-risk patients, of 118 high-risk patients diagnosed between 2000 and 2010 in the discovery cohort, we analyzed tumors from 89 patients with both WES and RNAseq data available. The total number of neoantigens in tumor was determined by filtering for those predicted to bind to MHC class I molecule HLA-A. We focused on HLA-A molecule because the prediction algorithm for this allele is the most reliable.

A median of 4 (range 1–30) candidate neoantigens were identified in 78 of 89 tumors. Seventy-four patients diagnosed between year 2000 and 2010 were included in survival analysis. We found that the neoantigen load was significantly associated with OS (p=0.00022, log-rank test) (figure 3A) and EFS (p=0.0044) (figure 3B), although there was no significant difference in neoantigen load between non-T cell-inflamed and T cell-inflamed groups (p=0.22, two-sided Wilcoxon rank-sum test) (figure 3C). We defined four patient groups (hereafter referred as, quadrants (Q)) split by the threshold of T cell-inflamed (Tinfl) gene expression in non-T cell-inflamed tumors and median of neoantigen load (Neo) (Spearman’s correlation coefficient ρ=0.053, p=0.65) (figure 3D): Q1 (n=8), Tinfl_lowNeo_high; Q2 (n=19), Tinfl_lowNeo_low; Q3 (n=20), Tinfl_highNeo_low; Q4 (n=27), Tinfl_highNeo_high. OS and EFS were significantly different according to quadrant assignment (OS: p=0.00083; EFS: p=0.0061, log-rank test) (figure 3E,F). Patients in Q1 and Q4, who had tumors harboring high level of neoantigens, had superior outcome compared with those in Q2 and Q3 (figure 3E,F).

To investigate if tumor-intrinsic transcriptional programs may play a role in inhibiting T cell infiltration in non-T cell-inflamed neuroblastomas, we first analyzed tumors from the discovery cohort for signaling pathways intrinsic to the neoplastic cells that were previously reported to impair the local immune response in other tumor types. This includes somatic activation mutations in CTNNB1 or damaging mutations in repressors of the pathway (APC/APC2/AXIN1/AXIN2),15 17 somatic copy number loss in PTEN or activation mutations in PIK3CA,18 activation mutations in VEGF-A,45 and loss of function mutations in B2M,46–48 STK11/LKB1,19 IDH1/2,49 and NRAS/KRAS/HRAS.50 Only three tumors harbored a missense mutation in AXIN2 (p.A113T), and two tumors had PIK3CA missense (p.K111N) or nonsense mutations (p.E888X), but none occurred at the known PIK3CA activation mutation positions (AA 345, 542, 545, 546, 1043, 1044, 1047).

We next took an unbiased approach15–17 to identify transcriptional programs that are activated in non-T cell-inflamed tumors by comparing the whole transcriptome RNAseq expression of 33 non-T cell-inflamed to 53 T cell-inflamed tumors from the high-risk patients in the discovery cohort. A total of 1730 genes were identified that were significantly differentially expressed between the two tumor groups, with 230 upregulated in non-T cell-inflamed group and 1500 upregulated in the T cell-inflamed group (FDR-corrected p<0.05, fold change ≥1.5 or ≤−1.5). Causal network analysis36 with Ingenuity Knowledge Base (Qiagen) identified activation of MYCN signaling in non-T cell-inflamed tumors (activation z-score ≥2.0, p<0.05), consistent with our findings showing enrichment of MYCN amplification in non-T cell-inflamed tumors (figure 4A). Immunohistochemistry (IHC) staining of a limited number of available intermediate or high-risk neuroblastoma tumors (n=17, 5 MYCN-amplified and 12 MYCN-non-amplified; images of all IHC slides are provided at https://github.com/riyuebao/NBL-TME-Immunogenomics) demonstrated lower infiltration with CD8⁺ T cell and Batf3⁺ DCs in MYCN-amplified tumors compared with MYCN-non-amplified neuroblastomas (figure 4B), although the difference did not reach statistical significance in this small cohort (p=0.22, two-sided Fisher’s exact test). In both the discovery and validation cohorts, the DC genes are highly correlated with the T cell-inflamed gene expression (p<0.05, Spearman’s correlation) (online supplemental figure 1).

To determine if activation of transcriptional programs other than MYCN signaling is associated with the non-T cell-inflamed phenotype, we repeated the differential gene expression and causal network analyses using only MYCN non-amplified tumors (n=91). Genes significantly upregulated in 16 non-T cell-inflamed neuroblastomas compared with 49 T cell-inflamed tumors were used to predict upstream regulators. Three pathways (ASCL1, SOX11, and KMT2A) were identified to be activated in non-T cell-inflamed tumors without MYCN amplification (activation z-score ≥2.0, p<0.05) (figure 4C). We next calculated an activation score for each pathway using previously described methods.17 21 The results showed that the three pathways operate in a partially exclusive manner (online supplemental figure 2), with activation of SOX11, KMT2A, and ASCL1 signaling detected in 66%, 30%, and 30% of non-T cell-inflamed tumors, respectively, compared with less than 5% of the T cell-inflamed tumors (figure 4D,E). Taking together, the activation of one or more pathways regulated by MYCN, ASCL1, SOX11, or KMT2A was found in 85% of the non-T cell-inflamed tumors. The inverse correlation between the expression of T cell-inflamed gene signature and the four pathways (MYCN, ASCL1, SOX11, KMT2A) was confirmed in the validation cohort (online supplemental figure 3A,B), providing strong evidence that the activation of the four transcriptional programs was significantly associated with a non-T cell-inflamed phenotype.