A systematic search to identify publicly available datasets was performed in January 2018 using the query “pleural mesothelioma” and filtering to keep only human samples and expression profiling experiments in multiple open access repositories. We also filtered datasets with less than 30 samples and kept experiments that covered most of the transcriptome. Raw data were downloaded whenever possible and subsequently processed accordingly as summarized hereafter: for whole-exome sequencing data, Broad’s Genome Analysis Toolkit best practices were followed, RNA-sequencing data were processed according to an in-house pipeline based on ENCODE best practices, and for expression arrays, robust multiarray average algorithm was used to normalize the data. A detailed description of data identification and data processing is available in online supplemental extended methods.

The immune infiltrate composition of 20 different cell types was deconvoluted from the obtained gene expression profiles using gene set variation analysis (GSVA) method.13 Briefly, for each sample and immune cellular fraction, we obtained a score between [−1,+1], with extreme values close to 1 or −1 indicating a relative strong presence or absence of a fraction in a specific sample, respectively. Similarly, GSVA scores close to 0 correspond to an unbiased distribution (ie, centered or randomly distributed) of the genes across the whole gene expression profile. A more detailed explanation on how this algorithm works or why its application is more robust to batch effect than other deconvolution methods can be found in Hänzelmann et al13 and Tamborero et al,14 respectively.

Gene signatures of the different immune fractions were obtained from a previous study15 in which signatures for a total of 24 different immune cell types were generated, including innate immune cells (eg, dendritic cells, eosinophils, mast cells, macrophages, natural killer cells, neutrophils) and adaptive immune cells (eg, B cells, T cells). Since GSVA can only deconvolute multigene signatures, we replaced two single-gene signatures (plasmacytoid dendritic cells and regulatory T cells) with their corresponding multigene signatures from another study.16 Additionally, when more specific categories were available for a cellular fraction (eg, NK CD56dim and NK CD56bright), general categories (eg, NK cells) were removed to avoid signature redundancy.

Once relative abundances for each immune fraction and sample were obtained, Cox proportional-hazards models adjusted for sex, stage, age, and histology were fitted using the largest dataset with available survival information as the discovery dataset,6 and setting the optimal cut point that maximized survival differences for each immune fraction using R packages survival (V.3.1.7) and survMisc (V.0.5.5). Using the GSVA cut-points defined in the discovery dataset, Cox proportional-hazards models adjusted for age, stage, sex, and histology were fitted in two additional validation datasets.7 10 After verifying there was no association of neoadjuvant therapy status with clinical outcome in Bueno et al6 (p=0.2, data not shown), we did not adjust for neoadjuvant therapy status due to lack of information in Bott et al10 and lack of diversity in Hmeljak et al7 (all patients are naïve) in order to have equivalent models to test the association of clinical outcome with the immune fractions. Fractions associated to clinical outcome were determined as those significant (p<0.05) after false discovery rate (FDR) adjustment in the discovery dataset and at least one validation dataset, and with a consistent HR across the three datasets.

In order to identify clinically relevant immune-cell fractions in MPM tumors, we collected data from seven MPM gene expression datasets that fulfilled our selection criteria.6–12 Overall, we collected 516 MPM gene expression profiles, with available survival information and covariates for 300 patients (58%). A summary of the clinicopathological characteristics of the samples in each dataset is available in online supplemental table S1.

For each sample included in the analysis, we quantified the abundance of 20 immune-cell populations by applying GSVA13 on a previously established set of immune-specific gene signatures (online supplemental table S2).15 16

Once the abundance of all immune fractions was quantified in the full cohort of samples, we aimed to assess whether there existed any individual fractions associated with overall patient survival using Cox proportional-hazards regression models adjusted for potential confounders (ie, age, sex, tumor stage, histology, and dataset). Thus, for each cell type, we identified the cut point that maximized the survival differences between high versus low using the largest available dataset (Bueno et al6). The cut points obtained in the discovery dataset were further validated in the two additional cohorts with available survival information (Hmeljak et al7 and Bott et al10). Out of the 20 studied immune cell types, T-helper 2 (T_H2) and cytotoxic T cells (T_C) showed a consistent association with patient survival across datasets based on our criteria (Cox p<0.05 in at least two of the three datasets, and the remaining dataset showing a consistent trend, although not necessarily significant). In our analysis, higher levels of T_H2 cells (GSVA score >−0.16) were associated with worse outcome (HR_Bueno=2.1, p=0.00015; HR_Hmeljak=2.6, p=0.0021; HR_Bott=2.3, p=0.066; HR_Combined=2.1, p=7.6·10^-7) (figure 1A), while higher levels of T_C (GSVA score >0.40) were associated with better prognosis (HR_Bueno=0.57, p=0.0091; HR_Hmeljak=0.73, p=0.39; HR_Bott=0.29, p=0.016; HR_Combined=0.59, p=0.00094) (figure 1B). Survival analysis results for the complete set of immune fractions can be found in online supplemental figure S1. When individuals were stratified in four groups based on the combination of the abundance (ie, low versus high) of both T_H2 and T_C cells, patients with MPM with low abundance of T_H2 and high T_C infiltration consistently showed the best overall survival when compared with patients with high levels of T_H2 and low levels of T_C (HR_Bueno=0.34, p=0.00022; HR_Hmeljak=0.20, p=0.0080; HR_Bott=0.12, p=0.020; HR_Combined=0.31, p=2.7×10⁻⁷) (figure 1C). Additionally, patients with either both high (IG2-H) or low abundance levels of both cell types (IG2-L) showed intermediate prognosis, suggesting some type of compensatory effect between these two immune cell types. Moreover, we do not observe statistically significant differences in survival when comparing IG2-H and IG2-L in any of the datasets. In Hmeljak et al cohort where although IG2-H patients (n=10) have a similar survival to IG1 patients, the HR and 95% CI between IG2-L and IG2-H (HR_IG2-L=0.39 (0.15 to 1.02)) is not statistically significant, with IG3 being the only group that shows an improved survival in that cohort. Therefore, we decided to merge these two groups into a single category to create a final classification of three immune-based groups: T_H2High-T_CLow (IG1, worse prognosis, 54.5% of the patients), T_H2High-T_CHigh or T_H2Low-T_CLow (IG2, intermediate prognosis, 37% of the patients, HR_Combined=0.52), T_H2Low-T_CHigh (IG3, better prognosis, 8.5% of the patients, HR_Combined=0.32) (figure 1D).

Once the immune-based MPM groups were defined, we aimed to identify any potential differences in the remaining immune fractions among them. Groups IG1 and IG2 comprised a mixture of samples with both strongly and less immune-enriched tumors, while on average, group IG3 showed significant higher abundance for 13 of the 20 immune cell fractions (figure 2A, online supplemental figure S2). These results suggest that the combination of these two immune fractions interplays with the remaining immune cell types and contributes to identify a subset of highly immune-infiltrated MPM tumors with improved prognosis, as already reported in previous publications.14

Regarding tumors in IG2 (figure 2A), we observe that samples from the cluster on the left-side are mainly from IG2-L, although 6.06% (n=6/105) is composed by IG2-H samples. On the right-side, consisting mainly of IG2-H samples, 41.0% of the samples (n=25/86) are from IG2-L, suggesting that samples with similar survival rates can have a heterogeneous landscape regarding immune cell abundance.

In order to explore potential additional clinical correlates of the three identified MPM immune groups, we analyzed their association with available clinical covariates: age, sex, tumor stage at the time of diagnosis, tumor histology, asbestos exposure, and neoadjuvant therapy. Out of all the variables explored, we identified a significant association with age and histology (table 1). Patients’ median age in group IG1 is 66 years, compared with 63.7 years and 63.2 years for groups IG2 and IG3, respectively (p=0.021). Regarding tumor histology, IG3 showed a strong enrichment in epithelioid tumors (90.70% in IG3 vs 62.04% in IG1 and 73.68% in IG2, p=0.001).

A multivariate analysis with the observations with complete data for all the covariates (n=231) confirmed the association with histology and age, and the lack of association of our classification with sex and stage (online supplemental table S3). A slight increase of asbestos exposure in IG2 as well as a marginal association with neoadjuvant therapy in IG2 was also observed.

To further exclude a potential confounding role of the tumor histology on the differential survival among the immune groups, we evaluated whether the immune-based classification still was associated with patient survival within both epithelioid and non-epithelioid tumors. In epithelioid tumors, we observed improved survival in IG2 and IG3 patients compared with IG1 (HR_IG2=0.47, HR_IG3=0.24, p=3.3×10⁻⁸) (online supplemental figure S3A). For non-epithelioid tumors, the very small number of IG3 patients (n=3) prevented us from obtaining reliable results for this subgroup, but we still observed a trend towards better survival in IG2 patients compared with IG1 (HR_IG2=0.60, p=0.074) (online supplemental figure S3B).

Finally, a comparison with previously published classifications6 7 9 was done. While the immune-based classification was not associated to De Reyniès et al9 classification, it was associated to Bueno et al6 classification, where IG3 is enriched in samples from Epithelioid subtype, and Hmeljak et al7 classification, where IG3 is enriched in S1 samples which is also enriched in epithelioid histology and has the best prognosis.

To further evaluate the potential clinical relevance of the identified immune groups, we evaluated the expression of a wide selection of ICI and activators. We observed a significant increasing expression pattern across MPM immune groups for most of the evaluated checkpoint markers (figure 2B). TNFRSF14, IL2RB, TNFRSF18 (GITR), CD27, and ICOS were among the strongest positively associated activator markers with the immune groups. Regarding ICI, LAG3, TIGIT, VSIR (VISTA), IDO1, CTLA4, and PDCD1 (PD-1) were the strongest correlated markers. In opposition, CD276 (B7-H3) showed negative correlation with the immune groups, being its expression higher in IG1 tumors than tumors in IG2 and IG3 groups. Detailed expression of CD274 (PD-L1), PDCD1 (PD-1), CTLA4, CD276 (B7-H3), and VSIR (VISTA) in each dataset is available in online supplemental figure S4A, showing the distribution of tumors in each group. An overall higher expression of VSIR with respect to the other immune checkpoints is observed, as previously reported in other publications.7 31 Importantly, immune groups still show significant differences in prognosis independently of the expression level of any of these five immune checkpoints (online supplemental figure S4B). Additionally, we also evaluated markers of effector and exhausted T cells correlated with the identified MPM immune groups (figure 2C). In our analysis, all evaluated markers—excluding CD44—showed a positive correlation with the immune group classification, with a majority being significantly enriched in IG3 (PRF1, GZB, TBX21, EOMES, PDCD1).

Once we identified the different immune-based groups in MPM tumors and established their potential clinical relevance, we wanted to characterize the different groups at the genomic level to identify any potential differential genomic patterns. These analyses were performed using Bueno et al6 and Hmeljak et al7 datasets.

Using whole-exome data from Bueno et al6 and Hmeljak et al7 datasets, we assessed if immune groups were different in terms of TMB, mutational signatures,32 or antigen immunogenicity (figure 3A–C). Although we did not find statistically significant differences among immune groups for any of the previous analyses, we did observe a slight decrease of mutational Signature 3 in IG3 tumors compared with IG1 and IG2 patients. Signature 3 is associated with homologous recombination deficiency and BRCA1/2 deficiency.32 Additionally, at the single-gene level, no gene was found to be preferentially mutated in any group (data not shown).

Copy number alterations were assessed for Hmeljak et al dataset.7 There were no significant differences between immune groups in the overall number of altered genes per sample, although tumors in IG3 tend to have fewer copy number alterations overall compared with the other groups as shown in figure 3D. Figure 3E shows the overall percentage of altered samples across the genome for each immune group. Regarding MPM landmark genes, genes like BAP1, SETD2, LATS2, TP53, and NF2 did not show differences between groups. CDKN2A is altered in 81% of IG1 patients, 47% of IG2 patients, and 14% of IG3 patients, but it does not reach statistical significance after multiple testing correction, probably due to small sample size. Online supplemental figure S5 shows expression of frequent tumor suppressor genes inactivated in MPM.3 For the case of CDKN2A, we see an increasing expression IG1<IG2<IG3, which is in agreement with the pattern observed at the genomic level.

We also evaluated differences in methylation data at the single-gene level and no significant results were found after adjusting for multiple testing (data not shown).

Next, we sought for gene expression differences among the three immune groups, to identify those genes with an increasing or decreasing pattern among the three groups. The results are depicted in figure 4A, which shows a high number of genes with decreasing expression from IG1 to IG3 tumors related to cell proliferation such as CENPF (centromere protein also known as mitosin), BIRC5 (otherwise known as survivin, preventing apoptotic cell death), or kinesins like KIF23, involved in cell division. On the other hand, genes positively correlated with the immune groups include immune system related genes such as HSH2D, a target of T-cell activation signaling pathways, and GNLY, which encodes for a protein present in cytotoxic granules of cytotoxic T lymphocytes.

In order to capture more precisely the enriched pathways and biological functions of the genes identified in the previous analysis, we performed a preranked GSEA using a set of transcriptional signatures covering a wide range of pathways, and the genes ranked according to their linear association with the immune groups. The results are shown in figure 4B and we observed that signatures more positively correlated with the immune group (ie, more expressed in IG3 tumors) were mostly related to immune system and inflammatory response processes. Contrarily, signatures negatively correlated with the immune groups mostly relate to cell cycle and epithelial mesenchymal transition, suggesting a trade-off between these molecular processes and the immune system.

Finally, since IG1 tumors appeared to be driven by cell cycle deregulation, we assessed the potential confounder effect among cell proliferation and immune groups. Therefore, we tested the association between the surrogate cell proliferation marker MKI67 and the immune groups. Notably, regardless of MKI67 expression levels, the association of the improving survival pattern across immune groups is upheld (online supplemental figure S6), meaning that our classification is not confounded with cell proliferation.

To further evaluate the potential clinical benefit of the immune-based MPM classification, we decided to assess the behavior of different treatment-response signatures among the three groups. We first wanted to identify whether there were any differences in the response to the currently available standard chemotherapy treatment based on the combination of cisplatin and pemetrexed. We obtained two expression-based signatures derived from non-small cell lung cancer that predicted resistance to pemetrexed20 and to cisplatin.21 Figure 5A displays GSVA scores for the two signatures across the three MPM immune groups. Overall, the results of the analysis suggested that on average, MPM tumors in IG1 might be more resistant to pemetrexed, while IG3 tumors show a trend to lower resistance. On the other hand, cisplatin resistance analysis showed no differences between immune groups.

Next, we tested a large set of gene-expression signatures predictive of benefit from ICI treatment24–30 to formally evaluate if MPM tumors in IG3 could benefit from an immune-based therapeutic approach (figure 5B). The overlap between these signatures and the classifier is almost non-existent (three shared genes at most). We observed that all signatures were positively correlated with the immune groups (ie, higher GSVA scores in IG3 for upregulated signatures (signatures 1 to 9) and lower GSVA scores in IG3 for downregulated signatures (signatures 10 and 11)), with IG3 tumors presenting a significant association for all but two of the signatures (signatures 9 and 10).

Given the increased activation of cell cycle and DNA repair processes in the IG1 group of MPM tumors, we wanted to explore whether any differences in sensitivity to cyclin-dependent kinase inhibitors and to drugs that inhibit DNA repair among the three groups. To do that, we collected a signature of resistance to palbociclib derived from breast cancer23 and another computationally-derived signature predicting sensitivity to DNA repair inhibition with PARP inhibitors22 (figure 5C). Despite a higher cycling nature of tumors in group IG1, they appear to be more resistant to palbociclib than tumors in other groups, based on this predictive signature. Strikingly, tumors in IG1 showed increased sensitivity to PARP inhibitors compared with the other groups.