A comprehensive study of alternative splicing in malignant pleural mesothelioma identifies potential therapeutic targets in a new cluster with poor survival

Cell culture and SRPK1 knockdown by SRPK1-specific siRNA

The human malignant pleural mesothelioma NCI-H2452 cell line was purchased from FuHeng BioLogy (Shanghai) and cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS), 100 mg/ml streptomycin, and 100 U/ml penicillin. The cells were cultured at 37°C in the presence of 5% CO₂. SRPK1-specific siRNAs were synthesized by RIBOBIO, Guangzhou. We designed two distinct siRNAs specific to SRPK1 with the following target sequences: GCCGTATCATCCACACTGA (siSRPK1-1) and GCGCCAGGCAGAATTACTA (siSRPK1-2). An equimolar mixture of the two siRNAs was used for the following assay. The RNAiMAX reagent was used to transfect H2452 cells with siSRPK1 and siControl following the manufacturer's instructions. The knockdown efficiency was verified by RT-PCR, using the following SRPK1-specific primers:

SRPK1-Forward: GGGCATCATCTGCTCAAGTGGA;

SRPK1-Reverse: GTCAGTGTGGATGATACGGCAC.

Apoptosis assay

The cells were collected 48 h after transfection. Cell apoptosis was detected using the FITC annexin V apoptosis detection kit (BD) following the manufacturer's instructions. Labeled cells were then subjected to flow cytometry (BD) to calculate the apoptotic cell fraction.

Transwell cell migration assay

Cell migration was analyzed using 8-μm pore filter chambers (Costar). The cells were collected 48 h after transfection, counted, and diluted to a density of 5.0 × 10⁵ cells/ml in PMI-1640 medium without FBS. The chambers were filled with 100 μl cell suspension and 650 μl RPMI-1640 medium supplemented with 10% FBS, and transferred to the plates. After 12 h, the cells that adhered to the membrane were fixed with methanol and stained with 0.05% crystal violet. Images of cells that migrated to the underside of the filter were acquired with a microscope and analyzed using ImageJ software.

Data sources and processing

RNA sequencing data and somatic mutation data from 84 MPM tumor tissues from the Cancer Genome Atlas (TCGA) database, processed by MuTect2, were downloaded from the UCSC Xena website (https://xena.ucsc.edu/). Clinical characteristics (age, gender, tumor node metastasis classification [TNM], and pathological type and stage) and survival information for all patients were collected from the TCGA database. Gene expression data were converted into transcripts per kilobase million (TPM), and log2(TPM + 0.01) was used for subsequent analysis. Neoantigen data of these TCGA MPM patients was obtained from the Genomic Data Commons website (http://api.gdc.cancer.gov/data/).

The microarray GSE42977 data, which were acquired on the GPL6790 platform, were downloaded from the Gene Expression Omnibus website (GEO, https://www.ncbi.nlm.nih.gov/geo/), and 40 MPM tumor tissues and eight normal pleura mesothelial tissues were used to calculate the differential gene expression between tumor and normal mesothelial tissues.

Percent-spliced-in (PSI) values of alternative splicing data (percentage of samples with PSI of more than 75%) were collected from SpliceSeq (http://bioinformatics.mdanderson.org/TCGASpliceSeq).¹⁸ A total of 328 and splicing-related genes were extracted from the SpliceAid 2 website (www.introni.it/spliceaid.html).¹⁹

MPM clustering analysis

Dimensionality reduction for RNA alternative splicing data using the t-distributed stochastic neighbor embedding (t-SNE) method was performed using the Rtsne package (version 0.15) using the default parameters, and visualization was performed by ggplot2 (version 3.3.5) in R software. To obtain a robust classification of MPM patients based on alternative splicing data, the consensus cluster plus package (version 1.56.0) was used under the implemented unsupervised consensus approach to identify MPM splicing subtypes. The patients were separated into four clusters (clusters 1, 2, 3, and 4).

Differential gene analysis and gene set variation analysis (GSVA) analysis

The Limma package (version 3.48.3) was used to identify differentially expressed genes and alternative splicing events between tumor and nontumor tissues, and between the four clusters. A fold change higher than 2 or lower than 0.5 and a false discovery rate (FDR) value lower than 0.05 were set as critical values for differential gene analysis. GSVA, a gene set enrichment method that estimates the variation of pathway activity over a sample population in an unsupervised manner, was performed using the GSVA package (version 1.42.0) on TPM RNA sequencing data.²⁰ The PSI value of each alternative splicing event with an interquartile range (IQR) greater than 0.05 and the mRNA expression levels of splicing factors were used for correlation analysis using the Hmisc package (version 4.6–0) in order to calculate the GSVA score of each cluster.

Survival analysis

Kaplan–Meier analysis was used to evaluate the overall survival of MPM patients in the four clusters. In order to further analyze whether classification in the splicing clusters could serve as an independent prognostic factor, we used the Cox proportional hazards model after adjusting for factors including gender, age, pathological type, TNM, stage, and CDKN2A mutation status to determine whether the splicing classification could predict survival rate of MPM patients.

Immunological score

The geometric mean of the expression of GZMA, GZMB, PRF1, GNLY, GZMH, GZMM, GZMA, and PRF, which are directly correlated with the cytolytic activity of T/NK cells, was defined as the cytolytic score. The geometric mean of the expression of HLA II genes (HLA-DMA, HLA-DMB, HLA-DPA1, HLA-DPB1, HLA-DRA, and HLA-DRB1) was defined as the HLA II score. The geometric mean of the expression of known HLA I genes (B2M, HLA-A, HLA-B, and HLA-C) was defined as the HLA I score.²¹

Tumor-infiltrating immune cell analysis

EPIC and CIBERSORT were used to calculate the infiltration fraction of seven and 22 immune cell types, respectively, in MPM tumor tissues. EPIC is based on the comparison of gene expression levels between a library of specific cell types and tumors, and can be used to predict the fraction of each cell type.²² EPIC analysis was performed online according to web instructions (http://epic.gfellerlab.org/). CIBERSORT is based on the principle of linear support vector regression and is a computational tool for the deconvolution of the expression matrix of human immune cell subtypes. CIBERSORT analysis was performed using the R software according to the authors' instructions.²³

Somatic mutation analysis

Somatic mutation data of MPM patients were processed by MuTect2, and analyzed and visualized using the maftools package (version 2.10.0).

Statistical analysis

All statistical analysis were performed using the R (version 4.0.2) and Graphpad Prism 8 software. The data analysis and visualization toolkit in R software included ggplot2, pheatmap, Rtsne, Upset, ConsensusClusterPlus, survminer, survival, limma, Cibersort, maftool, ComplexHeatmap, openxlsx, circlize, and pROC packages. All statistical tests are two-sided. Unless otherwise stated, a p-value less than 0.05 was considered to denote statistical significance.

SRPK1 deficiency promoted MPM cell apoptosis and inhibited cell migration

The activities of two key families of splicing factors, the family of serine/arginine splicing activators and the heterogeneous ribonuclear protein family of splicing suppressors, are both regulated by the SRPK protein kinase family.^{24, 25} SRPK1 expression is dysregulated in most cancers, and is higher in tissues from MPM patients than in their healthy counterparts (Figure S1A and Figure 1a). The Kaplan–Meier curves show that MPM patients with higher SRPK1 expression have a worse prognosis than those with lower SRPK1 levels (Figure 1b, p < 0.01). To verify the effects of SRPK1 in cancer in vitro, we carried out SRPK1 knockdown experiments in the human MPM NCI-H2452 cell line using SRPK1-specific small interfering RNAs. We confirmed that siSRPK1 transfection dramatically decreased SRPK1 expression in H2452 cells (Figure 1c). Compared to the control group, SRPK1 deficiency significantly promoted H2452 cell apoptosis and inhibited H2452 cell migration (Figure 1d and e). Collectively, our data support the notion that SRPK1 and RNA splicing-related events play a vital role in MPM patients. Although MPM histological types are an important factor affecting patient survival, SRPK1 expression showed no significant differences across the different histological types for both TCGA and GEO data sets (Figure 1f; TCGA data set, p = 0.54; GEO data set, p = 0.16). Evaluation of the splicing profile in MPM patients showed that most splicing events were shared among the three histological types (40 883/43433, Figure S1B). t-SNE analysis also showed that pooled samples from different histological types did not form discrete clusters (Figure 1g), suggesting that when analyzing MPM characteristics based on alternative splicing events, a new classification method might be needed.

Subtype clustering and clinical features related to MPM alternative splicing

We conducted a detailed analysis of the comprehensive splicing profile in TCGA MPM patients, covering seven typical splicing types: alternative acceptor site (AA), alternative donor site (AD), alternative promoter (AP), alternative terminator (AT), exon skipping (ES), mutually exclusive exons (ME), and retained intron (RI). A total of 43 433 alternative splicing events in 10 144 genes were detected in the 84 MPM patient cohort, comprising 7.99% AA, 7.07% AD, 20.13% AP, 19.51% AT, 38.53% ES, 0.42% ME, and 6.35% RI (Supplementary Table 1, https://figshare.com/s/e7ce15a90f7e235b5da9). We used the k-means clustering algorithm to perform an unsupervised analysis on MPM patients according to their alternative splicing profile and the elbow method to determine the optimal number of clusters. Although k = 5 appeared to be the optimal choice (Figure 2a), the fifth cluster contained only two samples, prompting us to eventually divide the patients into four clusters only: Cluster 1 (C1, n = 26, 31.0%), cluster 2 (C2, n = 17, 20.2%), cluster 3 (C3, n = 29, 34.5%), and cluster 4 (C4, n = 12, 14.3%) (Figure 2b and Figure S2A). The two samples in the original cluster 5 (when k = 5), were grouped in cluster 1 in the final classification. All 12 samples in cluster 4 belonged to epithelioid histological types and were associated with the best overall survival. Two of the five sarcomatoid samples were classified into cluster 3, and the remaining three were found in cluster 1. In cluster 2, except for one biphasic sample, the remaining 16 samples were all epithelial samples. Of the remaining 20 histological biphasic samples, nine were found in cluster 1 and 11 in cluster 3. Cluster 3 was associated with a worse overall survival rate compared to the other clusters (Figure 2c and d). Survival rates showed statistically significant differences across the four clusters (p < 0.001). Even after adjusting for histology and the CDKN2A mutation status,²⁶ two known molecular prognostic factors in MPM (Figure S2B), the survival differences were still significant (p < 0.01). Furthermore, the analysis of relevant clinical information (gender, age, T, N, and stage) showed significant differences in T, N, and stage that were not randomly distributed across the four clusters (Fisher's exact test, p < 0.05, Figure 2e). In conclusion, through unsupervised clustering based on alternative splicing profiles, we divided MPM patients into four clusters with different survival and clinical characteristics.

GSVA analysis in MPM clusters

Comprehensive analysis of splicing profiles in the above four clusters of MPM patients identified 41 861 alternative splicing events in 19 819 genes in cluster 1, 41 871 alternative splicing events in 19 814 genes in cluster 2, 42 056 alternative splicing events in 19 833 genes in cluster 3, 41 536 alternative splicing events in 19 801 genes in cluster 4, and 41 271 alternative splicing events shared by all four clusters (Figure 3a and b). Although the number of genes and alternative splicing events appeared to be similar across these four clusters, GSVA analysis revealed unique characteristics for each cluster at the transcriptome level. GSVA analysis uses gene expression profile information to evaluate changes in pathway activity in order to better annotate biological functions.²⁰ Here, we performed GSVA analysis on all MPM samples using a 7481-gene set from gene ontology (GO: biological process). Cluster 1 showed a high level of oxidative phosphorylation characteristics, with enhanced signals for ATP synthesis-coupled electron transport, mitochondrial electron transport, NADH dehydrogenase complex, and oxidative phosphorylation, and decreased signals for migration and tyrosine kinase receptor signaling pathways. Cluster 2, when compared with the other three clusters, had stronger signals related to complement-dependent cytotoxicity and fatty acid beta oxidation using Acyl-CoA oxidase, while the activation level of mononuclear macrophages and the synthesis of deoxyribonucleotide and pyrimidine were downregulated. It is worth noting that cluster 3, associated with the worst prognosis, was enriched in epithelial-to-mesenchymal transition, DNA replication, negative regulation of cell aging, and other related pathways that constitute cancer hallmarks,²⁷ showing strong malignant characteristics. Cluster 4, the cluster with the longest patient survival time, showed type I interferon response characteristics (Figure 3c and Figure S3). We found that the genes encoding PDGFRB, CSPG4, and COL1A2, which are highly expressed in MPM and correlate with MPM patient prognosis as reported earlier, were almost exclusively expressed in cluster 3 (Figure 3d).^28-30 All cluster 4 patients and 55% of cluster 3 patients belonged to the epithelioid pathology type, but showed significant differences in survival time. Therefore, we explored the differences between cluster 3 and cluster 4 only in epithelial-type patients and found that they exhibited similar overall characteristics, which indicated that the difference in pathological type was not so important after classification according to alternative splicing clusters. In summary, cluster 3, which was associated with the shortest patient survival time among all clusters, was enriched in antiapoptosis and EMT signaling pathways, which are frequent phenomena in MPM, whereas cluster 4, associated with the best patient survival, showed stronger type I interferon characteristics (Figure 3e). These data show that there was an enrichment of immune signals such as type I IFN in cluster 4 patients with longer survival, whereas cluster 3 patients with the worst survival showed increased EMT and antiapoptotic signals.

Correlation between alternative splicing events and splicing factors in MPM subtypes

In order to study the correlation between splicing factors and alternative splicing events, we further performed a Spearman correlation analysis on samples from the four clusters and identified highly expressed splicing factors in each cluster (Figure S4A). We found that in cluster 1, patients could clearly be divided into two subgroups based either on the expression of splicing factors or their correlation with alternative splicing events (Figure 4a and Figure S4B). Such classification was not obvious in clusters 2 and 3, and almost invisible in cluster 4. In cluster 1, RNA splicing genes with increased expression (PQBP1, THOC6, LSM7, FAM50A, and C19orf43) were negatively correlated with a greater number of alternative splicing events, and genes with decreased expression (ZC3H11A, RNF40, INTS3, RAVER1, and EIF3A) showed a strong positive correlation with alternative splicing events. In cluster 4, a similar number of alternative splicing events was involved in positive and negative gene regulation (Figure 4b). These observations are consistent with the heatmap results presented in Figure 4a. These results partially illustrate the differences in splicing regulation mechanisms among MPM clusters.

MPM patients in cluster 3 were sensitive to the GSK-3 inhibitor CHIR-99021

Although the alternative splicing profile could help us to clearly divide MPM patients into four clusters, this classification method is not feasible for practical applications. Therefore, we next asked whether specific gene expression patterns could distinguish between four clusters of patients. We established a marker signature based on 3–6 genes for each cluster (Figure S5). Our marker model achieved a sensitivity level of 88.5, 82.4, 89.7, and 91.7% and a specificity level of 84.5, 94.0, 85.5, and 94.4% for cluster 1, cluster 2, cluster 3, and cluster 4, respectively. Receiver-operating characteristics (ROC) showed that the area under the curve (AUCs) of the model was 0.912, 0.942, 0.956, and 0.983 for the four clusters, respectively, and all models exhibited good prediction effects (Figure 5a). We used this marker model to cluster the 84 MPM patients to further verify the classification effect (Figure 5b).

The development of new inhibitors for targeted therapy of mesothelioma is particularly important for improving patient survival, especially in cluster 3 patients, who have poor prognosis. The Genomics of Drug Sensitivity in Cancer (GDSC) database contains multiple omics data from more than 1000 tumor cell lines and responses to more than 200 drugs, and helps researchers to explore potential tumor treatment targets.³¹ We collected transcriptome data of 21 pleural mesothelioma cell lines in GDSC and clustered them together with the data of the 84 MPM patients from the TCGA database based on the expression of the marker signature. Pleural mesothelioma cells in GDSC were divided into two main groups: 16 cell lines highly similar to cluster 3 were labeled as GDSC-1, and the remaining five cell lines were labeled as GDSC-2 (Figure 5c). We selected 155 small molecule drugs with response records in at least 18 mesothelioma cell lines for the follow-up analysis. There are only three drugs, CHIR-99021, IOX2, and rTRAIL, whose responses were recorded as sensitive in more than five cell lines (Figure 5d-g). CHIR-99021 is an inhibitor of GSK-3α and GSK-3β, which are involved in the WNT signaling pathway. IOX2 is an inhibitor of the egl-9 family hypoxia-inducible factor 1 (EGLN). rTRAIL is a TRAIL receptor agonist and is involved in the regulation of apoptosis regulatory signaling pathways. We found that 50% of the cell lines in GDSC-1 were sensitive to CHIR-99021, whereas all cell lines in GDSC-2 were resistant to CHIR-99021, and there was a significant difference in IC50 between the sensitive and resistant groups. These results suggest that CHIR-99021, a small molecule inhibitor of GSK-3α and GSK-3β, would elicit a good therapeutic response in cluster 3 MPM patients but have no effect for MPM patients in the other three clusters. Personalized cancer treatment appears to be particularly important in the current customization of cancer treatment plans.

Immune characteristics of MPM patients

We next explored the characteristics of immune cell infiltration, antigen presentation, and cancer-germline antigens (CGAs) in MPM. CD8⁺ cytotoxic T lymphocytes (CTL) and NK cells are considered essential for effective immunotherapy. We used the EPIC and CIBERSORT algorithms to calculate the infiltration fraction of different immune cells in MPM (Figure S6A), and calculated the cytolytic score for each sample, which could reflect CTL/NK abundance.²¹ CD8⁺ T cell scores calculated by both EPIC and CIBERSORT and the combined scores of CD8⁺ T cells and NK cells calculated by CIBERSORT were correlated with higher cytolytic scores (r > 0.6), while the CD4⁺ T cell scores showed a weak correlation with cytolytic scores (Figure 6a and Figure S6B). The cytolytic score was highest in cluster 4 and lowest in cluster 3 (one-way ANOVA test, p = 0.014, Figure 6b). We also analyzed the scores of HLA I and HLA II in each cluster to evaluate the transcription levels of immune evasion drivers in each of them. The HLA I scores were basically the same in all four clusters (one-way ANOVA test, p = 0.993). The HLA II levels in clusters 2 and 3 were lower (one-way ANOVA test, p = 0.016), which indicated the downregulation of HLA II gene transcription and the avoidance of antigen presentation (Figure 6b). However, the cytolytic, HLA I, and HLA II scores did not differ significantly between histopathological subtypes (Figure S6C).

We next investigated the expression of ligands of T/NK cell costimulatory and coinhibitory receptors and other immunomodulators to identify potential targetable immune checkpoints in MPM. The inhibitory ligand CD274 (PD-L1) was preferentially expressed in 10 samples in cluster 3, showing lower expression levels in the other samples (Figure 6c). Other inhibitory checkpoint ligands, such as LGALS9 (TIM-3 ligand) and NK cell inhibitory receptor KLRK1 ligands (MICA, MICB, ULBP1, ULBP2, and ULBP3) were also highly expressed in cluster 3. All samples in cluster 4 highly expressed C10orf54, which encodes an inhibitory T cell checkpoint of the B7 family.

We further conducted a systematic study on the expression of CGA in MPM. MPM tumor tissues showed significantly higher expression of 11 CGAs than normal mesothelial tissues (Figure 6d). Each cluster showed a specific CGA expression pattern (Figure 6e). The expression levels of OIP5, ODF2, and ATAD2 were higher in cluster 3 than in the other clusters, whereas TSSK6 expression was lower. In conclusion, among all MPM patients, cluster 3 presented the lowest cytolytic score and enrichment in a series of inhibitory receptors, which may be a reason for the shortest survival time of patients in this cluster.

We ranked cytolytic, HLA I, and HLA II scores from high to low, and found no significant change in mutated genes and mutation frequencies in MPM patients in all four clusters (Figure S7A and S7B). Meanwhile, there were no significant differences between mutated genes and mutation frequencies and the number of neoantigens in all four clusters (Figure S7C, one-way ANOVA test, p = 0.68).