An extensive bioinformatics study on the role of mitochondrial solute carrier family 25 in PC and its mechanism behind affecting immune infiltration and tumor energy metabolism

The TCGA-PAAD cohort was collected (177 tumor tissues and 4 normal tissues) from the TCGA database, and 167 normal pancreatic tissue samples were pooled from the GTEx database. By analyzing the differential expression of genes between PC tumor tissue and normal tissue, a total of 19,712 differentially expressed genes (DEGs) were identified (up-regulated: 10,973, down-regulated: 8739). Finally, 38 DEGs of SLC25A members were screened using the Venn diagram (Fig. 1A). Compared with normal pancreatic tissues, 11 genes were down-regulated and 26 genes were up-regulated in PC tumor tissues, as demonstrated using heatmap (Fig. 1B, Additional file 1: Table. S1).

To investigate the correlation between SLC25A members and the clinical features of PC, the differences in the expression levels of DEGs of SLC25A members among various tumor pathological grades were examined in the TCGA-PAAD cohort. As shown, of the various tumor pathological grades, the expression levels of SLC25A4 (P < 0.01), SLC25A11 (P < 0.0001), SLC25A24 (P < 0.01), SLC25A37 (P < 0.01), SLC25A42 (P < 0.01), SLC25A43 (P < 0.0001), SLC25A44 (P < 0.05) and SLC25A45 (P < 0.05) were significantly different (Fig. 1C). Furthermore, the prognostic value of SLC25 DEGs in PC was investigated. The results showed that patients with PC who showed high expressions of SLC25A4 (P = 0.01), SLC25A6 (P = 3.9e−3), SLC25A11 (P = 3.9e−5), SLC25A14 (P = 1.8e−3), SLC25A27 (P = 1.5e−3), SLC25A29 (P = 3.6e−4), SLC25A34 (P = 8.9e−4), SLC25A42 (P = 9.7e−3), SLC25A44 (P = 3.8e−4), SLC25A45 (P = 4.1e−3) and SLC25A53 (P = 6.3e−5) had better prognosis. In contrast, those with low expressions of SLC25A24 (P = 9.3e−3), SLC25A37 (P = 0.02) and SLC25A43 (P = 3.63e−3) had longer survival (Fig. 1D, Additional file 2: Fig. S1C). Finally, we reveal that the DEGs of SLC25A members significantly correlate with each other (negative: blue; positive: red). Similar to that, there is a considerable protein-level interaction between them (Additional file 2: Fig. S1A, B).

To gain a more comprehensive understanding of the role of SLC25A members in PC, the DEGs were re-clustered based on SLC25A members to identify two novel subtypes while ensuring optimal stability (Fig. S2A, B). Survival analysis revealed that survival was significantly longer in the TCGA-PAAD cohort in Cluster 1 (blue) than in Cluster 2 (red) (Fig. 2A). However, principal component analysis (PCA) indicated a partial overlap between Clusters 1 and 2, and patients with PC could not be distinguished well based on mRNA expression levels (Fig. 2B).

Therefore, a prognostic risk model was constructed to evaluate the role of SLC25A members in PC more comprehensively. The expression data of 14 prognosis-related DEGs of SLC25A members in the TCGA-PAAD cohort were integrated (Fig. 1D, Additional file 2: Fig. S1C), and regression analysis was then performed using a lasso. Subsequently, tenfold crossover validation was done, and finally, 10 genes were screened out (Additional file 1: Table S2). These genes were then subjected to Cox multivariate regression analysis of subsequent stepwise regression, and the optimal model with five genes was obtained (Fig. 2C, Additional file 3: Fig. S2C). The samples were categorized into high- and low-risk groups according to the median of the model’s risk score. Survival analysis alluded that the patients in the high-risk group had shorter survival (P = 5.1e−6) (Fig. 2D). According to the receiver operating characteristic (ROC) curve results, the areas under the curve (AUC) for the 1-, 3- and 5-year overall survival (OS) were 0.72, 0.77 and 0.85, respectively, in the TCGA cohort. This finding suggests that the model has a good predictive value for the survival time of the TCGA cohort (Fig. 2E). The predictive ability of the model for the ICGC-PAAD cohort was validated, and the results showed good predictive performance.

To further elucidate the role of the SLC25A-based prognostic risk model in PC, the infiltration of immune-infiltrating phenotypes was explored in the high- and low-risk groups. The results showed that the immune-infiltrating cells were more abundantly characterized in tumors in the high-risk group than in the low-risk group (Fig. 3B, 3C). It is well known that the main energy metabolism of PC cells is oxygen-consuming glycolysis. Oxidative phosphorylation (OxPhos) and tricarboxylic acid (TCA) cycle pathway scores were greater in the low-risk group than in the high-risk group (Fig. 3A). Amino acid metabolism was further examined, which revealed that alanine_aspartate_glutamate metabolic pathway scores were significantly higher in the low-risk group than in the high-risk group. This finding suggests that SLC25A members may be involved in the metabolic remodeling of PC tumor cells.

After integration of age, gender, tumor grade, and risk score in the TCGA-PAAD and PAAD-CA cohorts, multifactorial Cox regression analysis showed that risk score (P < 0.001) was an independent factor for predicting PC prognosis (Fig. 4B). Calculation of the C-index showed that the model had good discrimination in the TCGA-PAAD [0.678 (95% confidence interval (CI) 0,614–0.747)] and ICGC-PAAD [0.647 (95% CI 0.600–0.694)] cohorts. Moreover, nomogram containing the risk scores were constructed using the clinical characteristics to obtain the predictive values of 1-, 2- and 3-year prognosis of patients with TCGA-PAAD (Fig. 4A). To facilitate scholarly applications, a web version of the dynamic nomogram (minnietree.shinyapps.io/DynNomapp) was deployed. Finally, the calibration curves showed good agreement between the training and validation sets in the predicted prognostic outcomes as well as the actual observed clinical outcomes of patients with PC at 1, 2, and 3 years (Fig. 4C, D).

SLC25A11, SLC25A29, and SLC25A44 were screened for differential expression in PC, with significant prognostic value, in previous univariate and multifactorial Cox regression analyses. Subsequently, separate in-depth analyses of the genes were conducted to explore their functions in PC. Functional module analysis based on LinkedOmics indicated that the transcriptome expression level of SLC25A11 in PC was positively correlated with the expression of 5057 genes but negatively correlated with the expression of 5069 genes (FDR < 0.05); the expression level of SLC25A29 was positively correlated with the expression of 4960 genes and negatively correlated with the expression of 5937 genes (FDR < 0.05); the expression level of SLC25A44 was positively correlated with the expression of 2106 genes and negatively correlated with the expression of 1835 genes (FDR < 0.05) (Additional file 4: Fig. S3A). The top 50 differentially co-expressed genes are portrayed in the heatmap (Additional file 4: Fig. S3B, C, D). Gene Ontology (GO) terminology annotation of gene enrichment analysis illustrated that SLC25A11 and its positively correlated genes were mainly involved in the processes of mitochondrial respiratory chain complex assembly, mitochondrial gene expression, translational elongation, and protein localization to the endoplasmic reticulum. Conversely, the negatively correlated genes were involved in the processes of cell junction organization, cell-substrate adhesion, positive regulation of cell motility, and epidermis development. In the SLC25A29 co-expression network, the positively correlated biological processes were essentially the same as those in SLL25A11, except for adaptive immune response, but it was negatively associated with multiple immune cell activation processes. Finally, SLC25A44 and its positively correlated genes were mainly involved in the synaptic vesicle cycle, multicellular organismal signaling, calcium-ion-regulated exocytosis, and regulation of membrane potential. Furthermore, the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis showed that the co-expressed positively correlated genes SLC25A11 and SLC25A29 were predominantly enriched in the ribosome and OxPhos pathways. SLC25A44 and its positively correlated genes, on the contrary, were enriched in pathways related to the synaptic vesicle cycle, nicotine addiction, maturity-onset diabetes of the young, and synaptic vesicle recycling. The negatively correlated genes SLC25A11, SLC25A29, and SLC25A44 were enriched in adherens junction, focal adhesion, and ECM-receptor interaction. Furthermore, SLC25A29 expression was negatively associated with various pathways, such as Th17, Th1, and Th2 cell differentiation. Additionally, SLC25A44 and its negatively correlated genes were involved in PC, cell cycle, HIF-1, P53, and Hippo pathways (Fig. 5A–C). The above findings imply that SLC25A11, SLC25A29, and SLC25A44 are likely to exert a significant impact on energy metabolism, tumor immunity, proliferation, invasion, and metastasis of PC cells.

The correlation between SLC25A11, SLC25A29, and SLC25A44 mRNA expression levels and tumor immune invasion was further explored in PC. SLC25A11 demonstrated significant positive correlation (Fig. 6a) with CD8 + T cells (Partial. Cor = 0.154, P = 4.39E − 02), CD4 + T cells (Partial. Cor = 0.314, P = 3.22E − 05), macrophages (Partial. Cor = 0.362, P = 1.13E − 06), neutrophils (Partial. P = 5.97E − 03) and dendritic (DC) cells (Partial. Cor = 0.279, P = 2.23E − 04). SLC25A29 presented significant negative correlation with B cells (Partial. Cor =  − 0.172, P = 2.45E − 02), CD8 + T cells (Partial. Cor = -0.31, P = 3.79E − 05), macrophages (Partial. Cor = -0.247, P = 1.15E − 03) and DC cells (Partial. Cor = − 0.282, P = 1.90E − 04). Moreover, a significant positive correlation was observed with the abundance of CD4 + T cells (Partial. Cor = 0.25, P = 1.07E − 03) (Fig. 6b). SLC25A44 was found to be significantly positively correlated (Fig. 6c) with B cells (Partial. Cor = 0.299, P = 7.21E − 05), CD8 + T cells (Partial. Cor = 0.304, P = 5.23E − 05), macrophages (Partial. Cor = 0.36, P = 1.31E − 06), neutrophils (Partial. Cor = 0.17, P = 2.61e − 02) and DC cells (Partial. Cor = 0.26, P = 5.99E − 04). The above results suggest that SLC25A11 and SLC25A44 are involved in antitumor immunity, whereas SLC25A29 is involved in pro-tumor immunity.

To determine the potential mechanisms by which SLC25A11, SLC25A29 and SLC25A44 affect the TME in PC, their expression levels were analyzed in the TME of different cell types using a common scRNA dataset. SLC25A11 and SLC25A29 were noted to be expressed in stromal cells, malignant cells, and immune cells. SLC25A44 was chiefly expressed in stromal cells (Fig. 7a). Interestingly, SLC25A11, SLC25A29, and SLC25A44 were significantly overexpressed in ductal, acinar and endocrine cells in tumor tissues compared with normal tissues (P < 0.001) (Fig. 7b). These findings allude that SLC25A11, SLC25A29, and SLC25A44 are closely related to the formation of the TME in PC.

To further understand the regulatory mechanisms of SLC25A11, SLC25A29, and SLC25A44, three common TFs were predicted using the CHEA3 database, namely, Top10 TFs of Mean Rank (SCX, NK2 Homeobox 1 (Nkx2-1) et al.), which are most likely to be involved in the TFs of the above mRNA (Additional file 1: Table. S3). The protein interaction network between TFs and the target genes was analyzed using the GeneMANIA website (Fig. 8b).

Finally, the sensitivity of SLC25A11, SLC25A29, and SLC25A44 to antitumor drugs was evaluated. The associations between SLC25A11, SLC25A29, and SLC25A44 expression levels and 20 current Food and Drug Administration-approved anticancer drugs were assessed. The results showed (Fig. 8a) that the sensitivity to SNS-314 increased with the up-regulation of SLC25A11. The sensitivity to cisplatin and nellabine increased with the up-regulation of SLC25A29. Similarly, the sensitivity to XL-147, ZM-336372, and OSI-027 increased with the up-regulation of SLC25A44. Furthermore, the resistance to stellaria and dasatinib increased with the up-regulation of SLC25A11. The resistance to the drug trametinib increased with the up-regulation of SLC25A29, and the resistance to the drug bugatinib increased with the up-regulation of SLC25A11 and SLC25A29.

We collected RNA-Seq (FPKM) and somatic mutation data on PC samples (TCGA-PAAD) from the TCGA database (https://​portal.​gdc.​cancer.​gov/​). For pairing with TCGA samples, we downloaded the normal pancreas sample sequencing data from the Genotype-Tissue Expression (GTEx) project (https://​xenabrowser.​net/​hub/​) and used Toil Pipeline downloaded from UCSC Xena for homogenization and unification. Furthermore, we collected PC samples (ICGC-PAAD) from the ICGC database (https://​daco.​icgc.​org/​) as the validation set. Finally, the RNA-Seq data were uniformly normalized [log2(X + 1)]. The samples with incomplete clinical data were censored.

We used the R software package limma [46] to analyze the gene expression differences between the TCGA-PAAD PC tumor group and the normal tissue group using Log2 [Fold Change (FC)] > 2; P < 0.05; FDR < 0.05 as the standard to identify DEGs of SLC25A members.

Interactions between the protein levels of DEGs of SLC25A members were assessed using STRING (http://​string-db.​org) [47]. DEGs among different tumor tissue stages were detected by ANOVA or two-sided Wilcoxon rank-sum test. Finally, according to the median expression level of DEGs, the SLC25A members were assigned to 2 groups: high and low—and their prognostic value in TCGA-PAAD was analyzed by Kaplan–Meier analysis.

First, we used the R package NMF [48] to undertake a hierarchical clustering analysis based on the DEGs of SLC25A members to identify new classes in PC. Subsequently, PCA of the new classifications was performed, and its prognostic role in PC was analyzed by Kaplan–Meier. Furthermore, we constructed a prognostic risk model for DEGs. Specifically, DEGs of SLC25A members with prognostic values were analyzed through univariate analysis. Integration of the survival time, survival status, and gene expression data was followed by regression analysis using the Lasso-Cox method. In addition, we set up tenfold cross-validation to obtain the optimal model. Finally, we used the Cox multivariate stepwise regression method to filter the final model.

Based on Bindea's research, we collected 29 gene sets that marked different tumor-infiltrating immune cell types [49]. The single-sample gene set enrichment analysis (ssGSEA) algorithm in the R GSVA package was performed to quantify the level of immune cell infiltration in each PC sample [50]. Subsequently, we obtained gene sets for OxPhos, TCA cycle, and alanine_aspartate_glutamate metabolism from the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway of GSEA [51]. ssGSEA was used to quantify the activity of each metabolic pathway by scoring and assessing the differences between the high- and low-risk groups.

To validate the clinical efficacy of the prognostic risk model for DEGs, we evaluated the risk scores in the TCGA-PAAD cohort through univariate and multivariate Cox regression analyses, and nomograms were constructed to assess the prognostic significance of these features. In addition, we built a dynamic nomogram based on the R package Dynnom, which forms an interactive interface for clinical applications, and deployed it on web pages [52]. The ROC analysis was performed using the R package pROC to obtain the AUC of the clinical prediction model. Finally, the survival significance of the model was analyzed by Kaplan–Meier [53].

We used the CellMiner database (https://​discover.​nci.​nih.​gov/​cell-miner/​home.​do) to investigate the expression levels of gene transcriptomes in NCI-60 cell lines and their sensitivity to antitumor drugs [54].

We analyzed the DEGs associated with the key SLC25A members transcription in PC based on the LinkedOmics (http://​www.​linkedomics.​org/​login.​php) functional module. Then, we used Gene Set Variation Analysis (GSVA) to analyze the biological processes and signaling pathways, where the DEGs were concentrated.

We also analyzed the differences in the expression levels of SLC25A11, SLC25A29, SLC25A44, and the Spearman correlation of different immune-cell infiltration levels in PC using the TIMER (http://​cistrome.​shinyapps.​io/​timer/​) database (purity-corrected) [55]. The Tumor Immunity Single-Cell Database (TISCH) (http://​tisch.​comp-genomics.​org/​) is a comprehensive scRNA dataset that enables interactive single-cell transcriptome visualization of the tumor microenvironment [56]. In the Datasets module, we visualized the expression levels of SLC25A members at the single-cell level in the PAAD_CRA001160 dataset. Furthermore, differences in the SLC25A members expression in immune cells between the tumor and normal groups were investigated.

TFs prediction was performed by ChIP-X enrichment analysis 3 (ChEA3) (https://​maayanlab.​cloud/​chea3/​) [57]. The ChEA3 database, which contains genomic libraries generated from a variety of sources, is a TFs-enrichment analysis tool that ranks TFs associated with user-submitted genomes.

Data were analyzed using R v 4.0.3. ANOVA was used to estimate the statistical significance of differences between the different tumor tissue stages for DEGs that conformed to a normal distribution or a two-sided Wilcoxon rank-sum test for non-normal distribution. The Chi-square test or Fisher's exact test was performed to investigate the statistical significance of differences between the categorical variables. Spearman's correlation test was employed to assess two gene correlations as well as gene and drug correlations. One-way and multi-way Cox regression analyses were performed to determine independent prognostic factors. The survival outcomes were assessed using Kaplan–Meier curves, with a two-sided P < 0.05 considered as the threshold of significance.