Human cancers display the PI signature discovered in mice
Observing PI in mouse tumors prompted us to look for a similar phenomenon in human cancers. To that end, we first analyzed the expression of 40 human homologs of the mouse PI genes in human cancers, utilizing data from the Cancer Cell Line Encyclopedia (CCLE) [22]. Whereas immune and inflammatory genes are normally expressed in hematopoietic cells, a wide range of carcinoma cell lines (?=?634) expressed the PI genes to a higher level than hematopoietic and lymphoid cancer cell lines (?=?180) (Fig. 3a; Additional file 3: Figure S2). Further investigation revealed that, compared with non-PI genes, many PI genes are broadly overexpressed in a subset of the cell lines. To quantify the distribution of PI genes in cell lines, we computed the expression distribution in all the carcinoma cell lines and identified the peak of the distribution. Then, by defining overexpression as twofold expression level over the peak (1 in log2 expression levels), we counted the number of cell lines overexpressing each gene (examples in Fig. 3b). Our analysis revealed that 74.4 % of the PI genes are overexpressed in more than 10 % of the carcinomas, significantly more than a non-discriminatory pool of random inflammatory response genes (40.5 %, chi-square test p value?=?2.9e-5) and all genes (29.6 %, p value?=?1.0e-9) (Fig. 3b). The median number of overexpressing cell lines was 125 (19.7 % of carcinoma samples) for the PI genes compared with 39 (6.2 %, U-test p value?=?2.7e-5) across random inflammatory response genes and only 31 (4.9 %, p value?=?2.6e-8) across random genes.
Parainflammation genes are overexpressed in carcinoma cell lines. a Heatmap of the expression of 39 PI genes in 634 carcinoma cell lines and 180 hematopoietic and lymphoid cancer cell lines from CCLE. One PI gene, IFITM3, is not represented in CCLE. Of the PI genes, 19 show significantly higher expression in carcinomas compared with cancers originated from immune cell types; 10 PI genes are more abundant in immune cancers. b
Left: the distribution of expression of two representative genes across 634 carcinoma cell lines from CCLE. We detected the expression peak and counted the number of samples that express the gene twofold (1 in log2 scale) more than the peak. The top example, the gene CSNK1A1 (CKI?), which here represents a housekeeping gene, is a typical normally distributed gene with low “overexpression” rate and the bottom example, BST2, which is part of the PI signature, shows a gene with a bimodal expression pattern, corresponding to a high “overexpression” rate. Right: the cumulative overexpression rates for the PI genes, all inflammatory response genes, and all genes. Of the PI signature genes, 29 (74.4 %) are overexpressed in at least 10 % of the carcinoma cell lines (?64 samples) compared with 29.6 % and 40.5 % of all and inflammatory response genes, respectively. The yellow curve (PI genes) shows remarkably higher levels of overexpression along the whole graph. c Spread plot of tThe PI score in 634 carcinoma cell lines grouped by tissue types. The dashed blue line differentiates PI+ and PI? samples as defined by the proportion of PI+ of tumor samples
We next computed a score for each cell line by performing single-sample gene set enrichment analysis (ssGSEA) [23] for the PI gene signature. We defined this score, which is an enrichment measure of the overexpression of the PI genes, as the PI score (Additional file 7). The PI score revealed major differences in PI both between cancer types and within cancer types, varying from high levels of PI in head and neck and pancreatic cancer to low levels in samples originating from prostate and liver (Fig. 3c). Thus, the PI expression and PI score patterns suggest that PI occurs across most cancer types, although not uniformly. It is important to note that while the PI score is based on a signature of 40 genes, the PI phenomenon is not restricted to these genes alone but has a wide effect on numerous genes over many different molecular functions (Additional file 8). Finally, we observed a high correlation between the PI score and scores derived from the downregulated PI genes (Spearman R?=?0.654, p value 1e-20; Additional file 3: Figure S3). This once again confirms the validity of our gene sets in human and the relevance of the PI mouse models to a phenomenon that is also observed in human.
Next, we explored the representation of the PI signature in primary human cancers. As differences in PI expression between individual samples may be explained solely by differences in purity levels of the samples [24], we designed a simple adjustment procedure for removing inflammatory gene expression originating from immune infiltrations (Additional file 3: Figure S4). This adjustment procedure consists of two steps: first, utilizing expression data of normal tissues from the Genotype-Tissue Expression (GTEx) project [25], we learned the normal association of gene expression with immune infiltrations in each tissue type. We then used the expression level of PTPRC (CD45), a pan-hematopoietic exclusive marker expressed in all leukocyte cell types but not in other tissues, as an estimate for immune infiltrations. Finally, we applied the gene-specific, tissue-specific learned slopes to adjust the expression levels of The Cancer Genome Atlas (TCGA) tumor samples. This procedure diminishes expression differences among samples which are most likely explained by differences in purity.
We collected gene expression data for 6523 primary tumors and 582 patient-matched normal samples of 18 cancer types from TCGA (Additional file 3: Table S3; http://cancergenome.nih.gov) and performed the adjustment procedure. As in the cancer cell lines, we computed the scores for the PI gene signatures using the adjusted expression of each tumor sample (Additional file 9). Cancer samples demonstrated a wide range of PI scores regardless of CD45 expression (Fig. 4a; Additional file 3: Figure S5). Furthermore, malignant tumors demonstrated significantly higher levels than adjacent normal samples, both overall and across most tested cancer types (Fig. 4b; Additional file 3: Figure S6). As in cell lines, the PI score analyses revealed considerable differences between cancer types and within cancer types (Fig. 4c). We next defined the PI threshold as the score that appears in only 5 % of the adjacent normal samples; cancer samples over the threshold were designated as PI positive (PI+). Strikingly, over all cancer types 25.9 % of the tumor samples were PI+, compared with a null expectation of 5 %, with varying proportions among cancer types, from 77.7 % in pancreatic adenocarcinoma (PAAD) to none in kidney renal clear cell carcinoma (KIRC). The PI score of a sample is highly correlated with the number of upregulated genes in the sample (R?=?0.649; Additional file 3: Figure S7). Accordingly, the median number of PI genes activated in PI+ samples is 17 (42.5 %; compared with eight in PI? samples), the same number we observed to be upregulated in the adenoma organoids. Interestingly, we observed that different sets of PI genes are activated in different cancer types (Fig. 4d). Finally, using the 25.9 % proportion of PI+ samples, we determined a threshold of PI+ samples in the cancer cell lines dataset (where there are no normal samples to determine the threshold). Remarkably, we found high concordance in the abundance of PI+ samples in cancer types between tumors and cell lines (Pearson coefficient R?=?0.875, p value?=?1.9e-4 (Fig. 4e). These results again suggest that the PI gene signature is being expressed by the cancer cells, distinct from immune infiltration.
Pan-cancer parainflammation in human cancers. a PI scores (y-axis) of all 6535 tumor samples (blue) and 582 adjacent normal samples (red) versus the CD45 expression level (x-axis). No correlation is observed between the PI score and CD45 after the adjustment. b PI scores in the tumor samples (blue) and the adjacent normal samples (red). The y-axis is the cumulative percentage of samples over the score in x; 25.9 % of the tumor samples (dashed blue line) are over a threshold which only 5 % of the adjacent normal tissues pass (dashed red line). The PI score is shifted accordingly, so PI+ samples have positive scores. c Spread plot of the PI score in 6535 tumor samples from 18 cancer types. The dashed blue line differentiates PI+ and PI? samples. PAAD pancreatic adenocarcinoma, BLCA bladder carcinoma, HNSC head and neck squamous cell carcinoma, LUAD lung adenocarcinoma, LUSC lung squamous cell carcinoma, COAD colon adenocarcinoma, OV ovarian serous cystadenocarcinoma, UCEC uterine corpus endometrial carcinoma, BRCA breast carcinoma, GBM glioblastoma multiforme, ACC adrenocortical carcinoma, UCS uterine carcinosarcoma, PRAD prostate adenocarcinoma, LIHC liver hepatocellular carcinoma, LGG lower grade glioma, KIRP kidney renal papillary cell carcinoma, KICH kidney chromophobe, KIRC kidney renal clear cell carcinoma. d Heatmap of expression profiles of PI genes across TCGA samples: left, PI+ samples; right, PI? samples. Different subsets of PI genes are expressed in PI+ samples. The expression levels presented are after adjustment to immune infiltrations and are standardized across all samples. e Correlation of the fraction of PI+ samples in each tumor type from TCGA with corresponding tissue origin in CCLE. The same types of cancers have low or high PI+ levels in tumors and in cell lines. The Pearson coefficient is presented
We noted earlier that the PI gene signature is enriched in genes related to the IFN signaling pathway but not to the NF-?B pathway. To test this hypothesis further we first expanded the list of PI genes, identifying 215 genes with expression levels highly correlated with the PI score (R??0.5) in the carcinoma cell lines (Additional file 8). We then performed transcription factor binding enrichment analysis using the ENCODE ChIP-Seq Significance Tool [26] (Additional file 3: Table S4). The analysis again affirmed our claim that PI has a distinct pattern of inflammation, where the top enriched transcription factors were STAT2, IRF1, STAT1 and STAT3, c-Fos, and PRDM1, but only modest enrichment for a key regulator of inflammation such as NF-?B. We further utilized enrichment scores of functional immune gene sets (Additional file 3: Table S5) and correlated them with the PI score in both tumors and cell lines. As expected, this analysis revealed high correlations in all datasets with the type I IFN response, which is a hallmark of PI, but a much weaker association with the tumor necrosis factor and NF-?B signaling pathways (Fig. 5a; Additional file 3: Table S6). We also did not detect any correlation with the immune cytolytic activity metric [27], which is a well-described anti-tumor immunity measure. It should be pointed out that the same correlations were also found in the CCLE dataset, again showing that PI is activated in the tumor and not in its microenvironment. These results again support our claim that PI is distinct from canonical chronic inflammation or other previously described inflammatory responses.
Cancer parainflammation resembles macrophage infiltration. a Heatmap of the Spearman correlations between the PI score and immune functional gene sets across different cancer types. Correlations in CCLE are shown as well (Additional file 3: Tables S5 and S6). b Heatmap of the Spearman correlations between the PI score and the immune subset enrichments calculated using gene sets (Additional files 6 and 9) across different cancer types derived from both TCGA (T) and CCLE (C). Similar correlation trends are observed for a cancer type whether the data were derived from TCGA or CCLE, suggesting that the correlation is due not to association of PI with immune subset presence but to shared functionality with the gene signatures. PAAD pancreatic adenocarcinoma, BLCA bladder carcinoma, HNSC head and neck squamous cell carcinoma, LUAD lung adenocarcinoma, LUSC lung squamous cell carcinoma, COAD colon adenocarcinoma, OV ovarian serous cystadenocarcinoma, UCEC uterine corpus endometrial carcinoma, BRCA breast carcinoma, GBM glioblastoma multiforme, ACC adrenocortical carcinoma, UCS uterine carcinosarcoma, PRAD prostate adenocarcinoma, LIHC liver hepatocellular carcinoma, LGG lower grade glioma, KIRP kidney renal papillary cell carcinoma, KICH kidney chromophobe, KIRC kidney renal clear cell carcinoma
Notably, certain PI genes are members of the Toll-like receptor (TLR) activation pathway (TLR2, CD14, and TIRAP) and, when upregulated, could have mediated an innate immune response to tissue-associated microbiota, igniting conventional inflammation with inflammatory infiltrate, secondary to PI. We therefore hypothesized that PI may enhance the recruitment of certain immune cell subsets to the tumors. To this end, we utilized hematoxylin and eosin (HE) estimations provided by TCGA and gene signature enrichments of immune subset types from Rooney et al. [27] (Additional file 3: Table S5) and associated them with PI scores in tumor samples across cancer types. HE estimations of major immune subtypes (lymphocytes, monocytes, and neutrophils) did not, however, show significant associations with PI scores across different cancer types (Additional file 3: Table S7). We further correlated PI scores of individual tumors with specific immune subsets based on gene signature enrichments in both TCGA and CCLE samples (Fig. 5b; Additional file 3: Table S8). Among the immune subsets the PI score demonstrated highest correlations across tumor types with the macrophage signature (average Spearman coefficient?=?0.362). However, we observed the same trend of correlation between the PI score and macrophages in cell lines (average R?=?0.407). This observation rules out the possibility that PI is dependent on macrophages infiltrating the tumor. Moreover, whereas the role of macrophages in orchestrating PI has been previously suggested [7], our finding supports this inference yet suggests that the tumors themselves may fulfill the macrophage inflammatory function in PI by expressing macrophage-relevant genes. Importantly, we did not observe any correlation with CD8+ enrichment, which is the main component of the “immunoscore” [28], thus suggesting that PI may represent a different immunotype, which, similarly to the immunoscore, may serve as a clinical parameter in evaluating tumorigenesis.
Thus, the PI signature is widely expressed in human tumors, distinguishing the tumor cells from adjacent normal tissue and the tumor microenvironment, with certain cancer types having stronger PI signatures than others. PI appears to be primarily a cancer cell-autonomous phenomenon, distinct from all other well-established cancer-promoting immune inflammatory responses.