Gene expression profiles in genome instability-based classes of colorectal cancer

Thirty-three tumor samples (clinicopathological data in Additional file 1: Table S1) were collected from a cohort of 32 patients who underwent resection of primary invasive CRC at “Centro Clinico Diagnostico S.r.l. G.B. Morgagni” in Catania (Italy). All patients gave informed consent for this study, which was approved by the Ethics Committee of ASL3 of Catania (Italy). All specimens were frozen and stored at − 80 °C until DNA and RNA extraction. In one case, a synchronous tumor located in another site of the colon was present and biopsied. The number of individual tumors, taking into account the patient with two synchronous tumors (considered as separated entities) was of 33. A biopsy of adjacent phenotypically normal colonic tissue (at a distance of 3–6 cm from the tumor) was taken for 25 patients (tumor/normal pairs).

Samples were tested for microsatellite instability with five markers belonging to the Bethesda panel (D2S123, D5S346, D17S250, BAT25 and BAT26) and one additional marker (BAT40) [5]. Tumors were defined as MSI if ≥30% markers were found unstable when comparing tumor versus normal colonic tissue. 5/32 (15.6%) patients had MSI tumors, whereas 27/32 (84.4%) patients had MSS tumors.

Genome-wide DNA copy number and SNP genotyping analysis were performed in 33 tumor samples and 25 normal tissue samples on Affymetrix SNP 6.0 arrays (Affymetrix, Inc., Santa Clara, CA, USA), using 500 ng of input DNA. Array scanning and data analysis were performed by Affymetrix® “GeneChip Command Console” (AGCC) and “Genotyping Console™” (GTC) version 3.0.1 software [13, 14] . Broad copy number abnormalities (BCNAs), defined as gains or losses involving more than 25% of a chromosomal arm or numerical aberrations involving whole chromosomes, were identified by using a bioinformatic tool called BroCyA, as described by Barresi et al. (2017). Briefly, an estimate of copy number for each DNA marker in each sample was obtained by calculating the log2ratio between fluorescent intensity signal in the sample and the corresponding median value in the reference group composed of 270 HapMap individuals. Log2ratio of each DNA marker is calibrated assuming the reference group value equal to 2 (diploid genome). DNA markers in autosomes and in female X chromosomes with calibrated log2ratios greater than 2.21 (average+ 2 S.D. of a control normal diploid group) have been considered as gains and markers with values less than 1.74 (average – 2S.D.) as losses. For the X and Y male chromosomes 1.24 and 0.81 were used as upper and lower limit, respectively. The algorithm searched all short segments, formed by at least 50 contiguous markers showing the same type of copy number variation (short isosegments) and then joined them in larger segments (BCNAs) and retained them in the final list if the following conditions were met: 1) their intersegment distance is lower than an established threshold [5], 2) the entire BCNA (short isosegments and intersegment intervals) has a mean calibrated log2ratio higher than 2.21 or lower than 1.74 for broad gains and losses, respectively, 3) the entire BCNA has a physical size higher than 25% of a chromosomal arm. Finally, a score was attributed to each chromosomal arm or to whole chromosome according to the following rules:

The sum of the scores per tumor sample corresponds to the “BCNA score”.

The logic behind our BCNA scoring system (p-q-w system) is to evaluate as a single mutational event (with a score of 1) the lost or the gain of a whole chromosome. For instance, in the p-q-w scoring system “n” segmental aneuploidies get a “BCNA score” equal to that of “n” whole chromosome aberrations. We are aware that other scoring system systems count p and q aberrations as separate events, even in the case of whole chromosome aberrations. We also recalculated BCNA scores according to a p-q scoring system (attributing a score of 2 to a whole chromosome aberration). Although the absolute values of BCNA scores are different in the two systems the subdivision of MSS tumors in two classes, as reported in the result section, is not affected by the choice of the scoring system.

Raw and processed data of SNP 6.0-array results have been submitted to public repository: “Gene Expression Omnibus-GEO” (www.​ncbi.​nlm.​nih.​gov/​geo) with the following accession number: GSE80460 [5]. Data of the most frequent DNA copy number changes in a cohort of 27 MSI samples were available from a study by Sveen et al. (2017) [15].

Whole-transcript expression analysis was performed from 100 ng of total RNA by amplification and target hybridization to the Gene-Chip Human Transcriptome Array (HTA) 2.0 (Affymetrix, Inc., Santa Clara, CA, USA), as previously described [16]. Array scanning and data analysis were performed by Affymetrix® Expression Console™ software version 1.4 and Affymetrix® Transcriptome Analysis Console (TAC v3) software (Affymetrix, Inc., Santa Clara, CA, USA). Transcript level analysis was performed using the normalization method based on the processing algorithm called robust multi-array average (RMA). Such RMA values are log₂ values. Average RMA values have been transformed in linear values and their ratios (linear fold changes) have been used in order to estimate differential expression between CRC groups and normal colon group. Fold changes < 1 have been reported as the negative of the reciprocal, so that e.g. a fold change of ½ is reported as − 2. Therefore, linear fold-changes (denominated FC in the rest of the text) were calculated in the following way: 2^{[CRC group Average RMA – Normal Colon group average RMA]} if CRC group > Normal Colon group, or − 2^{[Normal Colon group Average RMA – CRC group average RMA]} if CRC group < Normal Colon group. Statistical analysis of differential gene expression was performed as implemented in the TAC software using one-way ANOVA analysis and p-value correction for multiple testing according to Benjamini-Hochberg [17].

The data have been deposited to public repository: ‘Gene Expression Omnibus-GEO’ (www.​ncbi.​nlm.​nih.​gov/​geo) and are accessible through GEO: GSE73360 [16] and GSE84984.

Since in our series of 33 tumor samples there were only 5 MSI tumors, we did not include such samples for the expression analysis (TAC software), and used instead genome-wide HTA data of 34 MSI samples from Sveen et al. (2017) [15], deposited by the authors on the NCBI Gene Expression Omnibus (GEO) with the accession number GSE79959.

When analyzing whole-transcript HTA data, we ruled out genes with no gene symbol assigned by Affymetrix, genes which had a gene symbol of the type “OTTHUMG###”, genes described by Affymetrix as “uncharacterized LOC” (except for those with more than one description, the first one only beginning with “uncharacterized LOC”), genes encoding small nucleolar RNAs (snoRNAs), small Cajal body-specific RNAs (scaRNAs), small nuclear RNAs (snRNAs), small NF90-associated RNAs, RNA 5S ribosomal genes and pseudogenes, RNA 5.8S ribosomal pseudogenes, Y RNAs, mitochondrially encoded ribosomal RNAs (MT-RNR), microRNAs, olfactory receptor and histone cluster genes. We also did not consider transcripts on chr1_gl000191_random, chr4_ctg9_hap1, chr4_gl000193_random, chr4_gl000194_random, chr6_apd_hap1, chr6_cox_hap2, chr6_dbb_hap3, chr6_mann_hap4, chr6_mcf_hap5, chr6_qbl_hap6, chr6_ssto_hap7, chr7_gl000195_random, chr17_ctg5_hap1, chr17_gl000204_random, chr19_gl000209_random, chrUn_gl000211, chrUn_gl000212, chrUn_gl000218, chrUn_gl000219, chrUn_gl000220, chrUn_gl000222, chrUn_gl000223, chrUn_gl000228.

In their paper on the consensus molecular subtypes (CMS) of colorectal cancer, Guinney et al. (2015) [12] provided a downloadable R package (CMSclassifier), which included the Random Forest classifier and the Single Sample Predictor (SSP) classifier. SSP gives a prediction of the CMS of a tumor sample regardless of whether it is analyzed alone or within a series of samples. Our input data were formatted as requested by the software instructions (https://​github.​com/​Sage-Bionetworks/​CMSclassifier/​blob/​master/​README.​md) and included 5969 of the 5973 genes of the example data set. The 4 missing genes were due to the fact that in the transcript cluster (TC)-to-Entrez ID conversion table provided by Affymetrix NetAffx™ Analysis Center website (https://​www.​affymetrix.​com/), 4 TCs had two Entrez IDs each, and both the IDs were included in the example data set of the CMSclassifier R package as unique entries. In particular, in the Affymetrix conversion table, IDs 701 and 56,924 were both assigned to TC15000276.hg.1; IDs 6038 and 283 to TC14000067.hg.1; IDs 6560 and 3931 to TC16001199.hg.1; IDs 54,741 and 3953 to TC01000730.hg.1. Therefore, IDs 701, 6038, 6560, 54,741 were ruled out, in order to provide a unique Entrez ID for each transcript cluster.

MSS CRCs were subdivided in two groups according to the distribution of the BCNA scores: a low-BCNA (LB) group bearing a BCNA score lower than the first quartile (5.75) of the distribution (7 tumor samples from 7 patients with 0–5 BCNAs per tumor) and a high-BCNA (HB) group, including all the other MSS tumors (21 tumor samples from 20 patients with 6–21 BCNAs per tumor). The average BCNA score for the LB samples was 1.85 (SD = 2.03, n = 7) whereas for the HB group it was 12.76 (SD = 4.65, n = 21). BCNA scores of our series of MSI tumors were lower than the first quartile of MSS BCNA distribution, and had an average count equal to 2.6 (SD = 1.67, n = 5).

Apart from 16p, whose gains have been found only in the LB group, all chromosomal gains found in LB CRCs were also found in the HB group (Fig. 1). This was also valid for losses: chromosomes affected in the LB group were the same chromosomes commonly affected in the HB group (Fig. 1). Gains of whole chromosome 8 (8w) were more common among MSI samples compared to HB and LB tumors (Fig. 1). In Fig. 1 data obtained in our series of MSI tumors are shown. Indeed, broad gains of chromosomes 7, 8, 9, 13, 20 were also observed by Sveen et al. in a different series of 27 MSI samples [15]. Loss of 18w was far less frequent in LB and MSI tumors of our series, as well as in the MSI CRCs studied by Sveen et al. [15]

In the LB group, apart from 2 tumors of the rectum, 3/5 (60%) tumors were proximal (right colon) and 2/5 (40%) were distal (left colon and sigmoid colon). In the HB group, excluding 2 rectal tumors, 10/19 (53%) CRCs were proximal and 9/19 (47%) were distal. Of the 5 MSI samples of our series, 4/5 (80%) were proximal and only 1/5 (20%) was distal. Several CRC tumors showed mucinous histology or mucinous features. In the LB group 3/7 (43%) samples were mucinous adenocarcinomas (extracellular mucin > 50% of the tumor volume), 2/7 samples (29%) were adenocarcinomas with mucinous features (mucin 10–50%), 1/7 (14%) was a signet ring cell adenocarcinoma. In the MSI group 1/5 samples (20%) was a mucinous adenocarcinoma, 2/5 (40%) were adenocarcinomas with mucinous features. In the HB group 3/21 samples (14%) were mucinous adenocarcinomas, 2/21 (10%) were adenocarcinomas with mucinous features and 1/21 (5%) was a signet ring cell adenocarcinoma.

Differences in gene expression between each CRC group and the normal colonic tissue were assessed. A fold change (FC) > 2 was chosen as a threshold for upregulated genes, whereas a FC < − 2 was chosen for downregulated genes. A false discovery rate (FDR) p-value < 0.05 was used to establish statistical significance [17]. In these expression studies, results regarding MSI tumors have been obtained using raw HTA data from 34 MSI samples provided by Sveen et al. (2017) [15].

Results on the top 30 differentially expressed genes (DEGs) are shown in Additional file 2: Figure S1 (upregulated in a, c and e; downregulated in b, d and f). As expected, markers of normal colonocytes, such as Solute Carrier Family 26 Member 3 (SLC26A3) [18, 19], Chloride Channel Accessory 4 (CLCA4) and Membrane Spanning 4-Domains A12 (MS4A12) [20, 21], as well as other enterocyte markers such as Carcinoembryonic Antigen Related Cell Adhesion Molecule 7 (CEACAM7) [22], Guanylate Cyclase Activator 2A (GUCA2A) [23], Aquaporin 8 (AQP8) [24], isoforms CA1, CA2, CA4 of Carbonic Anhydrase and Keratin 20 (KRT20) [25, 26] were downregulated in the three CRC groups. Solute Carrier Family 12 Member 2 (SLC12A2), encoding the Na-K-Cl cotransporter isoform 1 (NKCC1), was the most highly upregulated transcript in the three CRC groups (first position in HB and third position in LB and MSI).

In order to identify those DEGs that are specifically upregulated in each of the three CRC groups (specifically upregulated DEGs) or that are concurrently upregulated in two CRC groups (shared upregulated DEGs), we used the thresholds for FCs and FDRs reported in Table 1. A similar analysis (Table 2) was also performed for specifically or shared downregulated DEGs. Results are shown in Fig. 2 and Fig. 3.

HB tumors were characterized by upregulation of amphiregulin (AREG) and epiregulin (EREG), which are members of the epidermal growth factor (EGF) family, and have both been reported as upregulated in CRC [27‐30] (Fig. 4). Moreover, a colonic epithelial marker such as Villin 1(VIL1) was specifically upregulated in HB tumors [31].

Both LB and, at a major extent, HB tumors overexpressed AXIN2, a negative feedback inhibitor of Wnt/β-catenin pathway [32] whose upregulation accompanies WNT signaling activation [33] (Fig. 5). The stem-cell marker OLFM4 was significantly upregulated in all CRC subgroups, but FC values were much higher in HB and LB (26- and 74-fold increase respectively) in comparison to MSI tumors (3.5-fold increase).

Analysis of the list of specifically upregulated genes in the MSI group by the “Functional Annotation Tools” of DAVID Bioinformatics Resources 6.8 [34] reveals a significant 3.66-fold enrichment of the term “immunity” (FDR = 6.36E-18), a 9.68-fold enrichment of the term “Immunoglobulin V region” (FDR = 1.57E-9), and a 4.35-fold enrichment of the term “regulation of immune response” (FDR = 6.69E-9), in agreement with the well-known increase in immune cells infiltrate in such CRC group [35]. The list of specifically downregulated genes in MSI tumors shows a significant enrichment in genes related to protein deubiquitation (members of the ubiquitin specific peptidase 17-like family; fold-enrichment: 17.8; FDR = 3.82E-4) and to the family of microbicidal and cytotoxic peptides defensin (DEFA1, DEFA1B, DEFB103A, DEFB103B, DEFB115, DEFB121, DEFB136; fold-enrichment: 25.65; FDR = 3.42E-4).

Interestingly, LB and MSI tumors shared upregulation of genes typical for secretory cells, such as Mucin 2 (MUC2), Mucin 1 (MUC1), Mucin 5B (MUC5B), Trefoil Factor 1 (TFF1), Deleted In Malignant Brain Tumors 1 (DMBT1), Regenerating Family Member 4 (REG4), although upregulation is generally higher for MSI tumors (Fig. 6). Using the “Functional Annotation Tools” of DAVID Bioinformatics Resources 6.8 [34], the list of shared upregulated genes in LB and MSI groups shows a fold-enrichment of 4.24 (FDR = 0.00286) of the category term “secreted” in the UP_KEYWORDS (genes: MUC1, MUC2, REG4, MMP9, SPINK4, POSTN, SERPING1, IGLV3–1, CD55, PI3, IL1B, TFF1, MUC5B, DPP4, DMBT1). LB and MSI tumors also showed upregulation of POSTN (Periostin), a gene involved in various signaling pathways and extracellular matrix (ECM) remodeling [36, 37].

As shown in Fig. 7, MSI tumors showed a specific upregulation of bone morphogenetic protein antagonist Gremlin 1 (GREM1). GREM1 encodes a protein whose expression is weak or absent in normal colorectal epithelium but increases in CRCs, especially those with serrated histology and low tumor stage [38]. A 40-kb duplication of a region upstream of GREM1 has been associated with hereditary mixed polyposis syndrome (HMPS) [39], and a 16-kb duplication in the regulatory region of GREM1 has been found as the disease-causing genetic alteration in a family with attenuated/atypical polyposis syndrome [40]. Gain-of function mutations in GREM1 impair BMP signaling, which normally exerts a negative control on intestinal epithelial cells [41]. Moreover, GREM1 epithelial expression has been shown to contribute to colonic carcinogenesis by reconferring stem-cell features to progenitor cells located at a distance from the stem-cell niche [42]. Finally, it has been demonstrated in glioblastomas that cancer stem cells express GREM1 to counteract BMP-driven differentiation, and promote their self-renewal [43].

We used the Single Sample Predictor (SSP) for CMS classification of HB, LB, and MSI samples (Table 3). About 80% (27/34) of the MSI samples were attributed the CMS1 subtype, as expected, since the CMS1 subtype has been defined as “MSI-Immune”, and most of its samples are MSI [12]. 76% (16/21) of the HB tumors were classified as CMS2, which is indeed the subtype with the highest number of somatic CNAs [12]. The LB group was instead heterogeneous with respect to the proportion of CMS subtypes: 43% of the samples (3/7) could not be attributed to any of the 4 CMS classes, whereas 29% (2/7) were classified as CMS3, 14% (1/7) as CMS1 and 14% (1/7) as CMS2.