Background & Summary

Colon cancer is a complex disease characterized by an accumulation of alterations at different molecular levels. Many studies focus on one molecular level, leading to a partial view of the carcinogenic process. Here we present Colonomics (https://www.colonomics.org/), a multi-omics dataset that comprises 100 stage II colon cancer patients from which samples of tumoral tissue (T) and paired normal adjacent tissue (N) were obtained from the surgical specimen. Also, 50 samples from healthy colon mucosa obtained at colonoscopy from donors (M) without cancer were included and used as reference (Fig. 1).

Fig. 1
figure 1

Scheme of Colonomics data. Number of samples that remained after quality control for each type of data. M are healthy colon mucosae, T is tumoral tissue and N is normal tissue adjacent to tumor.

Full size image

Multiomics data analysis allows to integrate different levels of biological and molecular information to understand complex diseases as cancer and to move forward towards precision medicine1,2. The Colonomics project aims to study the molecular basis of colon cancer through the integrative and large-scale analysis of different molecular data and to identify useful biomarkers of diagnosis and prognosis in colon cancer.

From each sample, DNA and RNA were extracted to obtain molecular data. Thus, extracted DNA was used to acquire methylation data (using Illumina microarrays), single nucleotide polymorphism (SNP) genotypes and CNV (using Affymetrix microarrays) and somatic mutations (through whole exome sequencing) in a subset. RNA was used to obtain gene and miRNAs expression, using Affymetrix microarrays and SOLID sequencing, respectively. Finally, patient clinical data from all samples were obtained, including follow up.

The analysis of these datasets allowed us to make a descriptive and a comparative snapshot of the molecular state of cancer cells (T) and, also, to find differences between healthy normal tissue (M), normal tissue adjacent to tumor (N), and tumoral tissue (T).

The Colonomics project web page (https://www.colonomics.org/) links to the digital data repository of the University of Barcelona (UB-DD)3 and provides four user friendly web tools that allow an agile and direct visualization of data.

Methods

Patients and samples

Colon cancer patients were retrospectively selected during 2012 from those that had received radical surgery in the Bellvitge University Hospital (Barcelona, Spain) between January 1996 and December 2000, and had fresh frozen tumor samples in the hospital biobank. One hundred patients were included, with the criteria of a confirmed adenocarcinoma of the colon, diagnosed in stage II and the tumor was microsatellite stable (tested with five markers). Patients should have not received adjuvant therapy and a minimum follow up of three years was requested at the time of selection. A pathologist reviewed the tumor blocks, confirming diagnosis and that there were no signs of tumor cells when margins were examined. The paired adjacent normal mucosa (N) was obtained from the proximal margin, at least 10 cm distant from the tumor (T). As a control group, 50 healthy mucosa donors (M) were prospectively invited to participate in this study when they underwent a colonoscopy indicated for screening or symptoms with no evidence of lesions in the colon or rectum. Table 1 shows the basic clinical characteristics of the samples. The study protocol was approved by The Clinical Research Ethics Committee (CEIC) of the Bellvitge Hospital. All the recruited individuals provided a written informed consent to participate in the genetic study. The approval number is PR178/11.

Table 1 Clinical characteristics of the 250 samples.
Full size table

From these 250 samples (100 T, 100 N and 50 M), data on gene expression, DNA methylation, SNPs, CNV and miRNAs were obtained. Whole exome sequencing was analyzed in a subset of 42 tumors and their paired adjacent normal mucosa. Figure 1 shows the number of samples for each type of data.

DNA extraction

DNA was extracted from colon tissue by the phenol-chloroform procedure, quantified using a NanoDrop ND 2000c spectrophotometer (NanoDrop Thermo Scientific, DE, USA) and stored at 4 °C. Bisulphite conversion of DNA (200–500 ng) was performed according to the manufacturer’s recommendations for the Illumina Infinium Assay (EZ DNA methylation kit. Zymo Research. Cat. No. D5004). The incubation profile was 16 cycles at 95 °C for 30 seconds, 50 °C for 60 minutes and a final holding step at 4 °C.

RNA isolation

Total RNA was isolated from tissue samples using Exiqon’s miRCURY™ RNA Isolation Kit (Exiqon, Denmark). For quantification, NanoDrop® ND-1000 Spectrophotometer (Nanodrop technologies, Wilmington, DE) was used. RNA quality was assessed by gel electrophoresis and RNA 6000 Nano Assay (Agilent Technologies, Santa Clara, CA). RNA purity was measured with the ratio of absorbance at 269/280 nm (mean = 1.96, sd = 0.04), no differences among tissue types were found. RNA integrity numbers showed good quality (mean = 8.1 for T, and 7.5 for N).

Gene expression data

Affymetrix Human Genome U219 Array Plate platform (Affymetrix, Santa Clara, CA, USA) was used to obtain gene expression data from isolated RNA. A block experimental design was performed to the 96-array plates to avoid batch effects and sample ids were blinded to laboratory technicians. Robust Multiarray Average (RMA) algorithm in affy package (version 1.28) from R/Bioconductor4 was used to normalize data. After quality control and normalization 49,386 probesets in 20,070 genes and 246 samples remain for the analysis: 50 M, 98 N and 98 T (98 paired N/T). Since the array provides multiple probes per gene, a summary gene expression value for each gene was obtained through a principal components analysis (PCA). We extracted the first principal component, that captures the highest expression variability, and was rescaled to have the average expression of the different proves and maximum observed standard deviation. This rescaling was required to avoid negative expression values of the first PC. Therefore, in addition of the dataset with the measured probes, a final dataset with a single expression value for each one of the 20,070 genes was obtained.

Methylation data

The Illumina Infinium HumanMethylation 450k BeadChip assay was used to obtain the DNA methylation profile of the samples. The array interrogates the methylation levels of 485,512 CpG sites covering 99% of RefSeq genes and 96% of CpG islands5,6. To minimize batch effects, samples were randomly distributed but matching paired samples (N, T) in the same array row (M samples were randomly paired in array rows).

Sample quality and sex concordance was checked using the SNPs of the 450 K array with those of the Affymetrix Genome-Wide Human SNP 6.0 array (Affymetrix, Santa Clara, USA). A total of 10 samples had to be excluded due to quality problems and the final dataset contained 240 samples: 96 N, 96 T (92 complete pairs of N/T) and 48 M.

Background correction was performed filtering out probes with a detection p-value greater than 0.01 in more than 5% of the samples. We discarded probes that ambiguously mapped to multiple locations in the human genome with up to two mismatches7. We also excluded probes that contained SNPs within 10 bp of distance. Finally, a set of 430,086 probes was obtained.

We used β-values, the Illumina’s standard, to measure methylation level at each locus. β-values are calculated using the average probe intensity and they range from 0 to 1, being β = 0 the absence of methylation and β = 1 a complete methylation. Subset-quantile within array normalization (SWAN) was used to reduce technical variation within and between arrays8 using the Bioconductor package minfi (version 3.0.1)9. A random forest approach was used for the imputation of missing values.

Genotyping data

Extracted DNA was hybridized to obtain the genotypes of the samples using Affymetrix Genome-Wide Human SNP 6.0 array (Affymetrix, Santa Clara, USA), which includes near 1 million of SNP markers. Genotype calling was performed with Corrected Robust Linear Model with Maximum Likelihood Classification (CRLMM) algorithm as implemented in R/Bioconductor package crlmm (version 1.8)10. A total of 4 samples were excluded due to quality or sex concordance problems (3 M and 1 N) so 146 samples (47 M and 99 N) remained for the analysis.

Copy number variation data

The Affymetrix Genome-Wide Human SNP 6.0 genotyping array was used to infer the copy number aberrations in the tumors. The average inter-marker distance of this array was 700 bp. Copy number estimate (CNE) was calculated using Affymetrix Power Tools (Version 1.16.1) software11 with default parameters.

Adjacent locus with a similar CNE were grouped in regions using a segmentation pipeline that had three steps. First, a smoothing spline was applied to normalize data. Second, change points in CNE pattern were located using the Vega R package (version 1.35)12 and split into discrete segments. To avoid masking effects, the stromal content of the tumor samples was taken into account to calculate the thresholds to split the segments. Stromal content was estimated using the ESTIMATE R package (version 1.0.13)13 using gene expression data14. A hierarchical cluster analysis was performed to divide the samples in 4 clusters reflecting their different levels of stromal content. A different threshold to split the segments was used for each cluster: ±0.5 for low stroma, ±0.4 for medium-low stroma, ±0.3 for medium-high stroma and ±0.2 for high stroma15. The third and final step of the segmentation process consisted in a t-test to compare the mean of consecutive segments and to merge them in case there are no statistical differences (p-value > 1e-04). For each segment, the mean of CNE was assigned as the representative value of the region. 26,423 segments of CNV were found in the 99 T samples (15,646 losses and 10,777 gains).

miRNA expression data

Isolated RNA was used to obtain the expression of miRNAs. Small RNA Assay of the Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA) was used to assess the quality control of the miRNA fraction following manufacturer’s recommendation. Small RNA-seq was performed through the SOLiD platform. The PureLink miRNA isolation kit was used to construct the libraries of compatible fragments with SOLiD from an enriched fraction of small RNA. Sequencing microspheres were obtained by applying an emulsion PCR into an equimolar mixture of 48 libraries followed by an enrichment process before charging in the reaction chamber. Finally, the reaction to obtain the sequences (of 35nt + 10nt barcode) from miRNA fraction was performed by blinded laboratory technicians through the Applied Biosystems SOLiD 4 System. The samples were randomly distributed among the different sequencing slides to minimize batch effects. The data quality was estimated using the SOLiD Experimental Tracking System (SETS) software. Parameters such as total number of reads, proportion of reads with miscalls or proportion of reads with low average quality value (QV) were evaluated for quality control with HTS SOLiD Preprocessing software16, removing reads containing a miscall as well as reads containing negative quality scores. All 250 samples showed adequate quality. Quantification of specific miRNAs was performed mapping the reads to the reference of mature miRNA sequences annotated in miRBase release 22.1 (October 2018) that contained 2,654 human miRNA sequences17. The FASTX-toolkit18 was used to preprocess the miRNA data and to provide compatible sequences for mapping with the Bowtie 1.2.2 aligner19. Cutadapt20 software was used to trim read adapter and the final table of counts was generated with SAMtools mpileup v1.821. Finally, 2,641 miRNAs remained for the analysis.

Whole exome sequencing

Genomic DNA from a subset of 42 N and T paired samples was sequenced in the National Center of Genomic Analysis (Barcelona, Spain; CNAG) using the Illumina HiSeq- 2000 platform. Samples selection was required due to budgetary constraints. Samples included were those from all patients that had experienced tumor progression (n = 21), and a matched sample from a patient with more than 5 years of follow-up without progression.

Exome capture was performed with the commercial Agilent kit Sure Select XT Human All Exon 50MB. Exomes from T were sequenced at 60x coverage (2 × 75 bp reads), and exomes from N were sequenced at 40x (2 × 75 bp reads). The sequences were quality controlled using FastQC software22 and aligned over the human genome hg19 version, with Bowtie 2.023. Unmapped reads, reads with unmapped mate, nonprimary alignments, and reads that were PCR or optical duplicates were discarded with Picard24. A local realignment around indels found in our samples and defined in dbSNP 135.b37 version25 and 1000 genomes project26 was applied. Variant calling was performed with GATK software27, and low-quality variants (mapping quality below 30, read depth below 10 or frequency <10%) were filtered out. Germline variants from our study that are the variants found in normal adjacent paired tissue but not in tumor were removed. Other germline variants found in normal tissue from 1000 genomes project were also filtered out. Only single-nucleotide variants (SNV) were included in this study. Finally, variants were annotated using the SeattleSeq Variant Annotation web tool28. No correlation was observed between the number of mutations per sample and the control quality parameters (number of reads, number of no matched reads, percentage of unique aligned reads, number of duplicate reads and coverage). Finally, 13,015 somatic variants (11,126 SNVs) were found.

Data Records

All the Colonomics data is publicly available, though it has been deposited in diverse repositories. Table 2 summarizes the accession for the different types of data. To facilitate performing multiomics analysis, downloading and merging the different sets, all samples have the same identifiers in all the upload datasets. We provide the processed data in the Colonomics website (https://www.colonomics.org/), and in the digital data repository of the University of Barcelona (UB-DD)3, to ensure permanent availability.

Table 2 Data accession summary table.
Full size table

Gene expression includes 246 samples: 50 M, 98 N and 98 T. Gene expression raw and normalized data are available in the Gene Expression Omnibus database (GEO) through BioProject PRJNA188510 and accession number GSE4407629. Raw data was upload as 246 CEL files and the normalized data is included within sample table as text file. The normalized data as log2 RMA signals include 49,386 probesets and can also be downloaded from UB-DD3. Finally, the gene expression summarized at gene level is available at the UB-DD3 and includes 20,070 genes.

DNA methylation data, both, raw and normalized are also available in GEO through BioProject PRJNA542323 and accession number GSE13101330. Methylation data includes 240 samples: 48 M, 96 N and 96 T. We provide 240 IDAT files that correspond to raw data and a text file with the normalized beta values of the 240 samples and 430,086 CpG probes. The normalized methylation data is also available at the UB-DD3 together with the annotation of the 430,086 CpG methylation probes. Both gene expression and DNA methylation data are part of the GEO super series GSE16642731.

Raw genotyped SNP data can be obtained, under controlled access, at the European Genome-phenome Archive (EGA) through accession number EGAD0001000125332. It includes 146 CEL files that corresponds to 47 M and 99 patients of colon cancer.

The processed CNV data of 99 T samples is accessible in the UB-DD3. It includes 54,002 segments: 15,646 losses, 10,777 gains and 27,579 segments that are not gains nor losses.

Raw small RNA sequencing data is also available under request in EGA through accession number EGAD0000100482733. We provide 250 fastq files, one for each sample (50 M, 100 N and 100 T). Count data of the 2,641 miRNAs and the 250 samples are accessible at UB-DD3.

Raw whole exome sequencing data of 84 samples (42 N and 42 T) with 2 fastq files per sample are accessible in EGA under request through accession number EGAD0000100482634. The tumors mutations identified are available at UB-DD3.

SNPs, miRNAs and whole exome data are in the EGA repository under the Data Access Committee through accession number GAC0000100066235.

Finally, somatic mutations of the 42 samples can be downloaded from UB-DD3. 13,015 somatic variants were included.

In addition to raw and preprocess multiomics data, clinical information of the samples is also upload at UB-DD3. This file is txt formatted and contains a row for each sample and a column for each variable. The variables included are: id_clx: the identifier of the Colonomics sample; type: tissue type (Mucosa, Normal or Tumor); id_clx_individual: the individual identifier; stage: the stage of the tumor (IIA or IIB); sex: (Female or Male); age; site: the site of the colon (Left/Right); event_free: a binary variable that indicates if the patient has a recurrence of colon cancer (1) or not (0); time_free: a continuous variable that indicates the time from surgery to recurrence of colon cancer; event_global: a binary variable that indicates if the patient is still alive (1) or not (0); time_global: a continuous variable that indicates the time from surgery to death; metastasis_site: the organ where a metastasis has appear; BRAF_mutated: a binary variable that indicates if the tumor is BRAF mutated (Yes) or not (No); KRAS_mutated: a binary variable that indicates if the tumor is KRAS mutated (Yes) or not (No); stromal_score: stromal score to predict the level of infiltrating stromal cells in tumor samples obtained using ESTIMATE package (version 1.0.13)13 from R and CMS: consensus molecular subtype classification of the tumoral samples in 4 groups (CMS1, CMS2, CMS3 or CMS4) obtained using CMScaller package (version 0.1)36 from R and based on the paper of Guinney et al.37.

To facilitate working with all these data and to avoid downloading each dataset individually, a MultiAssayExperiment object from R (version 1.22)38 has also been created. This object includes clinical data, data from gene expression, methylation, miRNAs and copy number variation and it is accessible in UB-DD3.

In addition to clinical and multiomics data, gene expression prediction models from SNP data, useful for Transcriptome-Wide Association Studies (TWAS), can be download from Zenodo through accession number 633476839. These models were built for normal colon tissue (samples N + M) using elastic net regression, following the PredictDB pipeline40,41. They include significant prediction for 1,758 genes, 17,281 CpG probes and 39 miRNAs. See the usage notes section for more details.

Technical Validation

We performed technical validations and results replication in other datasets for some of the Colonomics assays. For gene expression, a selected group of genes were assessed with multiplexed RT-qPCR using BioMark Dynamic Array 96 × 96 Plates (Fluidigm Corporation, San Francisco, CA). ACTB, TPT1, and UBC were used as control genes in the assay, see Solé et al.42 for details. Also, differentially expressed genes were validated using 45 N/T paired samples from The Cancer Genome Atlas (TCGA) data43, obtaining that the 97.86% of our differentially expressed genes were also differentially expressed in TCGA data44. Differentially expressed genes were found using limma R package (version 3.42)45. Genes with an absolute log fold change >1 and a Bonferroni adjusted P-value < 0.05 were defined as differentially expressed.

Regarding DNA methylation data, the list of the differentially methylated CpGs obtained using Colonomics data was compared with the one obtained using the 45 N/T paired samples from TCGA data obtaining that the 99.96% of the common CpGs were also differentially methylated in TCGA data44. Also, CNV was validated using TCGA data. In this case, 222 colon cancer samples were used to compare the obtained 13,279 minimal recurrent regions and 66% of these regions were also found in TCGA data15.

In the analysis of whole exome sequencing data, 13 SNPs in the 84 samples were genotyped using KASPar genotyping assays (KASP-By-Design; LGC Group) on the Fluidigm genotyping platform (48.48 Dynamic Array IFG, Fluidigm). These data were used to assess concordance between samples. All 42 N samples correctly matched with their corresponding T sample46. Also, to validate recurrent mutations found in AMER1 gene, sanger sequencing was used. All mutations were validated.

As a validation of the discovery pipeline of SNVs, 6 mutations in KRAS (Q61H, A146T, G12V, G12D, G12S, G13D) and 7 in TP53 (G245D, R248Q, R237H, R273C, R175H, R282W, R213_, G245S) were tested using KASPar genotyping assays in the Fluidigm Biomark platform (dynamic arrays). A 65% of concordance were achieved. It is noteworthy that 10 out of 11 no concordant mutations were only found by exome sequencing thus confirming the higher sensitivity of this technique.

Usage Notes

Re-use of the data and study limitations

All the data in the Colonomics project has been analyzed and used in published papers except for miRNAs data. The first paper involving Colonomics data was published eight years ago in 2014. The data were analyzed with the best algorithms and programs at that time, but nowadays there may exist improved analysis pipelines. That’s why we have shared raw data for each omic layer in addition to preprocessed data. Specifically, for the case of whole exome sequencing data that was processed 8 years ago, we are aware that the calling algorithms may have produced some false positive mutations.

When re-using Colonomics data we also need to consider that all samples were confirmed adenocarcinoma of the colon, diagnosed in stage II and the tumor was microsatellite stable. Colonomics data is a very homogeneous cohort, but the absence of rectal tumors, colon tumors in other stages rather than stage II and microsatellite instable samples may limit building generalizable hypothesis of colorectal cancer. Patients were selected to have received only radical surgery, without adjuvant chemotherapy, which also reduces potential analysis to prognosis but not prediction.

Web browsers

In order to visualize the different types of data, we developed different intuitive and user friendly shiny47 web applications that are freely accessible at https://www.colonomics.org/data-browser. There are four different web applications: the expression browser, the methylation browser, the expression quantitative trait loci (eQTL) browser and the regulatory networks browser. In these browsers, expression, methylation, and genotyping data have been exploited developing different types of analysis. A web browser integrating all the types of omics datasets available would be an interesting future application.

Genetic prediction models for normal mucosal biopsies

The genetic prediction models represent reference imputation panels for normal colon tissue methylation and gene and miRNA expression, which are of high interest for performing TWAS studies for colon-related diseases.

We provided significant prediction models for 1,758 genes, 17,281 CpG probes and 39 miRNAs obtained from normal biopsy samples. These features can be predicted from SNPs located within ± 1 Mb, which we assumed they act through cis mechanisms. We included the model’s summary statistics and corresponding SNP weights in SQLite objects39. Models were trained using the elastic net procedure employed in the PredictDB pipeline40,41, according to which only models with a predictive performance p-value < 0.05 and R2 >0.1 are considered significant. We adjusted the models by basic covariates, i.e., sex, age, tissue type and colon anatomic location where biopsies were collected (left and right colon). Genome coordinates refer to GRCh37/hg19.

Prediction models for gene expression were trained using 144 normal mucosal biopsy samples (97 N and 47 M), the common samples between gene expression and SNPs. We trained models for 13,939 genes (including protein coding, long non-coding and pseudogenes), which had more than 3 sequencing reads in more than 10% of the samples. In cases where multiple probes mapped to a single gene, the first principal component from PCA was used to capture the largest common variability. Trimmed Mean of M-values (TMM) normalization was applied. TMMs were transformed with inverse normal transformation.

Prediction models for CpG probes were trained using 132 normal mucosal biopsies (95 N and 37 M). Only CpG annotated in islands, shores and shelves were taken into account as CpGs in open-sea are not functionally interesting. We also filter in CpGs which had more than 3 sequencing reads in more than 5% of the samples. After these filters, 257,809 CpG remained for the analysis. Inverse normal transformation on Beta MIxture Quantile (BMIQ)-normalized values was applied. In addition to the basic covariates for adjustment, we adjusted the models by 10 probabilistic estimation of expression residuals (PEER) factors48 to capture additional technical variability.

Prediction models for miRNAs were trained using 146 normal mucosal biopsies (99 N and 47 M), the common samples between miRNAs and SNPs. From the 2,641 miRNAs, we included 739 miRNAs, which had more than 3 sequencing counts in more than 10% of the samples. TMM normalization was applied and transformed with inverse normal transformation. As SNP data is in hg19 genome built, annotations of miRNAs refer to the miRBase release 2017,49.