hGSuite HyperBrowser description

The hGSuite HyperBrowser web platform is a set of independent tools that can be used together to cover a full analysis scenario, from data preparation to exploration of results. The pipeline is divided into four major steps that can be used to analyse a collection of genomic tracks based on metadata of interest.

Data collection

Data can be uploaded or imported from external repositories using the tools provided by GSuite HyperBrowser [9]. If needed, additional metadata can be added to these GSuite tracks. The GSuite files can then be converted to hGSuites using the available metadata. One of many alternatives is to organize data on a local computer into a zip/tar file and directly create a hGSuite from this file.

Data preparation

Within the web platform, collections of datasets are represented in the GSuite format [9]. The platform includes tools for preprocessing the GSuite tracks into a binary, indexed format in preparation for statistical analyses.

Statistical analysis and visualization

Analyses are provided for a collection of genomic tracks, as well as for pairwise combinations. Analyses are restricted by the defined hierarchical structure associated with the metadata of the genomic tracks. (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Tools for descriptive statistical and visual analyses of hierarchical genomic tracks.

https://doi.org/10.1371/journal.pone.0286330.t001

Exploration of results

Data and analysis results are primarily represented as a data cube Fig 1. This format is suitable for analysing the data from different perspectives, based on the metadata, through operations like slicing, dicing, pivoting, and aggregation [10]. The final results can be viewed as a two-dimensional array, which would enable the user to draw overall conclusions in a time-efficient manner.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Data cube.

Schematic overview of the hGSuite tool that uses the concept of a data cube and its different dimensions. The figure indicates the initial collection and preparation of data, through customization of data using the metadata, to the statistical evaluation for aiding a biological hypothesis. The input is a set of genomic tracks that are converted to GSuite with a hierarchy defined based on the metadata. The data along with metadata can be fitted in a cube that has dimensions based on the hierarchy. The second part of the figure illustrates the different ways in which the data can be viewed in a cube. The cube can then be sliced from different dimensions based on the user query to get the desired output.

https://doi.org/10.1371/journal.pone.0286330.g001

Case studies

To demonstrate the possible applications for the hGSuite HyperBrowser web platform, two case studies are presented. Data, tools, and a detailed history are provided at https://hyperbrowser.uio.no/hgsuite/, under the Examples section. A visual summary of both studies can also be found in Figs 2 and 3 and S5–S18 Figs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Mouse data.

Use case 1 (A-C) exemplify the hGSuite usage. Use case 1 shows the analysis of mutation profiles of different mouse genotypes. A: BED files are converted to GSuite track and metadata information is added to define the hierarchy as a hGSuite. This is then given as input to the tool “Compute data cube for hGSuite” which has the hierarchy of the query track as each dimension. B: Mutation profile of each mutation type in each genotype of the mouse data. C: Rainfall plot of mutational frequency along the genome and mutational frequency plot for one single chromosome.

https://doi.org/10.1371/journal.pone.0286330.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Human data.

Use cases 2 (A-G) exemplify the hGSuite usage. Use case 2 shows an analysis of the mutational profile for each patient and the average mutational profile across 3 cancer types—colorectal, breast, and skin cancer. It also shows how the data cube can be filtered to show the frequency of mutation for one patient. A: BED files are converted to GSuite tracks and converted to hGSuite using metadata. The analysis can be done either on just the query track or it can be analysed for overlap with other tracks containing other regions of the genome such as DNA repair genes. B: Average number of mutations across all cancer type. C: Average number of mutations across patients in each cancer type and the average number of mutations of each mutation type across each cancer type. D: Number of mutations for each mutation type across different cancer types in coding regions of reference genome. E: Average number of mutations in DNA repair pathways for every patient with colorectal cancer. F: Average number of mutations of each mutation type for a single patient’s DNA (patient006). G: Frequency of mutations in chromosome 6 for a single patient.

https://doi.org/10.1371/journal.pone.0286330.g003

Firstly, we analysed the contribution of DNA repair to antimutagenesis. We have generated three types of murine knockouts (KOs) [12, 14] in the two major U repair enzymes; UNG and SMUG1 [12, 13]. SV40-transformed mouse embryonic fibroblasts (MEF) were serially cultured over 19 passages. MEFs from passage 1 and 15 were subjected to whole genome sequencing. The types of mutations accumulating from passage 1 to 15 were extracted as previously described [14]. BED files from wild type (WT), UNG KO, SMUG1 KO, and UNG-SMUG1 double KO mice were used to look at the mutation profiles with the data cube (Fig 2A–2C). The genotypes of the mice were used as one dimension of the cube. We also split the mouse tracks based on the mutational types, which were used as another dimension of the cube. From S1 Fig, the first observation that can be made is that there was little to no difference between the number and type of mutations accumulating in WT and SMUG1 KO MEFs. We also observed that more mutations accumulated in UNG KO MEFs. As expected, C>T and G>A transitions, most likely originating from spontaneous cytosine deamination, were by far the dominating class of mutation. However, mutations at AT base pairs were also higher in the UNG KO. Uracil misincorporated opposite adenine are read as thymine and are therefore not directly mutagenic. Mutations at AT pairs might therefore suggest that these lesions might be misprocessed. Interestingly, all mutation classes except A>C-T>G were reduced in UNG-SMUG1 double KO. UNG and SMUG1 both contribute to uracil repair [13], but the reduced mutation accumulation in double KO cells challenges the idea that these enzymes are fully compensating for each other. Instead, our data suggest that SMUG1 initiates error-prone processing of uracil residues that are preferentially repaired by UNG. Figure S3 Fig shows the rainfall plot of the genome from which we see a dominance of C>T and G>A mutations (green), which are found in high density within chromosome 13, as well as two different locations within chromosome 5 for the UNG KO. Premutagenic uracil residues are not expected to be introduced homogeneously along chromosomes as the cytosine deamination rates are higher in single-stranded DNA [15]. Similarly, the accessibility of DNA repair enzymes are likely to depend on chromatin compaction. To visualize mutation distribution along chromosomes we developed a tool where mutation frequency (number of mutations per bin) were mapped together with other genomic features, such as exons or CpGs. From figure S4 Fig, we observe that our peaks of interest in chromosome 5 primarily coincide with exons (blue line) and not CpG islands (orange).

For the second biological example, we obtained BED files of somatic nucleotide variants of 30 samples from the Hartwig cancer database [1]. The dataset consisted of 10 patients each for colorectal, breast, and skin cancer. The mutation profile of each patient was studied for preferential occurrence in specific regions of chromosomes, such as coding, non-coding, and regulatory regions, both individually and also based on cancer type. The results are shown in S2 Fig. To compare general mutational profiles, such as the average number of mutations based on cancer type, we aggregate the results across the “patientID” dimension and display the value for each based on the primary tumor type. Fig 3(A)–(3E), S5–S7 Figs show that consistent with large datasets [15], the mutational burden of skin cancer patients is relatively higher compared to breast and colorectal cancer patients. Similarly, when we aggregate the results based on mutation type for each cancer type dimension and the transpose of it (S6 and S7 Figs, respectively), we observe the relative occurrence of different mutation types across cancers. As expected, the C:G to T:A mutation is the most common across all cancer types due to the deamination of cytosine or 5-methylcytosine. For the next analysis, we observed how the frequency of mutations varies along chromosomes, and correlates with functional features such as coding and non-coding regions. To do this, we compare the GSuite of the patients against a GSuite containing information about coding and non-coding regions of the reference (hg19) genome. The results for the coding region mutational profile for different mutation types are shown in S8–S10 Figs. S8 Fig shows the result of the aggregated average overlap in the coding region for each mutation type across cancers. From S10 Fig, we observe that skin cancer patients have the highest average of C>T-G>A mutations in the genome and in the coding regions. Ultraviolet light is the major driver of mutagenesis in cutaneous melanomas, which produces a distinct mutational signature with C>T and CC>TT dinucleotide transitions [15] which explains the high occurrence of the C>T-G>A mutation type. Similarly, in colon cancer patients, after the C>T-G>A mutation, it is C>A-G>T that appears most frequently, a common cause of driver mutations in e.g. KRAS, in colorectal cancer patients [16]. As a next step in understanding our dataset, we wanted to study the co-occurrence of mutations in the chromatin state of a given cell type, since mutations in open chromatin regions are a prognostic factor in various cancers [17]. Hence, understanding the effects of individual mutations in the chromatin state of a given cell type may allow the identification of novel cancer-associated regulatory genes or sequences [17]. GSuites of DNAseq and ATACseq for cell lines associated with each cancer type were obtained from ENCODE and used in the hGSuite. The compute relations between two hGSuite tools were used to obtain an average occurrence of mutations in the open chromatin regions (S11 and S12 Figs). This provides a broad view of co-occurrence in each cell line for mutations of a patient with a particular type of cancer. A failure in DNA repair pathways and cell cycle checkpoint networks that cooperate to ensure the maintenance of genetic stability can lead to tumor growth [18]. Hence, the next set of questions we want to ask about our data is based on DNA repair genes and their associated pathways. To do so, we downloaded the location of DNA repair genes. The genes are classified based on the DNA repair pathways such as mismatch repair (MMR), base-excision repair (BER), non-homologous end-joining repair (NHEJ), nucleotide excision repair (NER), homologous recombination (HR), and the set of DNA polymerase genes. The GSuite of patients was compared with the GSuite for DNA repair pathway genes consisting of all 6 different pathways. The result for colorectal cancer is shown in S14 Fig, where we get the average normalized count for each mutation type across patients with colorectal cancer. When we aggregate along the first GSuite of patients and select colorectal cancer as the primary tumor location, we notice that MMR and NHEJ are the two repair pathways that show a higher number of mutations. It has been shown that approximately 15–18% of colorectal carcinoma show mismatch repair deficiency [19–22] and up to 3% of tumors develop in association with Lynch syndrome that also results from mutations in MMR genes, most commonly MLH1, MSH2, MSH6 or PMS2 [23]. DNA double-stranded breaks (DSBs) are genotoxic carcinogens, and the NHEJ allows DNA recovery by direct joining [24]. It has been shown that increased mutations within NHEJ genes increase the risk of colorectal cancer [23]. Among breast cancer patients (S15 Fig), we observed a high frequency of mutations in the base-excision repair pathway (BER). While observing skin cancer (S16 Fig), it clearly shows an increase in mutations along all the pathways compared to the other two types of cancer which fits well with the previous observations that show the highest mutation burden.

The above case study can also be done for each patient (illustrated in Fig 3(D), 3(F) and 3(G)) and can be useful in developing treatment strategies. We have done this with a breast cancer patient having an odd mutation profile and found that a highly mutated region within this patient mapped to the DNA polymerase genes. S17 and S18 Figs shows the rainfall plot and frequency plot profile for patient 6 for each chromosome. Overall, these results show the hGSuite’s ability to aid in both basic and personalized research.

Comparing hGSuite to other programming tools

The main approach in the field of bioinformatics has been to download data from multiple databases and restructure the data to define the hierarchy. Downloading and re-integrating the tracks or collection of tracks as needed quickly becomes time-consuming. To illustrate the benefits of using the hGSuite, we provide an example of an analysis to get an overall average mutation profile of the tracks based on any dimension such as cancer type, mutation type, or per patient. The results of the analysis of overlap between certain genomic regions of the chromosome and the tracks from any of the above-mentioned dimensions are shown and explained. The screencast (https://bit.ly/hGSuite) illustrates the analysis using the hGSuite while the R code shows the equivalent analysis. We compare it to the equivalent R analysis (S1 File).

The different statistical and visualization tools in the hGSuite enable performing analysis on multiple tracks by simple repetition of computations on each track. The results of the analysis are saved in the data cube making it flexible for the user to analyse and visualize the results in any dimension defined using the metadata. In the hGSuite this takes 5 clicks as shown in the screencast. In R, the data require a manual download, and the code is written for every dimension separately. The code presented does not necessarily represent the optimal way of solving it with R, but we do believe it to be representative of how an analyst might typically approach such an analysis. Notably, it would take more than 100 lines of code to achieve the same result we would get in 4 clicks with hGSuite.