CRI iAtlas: an interactive portal for immuno-oncology research

Implementation

iAtlas is a web-based application to enable data exploration for clinicians, biologists, and informaticists. The inputs and architecture of the application are described below.

Data

The iAtlas app uses structured data and outputs from the Immune Landscape⁶ study and the TCGA PanCancer Atlas initiative⁸, which harmonized TCGA data, ensuring uniform quality control and sample inclusion, batch effect detection, normalization across platforms, combination mutation calling from multiple centers, and robustly compiled clinical and outcome data. A key source of iAtlas data is the table summarizing tumor-sample and immune characterizations for 11,080 TCGA patient participants of the TCGA, Table S1 of the Immune Landscape⁶ manuscript, here termed the “PanImmune Feature Matrix”. Auxiliary data were sourced from files available on this manuscript’s data page at the NCI Genomic Data Commons, from the TCGA PanCancer Atlas Data Mirror, and from the TCGA PanCancer Atlas working space in Synapse (see Data availability). iAtlas data were formatted as data frames (tables) and stored as “feather” files (https://github.com/wesm/feather) on the application server for fast loading (Table 1).

Annotation and browsing of the PanImmune Feature Matrix: iAtlas includes a Data Description page with details on all variables presented in individual modules, with the ability for users to “drill down” on related groups of variables to understand how values were derived. Variables are listed in a text-searchable table containing the name of the variable, the ‘Variable Class’, the unit (if applicable), and whether the variable is numeric or categorical. A ‘Variable Class’ is the name of a group of variables that are of similar type and are often the result of one particular analysis. Clicking on a row exposes a list of all variables in the ‘Variable Class’ and provides links to text descriptions of the analysis methods used to generate the variables.

R/Shiny architecture

iAtlas is powered by Shiny⁹ and makes extensive use of Shiny Modules¹⁰ to organize code into composable units (Figure 1). Each iAtlas Analysis module is designed as a Shiny module, allowing simple integration of new analytical functionality. iAtlas uses the tidyverse¹¹ family of R packages (e.g., dplyr¹², tidyr¹³, purrr¹⁴, stringr¹⁵, tibble¹⁶) as well as the wrapr¹⁷ package to assist with tidy evaluation. These functions power the data transformations of internal tabular data that are then used to create the interactive plots (i.e., with the plotly¹⁸ graphing library) and data tables (via the DT¹⁹ wrapper to the DataTables library) seen through the iAtlas modules. We also make heavy use of the crosstalk²⁰ package to enable event-driven updates to the application state. The core iAtlas application is hosted on https://shinyapps.io.

Figure 1. iAtlas architecture overview.

(A) Structured data from immunogenomic analyses, including the Immune Landscape⁶ study and expanding over time, are organized and stored as flat (i.e., feather) files within Synapse and made available alongside the application code in GitHub. (B) Tabular data from feather files are read from disk into memory to drive all operations related to sample groupings, sample-level immune characterizations (readouts), and more granular—and high dimensional—assay measurements. (C) The core application code is built as a catalog of Shiny Modules, each of which encapsulates logic for data transformation and visualization related to a scientific theme or assay type. (D) Analysis modules, tools, and data description views are hosted in a unified application on shinyapps.io; the layout and connectivity between modules in the iAtlas Explorer space are managed by the shinydashboard²¹ library.

Analysis modules

The main feature of the iAtlas interface is the iAtlas Explorer (Figure 2, found under the EXPLORE tab), which provides several Analysis modules to explore and visualize results. Each module supports a type of exploration, with interactive views and controls to enhance and extend the results and analytics as initially described in the Immune Landscape⁶ study. The layout of pages and sections within the iAtlas Explorer is driven by the shinydashboard²¹ package.

Figure 2. iAtlas Explorer.

A range of Analysis modules (blue boxes above) are available that span from clinical to molecular and imaging data types. Within each module, interactive controls allow researchers to expand views, exposing underlying data and results. Settings are available (green box above) to select the sample groupings (TCGA Study, Disease Subtype, or Immune Subtype) which then propagate through modules.

Within each module in iAtlas, results are displayed relative to Sample Groups, corresponding to defined study cohorts. Several Sample Groups options are pre-loaded in the tool: first, TCGA tumor type (TCGA Study), which are the standard TCGA tumor types collected and designated by the TCGA. Second, TCGA tumor subtypes (TCGA Subtype), a compendium of further subdivision of TCGA studies into molecular subtypes according to publications by the TCGA Research Network²². Finally, a division of tumor samples into distinct patterns of immune response in cancer (Immune Subtypes) is provided⁶. The choice of Sample Groups is global across all modules but can be updated at any time via the Select Sample Groups element in the side menu. We also allow users to upload custom-grouped samples and analyze those with iAtlas modules. The selection of a sample group defines the samples utilized in all analysis modules. For convenience, group annotations can be displayed in visualizations within each module.

Integration with Landscape of IO Drug Target Development: CRI has compiled and published comprehensive overviews describing ongoing immunotherapy drug trials, including targets, agents, and tumor sites and has made summaries available in an online resource, the Immune-Oncology Landscape (IO Landscape ) (www.cancerresearch.org/IO-landscape)^25–28. The iAtlas and the IO Landscape resource have been interlinked, enabling researchers to more readily understand the relationship between targeted proteins in IO therapy and the behavior of those targets in tumor tissue.

In the opposite direction, clicking on targets in the barcharts in the IO Landscape on CRI web pages brings up the target gene expression in iAtlas.

Tools

iAtlas Tools are accessible via the TOOLS tab on the iAtlas Portal. Modules in this space of the portal enable users to “bring their own data” for processing through immunogenomic algorithms that drive some of the results presented in the Analysis modules described above.

Immune Subtype Prediction: This tool performs classification of RNA-seq data into one of six immune subtypes as described in the Immune Landscape⁶ study. Using a new ensemble model based on XGBoost²⁹, researchers can upload their own data for classification³⁰. Each member of the ensemble was trained on a random subset of previously reported immune subtypes⁶ and features (described below) based on gene expression data from the TCGA PanCancer Atlas Initiative⁸. All code and methods have been confirmed as reproducible. An R package is available on GitHub (https://github.com/CRI-iAtlas/ImmuneSubtypeClassifier)³⁰.

The submitted expression data—subsetted to the 485 genes that comprised the 5 signatures that produced the immune subtypes—are used to generate robust features of three types: quartiles, binary gene-pairs, and signature-pairs. For example, given a single sample, genes are binned into quartiles and given a bin label (quartile features). Then, similar to the “Top Scoring Pairs” classifier³¹, genes are paired, and given binary values depending on whether (g_i > g_j) for two gene expression values, g_i and g_j. Lastly, signature-pair features are calculated using the five immune subtype signatures, where s_mn = ∑_ij(g_im > g_jn)/k, where g_im is gene i from signature m, g_jn is gene j from signature n, and k is the number of gene pairs considered resulting in a value between 0 and 1. The features are computed independently for each sample, and do not require normalization across samples. These features are given to a trained XGBoost classifier which returns a probability of being in any of the six subtypes. Lastly, a “best call” is made with a final trained XGBoost classifier using the six probabilities as input. To validate the robustness of the classifier, TCGA data were processed using four different software pipelines and normalization, showing that classification performance was independent of the gene expression quantification method³⁰. Along with a downloadable table of results, visualizations are also provided. This tool is a convenient way for researchers to apply the methods of the Immune Landscape⁶ study to their own data without difficult statistical coding.

Reproducing published results and gaining information on underlying data

One of the initial motivations behind iAtlas was to provide an interactive platform that is able to reproduce figures published in the Immune Landscape⁶ manuscript but expands that with the ability to generate variations of those figures, for other choices of tumor samples and immune readouts of interest. As an example, in order to reproduce Figure 4A from the Immune Landscape⁶ publication, which shows the correlation of DNA damage measures with the fraction of leukocytes in the tumor, we began by selecting the EXPLORE tab. We then opened the Immune Feature Trends module and selected the “Immune Subtype” option under Select Sample Groups in the Explorer Settings panel in the left menu. In the ensuing module page, at the Correlations section (Figure 3), we selected the “DNA Alterations” under Select or Search for Variable Class, “Leukocyte Fraction” under Select or Search for Response Variable, and the “Spearman” method under Select or Search for Correlation Method (each a separate dropdown menu). This produced a heatmap identical in content to Figure 4A in the Immune Landscape⁶ publication. However, the heatmap provides additional information on underlying data via interactivity: by clicking on a heatmap-cell, the underlying data is displayed in a scatterplot. Hovering a cursor over a point in the scatter plot reveals sample-level information.

Figure 3. Visualization of the correlation of DNA damage in cancer with the degree of immune cell infiltration.

Top right: Original manuscript figure panel from the Immune Landscape⁶study. Bottom left: Equivalent figure generated in iAtlas. By selecting a specific heatmap cell (highlighted), the underlying data is displayed (Bottom right), using the selections shown. Individual points can be selected to get sample IDs and additional information (blue box).

Figure 4. Visualization of the PD-L1 immunomodulator expression patterns in breast cancer subtypes.

Selection for PD-L1 expression distributions displayed as violin plots within molecular subtypes of breast cancer, according to “PAM-50” classification. Elevated expression is seen in the HER2 subtype.

Table 2 lists the particular manuscript figures (from the Immune Landscape⁶ publication) that can be reproduced or adapted to specific research questions.

Exploring new IO results

With the iAtlas portal, scientists can explore and answer new questions based on specific research interests. For example, we asked: “What is the expression level of PD-L1, a therapeutically important protein, in subtypes of breast cancer?” To answer this question, from the landing page, we first selected the “TCGA Subtype” sample group, followed by the “Breast Invasive Carcinoma (BRCA)” study subset. Next, we selected the Immunomodulators module (Figure 4). Based on a very quick scan of the drop down, we didn’t see any names that matched our gene of interest, so we scrolled further down on the page to view the table of ‘Immunomodulator Annotations’. By typing in the first few letters of a gene name (e.g., “PD...”) into the ‘Search’ field, the table was filtered to a set of matching genes, and we could see that “PD-L1” is the Friendly Name for the gene “CD274” (the approved gene symbol on genenames.org). After returning to the Select or Search for Variable drop down menu above and selecting “CD274 (PD-L1)”, we were able to see a display of violin plots showing the distributions of gene expression across BRCA molecular subtypes. We could then visually compare distributions between subtypes, noticing for example the elevated expression level in the Her2 subtype compared to Basal breast cancer. These comparisons can guide further characterization not only of how gene expression can differ between TCGA subtypes of breast cancer, but also how these subtype-specific differences might correlate with clinical outcomes, as investigated in other studies^32–34. Using this module and others, the researcher has the ability to answer new questions which could lead to developments in oncology research.

Classification of immune subtypes on new data

In order to classify any tumor-derived gene expression samples into immune subtypes^6,30, users can select the TOOLS tab (top right), which leads to an interface containing notes, several links and the controls. In order to classify new data, we submitted data as a text file, in this case tab separated, with the first column containing gene IDs and later columns containing samples. A provided example file can be found in the description text. The first row of the data was a header containing sample IDs. Gene IDs can be either HGNC gene symbols (preferred), Entrez ID, or Ensembl identifiers. The locally available data was selected using the Browse button, and the file delimiter was selected, along with gene ID type, using drop down menus. Hitting the GO button produced classifications, signature scores, and cluster probabilities, which were reported in a table that was downloaded as a csv, xlsx, or pdf file. In addition, a barplot with the frequency of predicted subtypes for the submitted data was displayed.

All data required to run the application and describe the Use Cases are available in GitHub and archived with Zenodo⁷.

Source data

Original data files from the TCGA PanCancer Atlas publication can be found in the NCI Genomic Data Commons (https://gdc.cancer.gov/about-data/publications/panimmune) or the TCGA PanCancer Atlas Data Mirror (https://isb-cancer-genomics-cloud.readthedocs.io/en/latest/sections/PanCancer-Atlas-Mirror.html.

Extended data

Zenodo: CRI iAtlas (Version 1.2.0). https://doi.org/10.5281/zenodo.3926757⁷.

Folder ‘Data’ contains all data required to run the application and describe Use Cases. This is also available on GitHub.

License: Apache License 2.0.