iSEE: Interactive SummarizedExperiment Explorer

Using a multi-panel layout

All data aspects stored in a SummarizedExperiment can be simultaneously examined in the multi-panel layout of the iSEE interface (Figure 1A). The interface layout is built using the shinydashboard package (Chang & Borges Ribeiro, 2018), with colour-coded panels to visualise each data aspect. Individual panel types include:

Figure 1.

iSEE uses a customisable multi-panel layout (A) that simultaneously displays one or more panels of various types, where each panel type visualises a different aspect of the data. New panels of any type can be added (i), and all panels can be removed, reordered or resized (ii). Panel types are available to visualise sample-based reduced dimensionality embeddings (iii), sample-level metadata (iv), and experimental observations across samples for each feature (v). Other panel types include row statistics tables (vi), to facilitate searching across features and their metadata; heatmaps (vii), to visualise experimental observations for multiple features; and feature-level metadata plots. Panels of each type are colour-coded for ease of interpretation. (B) Information can be transmitted between panels according to a user-specified scheme. Here, the selection of feature X in the row statistics table determines the y-axis of the feature assay plot, and colours the samples in the reduced dimension plot by the expression of X. Selection of points in the reduced dimension plot (dotted blue line) also determines the samples that are shown in the column data (i.e., sample metadata) plot; further selection of points in the column data plot determines the samples that are shown in the heatmap.

Each sample is represented as a point in column data, feature assay and reduced dimension plots. Similarly, each feature is represented by a point in row data plots. For these panel types, a scatter plot is automatically produced if the selected variables on the x- and y-axes are both continuous. If exactly one variable is categorical, points are grouped by the categorical levels and a (vertical or horizontal) violin plot is produced with points scattered within each violin. If both variables are categorical, a “rectangle plot” is produced where each combination of categorical levels is represented by a rectangle with area proportional to the frequency of that combination. Points are scattered randomly within each rectangle. For ease of interpretation, the rectangle plot collapses to a mirrored bar plot when one of the categorical variables only has one level.

Custom panel colouring

Sample-based points can be coloured according to the values of any sample-level metadata field in the colData slot or by the assay values of a selected feature. Similarly, feature-based points can be coloured according to any feature-level metadata field in the rowData slot. Heatmaps are coloured according to the expression values of the selected features in the chosen assay, with additional colour annotation for each of the colData fields used to order the samples. In all cases, the variable to use for colouring can be dynamically selected for each plot. This enables users to easily examine relationships between different variables in a single plot.

By default, colour maps for categorical and continuous variables are taken from the ggplot2 (Wickham, 2009) and viridis packages (Garnier, 2018), respectively. However, iSEE also implements the ExperimentColorMap class, which allows users to specify arbitrary colour maps for particular variables. Each colour map is a function that returns a vector of distinct colours of a specified length, and will be called whenever the associated variable is used for point colouring in a particular panel. The returned colours will be mapped to factor levels for categorical variables, or used in colour interpolation for continuous variables. For categorical variables, the function may also return a constant vector of named colours corresponding to the levels of a known factor. Colour maps can be specified for individual variables; for all assays, all column data variables, or all row data variables (with different functions for continuous or categorical variables); or for all categorical or continuous variables. This provides a convenient yet flexible mechanism for customisation of colouring schemes within the interface.

Dynamic linking between panels

A key feature of iSEE is the ability to dynamically transmit information between panels (Figure 1B). Users can define and reorganise arbitrary links between “transmitting” and “receiving” panels, whereby selections in transmitting panels control the inclusion and appearance of the corresponding data points in receiving panels. This feature facilitates exploration of the relationships between different aspects of the data. For example, users can easily determine co-expression patterns of genes in a particular region of a reduced dimensionality embedding – this is achieved by selecting points in a reduced dimension plot (using the standard rectangular brush or a lasso selection) and transmitting that selection to any number of feature assay plots.

This linking paradigm extends to multiple panels, whereby a panel can transmit to multiple receivers, and a receiving panel can transmit its own selection to another plot. Chains of linked plots allow users to mimic the arbitrarily complex gating strategies often found in analyses of flow cytometry data (Finak et al., 2014). With iSEE, this concept is extended to any assay data, feature-level or sample-level metadata present in a SummarizedExperiment object, providing a powerful framework for interrogating multiple interactions between data aspects. Row statistics tables can also transmit to various plot types, by selecting a table row to control the colouring of sample-based points; or by defining a subset of features to visualise in a heatmap. Furthermore, row data plots can transmit to row statistics tables, whereby selection of points in the former will subset the latter.

Code tracking and reproducibility

iSEE automatically memorises the exact R code that was used to generate every plot, extending previous work by Marini & Binder (2016). This code is fully accessible to users at any time during the run-time of the interface. By integrating the code reported by iSEE into their own scripts, users can easily reproduce the results of any exploratory analysis. Similarly, the code required to reproduce the current state of the interface can also be reported. This can be used in startup scripts to launch an iSEE instance in any preferred layout, including the panel organisation, variable selection, colouring schemes, links between panels and even individual brushes and lasso selections.

Additional functionalities

Row statistics tables can be augmented with dynamic annotation based on the selected row, linking to online resources such as Ensembl (Zerbino et al., 2018) or Entrez (Coordinators, 2017). For large data sets, points can be downsampled in a density-dependent manner to accelerate rendering of the plots, improving the responsiveness of the interface without compromising the fidelity of the visualisation. Users can also include a bespoke step-by-step “tour” of their data set via the rintrojs package (Ganz, 2016), guiding the audience through an examination of the salient features in the data.