DataViz: visualization of high-dimensional data in virtual reality

Implementation

In the design of this application, special consideration was given to the following elements: the method of data analysis, the format of the input data, the limitations in computing power of the selected platform, and the mitigation of motion sickness.

Data analysis

The primary method of data analysis is PCA. The rationale behind this decision is that because humans live in three dimensions, the most intuitive manner of visualization is one that plots in that space. In this sense, PCA is excellent at taking large dimensional data and reducing them to plottable 3D coordinates, making the resulting graph more intuitive, and helping users discover patterns and develop scientific intuition.

Input data

DataViz only accepts data in the table format (CSV or TXT). Occasionally, the user would want to analyze the transpose of the provided data. Although the transpose of a table could easily be found using specialized functions in Numpy or R, we decided to add the transpose functionality into the application.

In addition to transposition, DataViz also allows the user to omit specific columns from the file. This may be due to a variety of reasons including an unwanted dimension of data or column names. This functionality allows researchers to analyze only the columns they are interested in.

The user may also have a column that labels the points. Users can designate a specific column that differentiates the data with various tags, and these groups will show up in a graph legend during runtime.

Limitations in computing power

The engine used in developing this application is Unity®. Unity is one of the most popular platforms for VR development but is not specifically designed for statistical analysis. Therefore, PCA on large datasets may result in slow run times, especially when there is a lack of an appropriate graphics card or other computational power involved. To overcome this limitation, the application can also accept coordinate data derived from PCA or other dimensionality reduction methods such as t-SNE⁷. In this manner, users can circumvent the slower computations associated with Unity.

VR considerations

When implementing the VR aspect of the application, we concentrated on two main considerations: immersion and motion sickness. For the former, the primary goal was to allow the user to focus on the graphical rendering of his/her data without being bothered by the complicated details on how to use the tool. In pursuit of this, we designed an intuitive interface and series of menus, with clear instructions on the associated GitHub page in ‘Software Availability’.

Another concern when designing for VR was motion sickness. Motion sickness is a consequence of conflicting input between visual and inner ear senses and is a major problem in current VR simulations⁸. It has been found that motion sickness is a consequence of the action of motion and not displacement itself, and as a result, we designed our movement to be in short bursts of teleportation.

The application is built using the Unity® engine with scripting done in C#. The PCA and transpose implementation is from the Accord.Net 3.8 framework (http://accord-framework.net). The mouse embryonic development data used in the case study is from Ref 9.

Operation

DataViz was designed to be an intuitive application for graphically rendering large datasets. Upon opening the software, a user should follow the onscreen prompts and fill out the appropriate parameters to input their dataset as well as use the extra functionalities described above. DataViz automatically runs PCA on the input dataset according to user configurations. If needed, more detailed instructions can be found on the associated GitHub page.

VR is a resource intensive activity. The following are guidelines for ensuring the quality and performance of DataViz.

System Requirements (https://www.vive.com/us/ready/):

Mouse embryo development

The primary goal in the development of this application was to determine the viability of using VR to graphical render and analyze complex data sets. After development, we tested the DataViz by analyzing a high-dimensional dataset regarding mouse embryonic development⁹. Using single-cell RNA sequencing (scRNA-Seq), Deng et al. generated hundreds of expression profiles of individual embryonic cells from zygote blastocyst stages.

By graphically rendering the 3D PCA projection of the data and subsequent analysis, we were able to verify an expected trend of embryo development; initial cell division (zygote stage to 16-cell stage) results in large-scale physical changes inside the embryo. This is in contrast to later cell division where the various stages of embryo development are more similar to one another. We can also see the developmental trajectory in the transcriptomic landscape (Figure 1).

Figure 1. The application when plotting provided mouse embryonic development coordinate data.

The graph displays the similarities among the blastocyst stages in comparison to changes in earlier stages of development. We can identify categories and general trends of the data using this method.

This method of analysis has some limitations, the foremost being an inability to account for all the data present. While reducing high-dimensional data to three dimensions simplifies the resulting plot and may help formulate testable hypotheses through further research or build intuition and comprehension regarding the data provided, it is inevitable that we lose some of the variance present in higher dimensions. In this test case, Table 1 reveals the proportion of the data retained per principal component. One way of overcoming this would be to use non-linear dimensionality reduction methods like multidimensional scaling (MDS) or t-SNE.

Despite the shortcomings involved in the provided analysis and plotting approach, DataViz is still useful for categorizing the data into disjoint groups.