From command-line bioinformatics to bioGUI

Existing (workflow) systems

There are several (workflow) systems already available. Most prominent in bioinformatics are the Galaxy server and Yabi. In addition, workflow specification languages such as the common workflow language (CWL) or Nextflow exist. These workflows do not directly compare to bioGUI because they (usually) require a server infrastructure and are not aimed to run on a local computer. However, they have in common that no CLI is needed to run (bioinformatics) applications.

With the R Gui Generator (RGG) a general GUI framework for R already exists. Recently, specialised GUI frameworks, like SEED 2 (Větrovský, Baldrian & Morais, 2018) or RNA CoMPASS (Xu et al., 2014), have been presented.

Use-case study

One of the main goals we had in mind when developing bioGUI is to create a powerful framework, which is easy-to-use for scientists/users and which does not create significant overhead for the application developer. In order to study this, we introduce two classes of possible users: The first class represents a general user of the software who generally prefers a GUI for performing a research task, for example, data analysis after sequencing. The second class describes a software developer releasing an application of a new algorithm to solve the alignment of sequencing reads. This class thus depicts a typical developer.

From these two use-cases (see also Appendix section “Use-cases”) we identify several requirements/goals for bioGUI:

1. installing new programs must be simple and should not require system administrators

2. creating a GUI for a program must not take a lot of time

3. templates must bring a basic GUI to run the programs, output must not be interpreted

4. templates must be saveable for later re-use and reference, and also searchable

5. the system must be lightweight (runtime overhead, disk-space) to even run on laptops

6. installing a program may require additional (protected) external files

Finally, we developed a paper mockup with which we went through the anticipated workflow of the user. We identified several input components and features the bioGUI program has to include (Fig. A1).

bioGUI approach

“The accessibility of bioinformatics applications is crucial and a challenge of contemporary biology” (Morais et al., 2018). Particularly the usage of CLIs poses a problem. Since most bioinformatics applications require the execution of commands on the CL for installation (such as for compilation, adding dependencies to the path variable, etc.), we estimate that also the installation poses a problem.

During the use-case study, and interviews with wet-lab scientists without a computational background (Q Emslander, 2019, personal communication; L Jimenez, 2019, personal communication), we found two main problems with bioinformatics applications for scientists which we want to address with bioGUI: first the installation of potentially useful applications and second its usage. Both problems have in common, that they are expected to be performed on the CL. A GUI for achieving the respective tasks in bioinformatics (and beyond) is missing.

Especially the first task, installing bioinformatics applications on a user’s machine, poses a few problems. Most bioinformatics applications are written for a UNIX operating system, like Linux or Mac OS, while in general Microsoft Windows is the dominant operating system. In order to overcome this problem, bioGUI makes use of WSL on Windows. Even if the user’s OS is already Unix-like, using the CL to install software might be strugglesome. Thus, in order to support all users, bioGUI uses a cross-platform approach. bioGUI is developed in C++ using the Qt framework.

The general workflow for any program using bioGUI is shown in Fig. 1. Given a CL application, the software developer (blue) writes the specific template in a XML-based DSL and can then make this template available, for example, in the bioGUI repository (cyan). Such templates can be automatically retrieved by bioGUI. Upon selection of a template by the user, bioGUI displays the input mask as defined in the template. When the user (yellow) has filled in all parameters, the parameters are collected by bioGUI and assembled into CL arguments which are used to execute the original CL-only application. Upon completion, simple results (like text-output or images) can be shown in bioGUI directly, or an external application is opened.

Install modules

Install modules are designed to install applications such that bioGUI can access them. Essentially, install modules are bash scripts which allow an automatic installation of applications into a predefined location. For this purpose, install modules receive several arguments from bioGUI when launched, for example, where to install the application to, the sudo password to fetch packages via a system’s package manager (e.g., aptitude, conda, …), whether the application should be made available to the user via the system’s PATH variable, etc. Install modules download and install applications and make them available to the user and bioGUI. However, some applications cannot be simply downloaded, but are distributed by installers. For this purpose, the install module template can be extended by further input fields. These must be specified by bioGUI elements and their values are added to the end of the CL arguments of the install module. An install module can then execute the referenced installer.

Finally, an install module should contain the specification of its bioGUI template and could hard-code the path to the installed application. Other constant values, which can already be derived during the installation (e.g., absolute paths to dependencies), could also be defined in the template during this stage.

bioGUI templates

bioGUI templates are the actual end-user-interface to programs. A bioGUI template defines the look and functions of the UI. Thus it can define how the CL-application is called (with corresponding parameters).

Each bioGUI template consists of two parts (Fig. A2). The first part (<window> model) defines the visual appearance of the GUI. The second part (<execution> model) defines the processing logic of the template. Input values from the GUI components are collected and assembled (e.g., pre-/post-processing steps) to call CL applications. As part of this assembly, input values from the GUI may be transformed using (multiple) predefined nodes. Concatenations are possible using the <add> node, and constant values can be inserted using the <const> node. System environment properties, such as the operating system, the computer’s IP address or specific directories can be collected using the <env> node. If the regular nodes are not sufficient, for example, because more complex string manipulations should be made (see use-case study), script nodes may also accept functions written in LUA (Lua, 2019) or JavaScript (JavaScript, 2019).

In general, the execution part infers a network with inputs (e.g., GUI elements, other nodes within the execution part) and actions (if, add, …). For example, the execution network for an application with many sub-commands is exemplarily shown in Fig. A3.

The time to template varies with the application as well as the number of options to be included. A simple template, like the one for MS-EmpiRe (Ammar et al., 2019), can be created within 10 min. More comprehensive templates, like the one for HISAT2, usually take about 30 min. Time can be saved if only the most important command line options are shown in the GUI. This can be achieved by adding an “optional parameters” input field, where users can insert CL arguments themselves. This is, for instance, shown in the wtdbg2 (Ruan & Li, 2019) and spades (Bankevich et al., 2012) templates. Adding the install part to a template usually can be done within 15–30 min, depending on how detailed the build process is documented. The creation of an install module thus takes approximately 1 h.

bioGUI integration with CWL and argparse

The CWL (Amstutz et al., 2016) only describes the CL workflow and neither provides a GUI nor means to install the desired tool. Due to this more general specification, CWL fits most problems, but specific annotations of inputs, explanations or the embedding of images is not supported in CWL.

While developers can always create templates manually, bioGUI supports developers by offering a template generator from CWL templates or python3 argparse CL parsers. Since there are already many CWL templates available for bioinformatics CL applications, CWL files can be used as a base to automatically generate bioGUI templates from. Using the bioGUI template generator for argparse, it is also possible to automatically generate templates from CWL files (making use of the cwl2argparse program provided by CWL). Our generator takes as input the argparse parser or CWL file and creates input elements for all elements. In case the type of an input is unclear or not supported, the generator falls back to a regular text-input element.

bioGUI templates

Currently more than 25 (install) modules exist for bioGUI. These represent basically three groups of bioinformatics tasks: next-generation sequencing data analysis and transcriptomics, long read sequencing analysis and assembly as well as more general sequence analysis. In general these install modules will install the respective application on the local machine. The Circlator (Hunt et al., 2015) template allows to pull and use the corresponding Docker image. The available tools, as well as their respective categorization, are listed in Table 1.

Benchmarking bioGUI templates

Our benchmark comprises of four tasks. The first task is to assemble a bacterial genome from Oxford Nanopore long reads, for which the Minimap2 (Li, 2018)/miniasm (Li, 2016)/Racon (Vaser et al., 2017) pipeline (available as install module from bioGUI) is used. The second task is the quantification of reads from a yeast mRNA sequencing project using Oxford Nanpore Reads and Illumina Reads (EMBL ENA studies PRJNA398797 (MinION) and SAMN00849440 (Illumina)). The quantification is performed using featureCounts from the subread package (Liao, Smyth & Shi, 2014). The third task uses these results to compute differential gene expression. Differential gene expression analysis is performed using MS-EmpiRe in R (install module available, Ammar et al. (2019)). Finally the fourth task uses RNAhybrid to predict miRNA binding sites (1,978 murine miRNAs) in 170 sequences of each 200 nt.

The results are shown in Table 2. The given runtimes are wall clock times. The peak RAM consumption has been sampled from the process viewer on the given operating systems (Task Manager on Windows, top on Linux and Mac OS).

Use-case analysis

Our use-case analysis (Appendix section “Use-cases”) has revealed several requirements for bioGUI (see the section “Methods”) to enable the user to perform the sequencing analysis and to allow the developer a fast template creation (Fig. 2).

An easy installation (goal 1) is given through the availability of install modules, which can be downloaded from the bioGUI repository and started via a GUI. These also allow additional inputs (e.g., Python wheels for albacore, goal 6).

The install modules combine the installation of an application and the creation of the actual GUI template. If the developers employ automatic testing of their software (e.g., build checks with Travis (https://travis-ci.org/)), the install part resembles a Travis container setup (goal 2): dependencies and the application itself are installed into an operating system. Even if not, most bioinformaticians extensively use Ubuntu and/or bash-scripts. Thus writing a script to install dependencies is not a significantly hard workload. We have reached a seamless and time-efficient creation of templates using an XML-based DSL. XML is particularly helpful as it allows to specify hierarchies and attributes to objects. Using our template generator for CWL and python3-argparse, bioGUI templates can be created even faster (goal 2). The templates are highly flexible in the creation of CL parameters, also due to providing script nodes. By providing install modules and templates, high-quality open-source bioinformatics applications become more accessible to the community.

The bioGUI application is cross-platform compatible and only requires few MBs of disk space (goal 5). bioGUI implements several possibilities to execute applications (see Fig. A4). In general, the only runtime overhead involved is the creation of a bash-process which starts the actual program with the assembled CL arguments (goal 5). bioGUI, being a local stand-alone application, has the possibility to target both, locally installed and web-based applications, reachable within a controlled environment and with large data. In addition, bioGUI also supports the use of Docker containers, for cases where all other options fail.

The UI can be made easily understandable (goal 3). Using text-labels, the user can get help on inputs (if specified by the developer), links can be used to provide further information and most importantly tool-tips could also hint the user to which information is needed at a certain step.

Finally, filled out templates can be saved via the Save Template button in bioGUI, and all available templates can be filtered (goal 4). This enables to keep track of performed analyses, and makes results more reproducible, because parameters are saved. Having the possibility to save templates also allows to easily repeat analysis with the same parameters. Additionally, using the bioGUI repository, templates can easily be shared among users, making it easier to standardize runs among different users or even institutes.

An anonymous survey (with 10 participants) about common problems in using bioinformatics software has been conducted among colleagues (n = 4) and under-graduate students or collaborators from life sciences (n = 6, short: collaborators). The results are available in the Appendix section “User Survey.”

We asked “What were the most struggle-some tasks in accessing and using the software?” referring to recently used bioinformatics software by the participants. Eight of the 10 participants answered that finding parameters or using the software has been strugglesome. This shows that the selected goals for bioGUI address actual problems faced by both experts and regular users. We further asked the participants to install and use graphmap (Sović et al., 2016), which has been selected because it is reasonably easy to install and use. First the participants were asked to install the tool using the CLI as well as via bioGUI. For this task, all instructions have been provided. The question “Has the installation process been easy?” (0 = No, 5 = Yes) has been answered with an average score of 4.4 for the CLI and 4.8 for bioGUI. Then the participants were asked to align the given reads against a given reference genome—this time without giving the instructions. Again we asked “Has it been easy to align the reads?” (0 = No, 5 = Yes). Here the CLI scored a 3.5 on average and bioGUI a 4.9. (Fig. A6). This coincidences with the answer to our question “Overall: Which interface was easier to use in your opinion?” (0 = CLI, 5 = GUI). Here the average score has been 4. Bioinformaticians and collaborators answered differently: the average bioinformatician has been undecided on which interface has been easier to use (average 3), but non-bioinformaticians preferred the GUI over the CLI (average 4.5).

The survey indicates that there are problems with bioinformatics software regarding installation and usage of CLI tools. These problems can be reduced by providing a GUI for these programs. The more experienced a user is on the CL, the less impact a GUI has. But particularly for non-experts on the CL, a GUI makes it easier to use a program.

bioGUI repository

We provide a repository of preconfigured templates on our website (Fig. A5), where authors and users can search for and browse existing templates, or submit new ones. bioGUI can access uploaded templates and save them directly for use. Specifically for WSL and Ubuntu users install modules are provided, which take care for dependency resolution and install applications (locally) into the user’s home. This currently works in any environment using the aptitude (apt) package manager, but users can submit templates which also support other environments, since install modules are versatile bash scripts. On Mac OS, some install modules support Homebrew for template installation. Install modules download and potentially severely alter a system (especially if the sudo password is supplied). Thus, submitted install modules are manually curated and are only accessible when no security threat has been identified.

The major goal of bioGUI is to enable any scientist to use bioinformatics applications. While we extend the repository on a regular basis depending on our own use, users can also request new templates for applications relevant to them.

Availability and extensibility

bioGUI is open-source software and users or institutions can either use our global bioGUI repository or deploy a custom repository, for example, one which is only reachable within an institution.

bioGUI is available on GitHub (https://github.com/mjoppich/bioGUI). Both source code as well as pre-built binary distributions (for Microsoft Windows, Linux and Mac OS) are available. While bioGUI will run on any Linux distribution, install modules currently use mainly aptitude as package manager (e.g., Ubuntu, debian-based distributions). If used on Windows, the same applies for the used WSL-application (Ubuntu 18.04 is recommended). bioGUI has been tested on Microsoft Windows 10, build 17763. On Mac OS bioGUI uses Homebrew (https://brew.sh/) to install dependencies. Homebrew does not support a silent, non-interactive installation: the user has to install Homebrew before running the First time setup for Mac OS install module which will then install the most common dependencies.

While a number of use cases and corresponding components are already included in bioGUI, we encourage users to contribute on GitHub by either pushing their own extensions, or opening feature requests. Further documentation (installation & setup guide, how to write templates) is also available via ReadTheDocs (https://biogui.readthedocs.io/en/latest/index.html).

Benchmarking bioGUI templates

bioGUI starts a sub-process for each executed program. Thus, the only overhead created by bioGUI itself is the one for running the GUI, which creates less than 1% CPU usage, allocates less than 50 MB and only performs IO operations when loading a new template (assessed via Sysinternals Process Explorer (Microsoft Sysinternals, 2019)).

Nonetheless we have been interested in demonstrating that many bioinformatics tasks do not require a dedicated server setup but can be performed on regular laptop computers. We thus benchmark four typical tasks performed using bioGUI.

The selected tasks allow a good overview of different demands: Tasks 1 (assembly) and 3 (differential expression analysis) are CPU-bound tasks, while tasks 2 (feature counting) and 4 (miRNA target prediction) are IO-bound. Particularly task 2 has a high load of read operations, and task 4 has a high demand of write operations. We compare these tasks on a dedicated (Linux) server, one (rather) powerful Lenovo T470p laptop computer, one Surface Book laptop computer (resource-wise a typical laboratory laptop) as well as one MacBook Air. The computer specifications are listed in the Table A1 and results are shown in Table 2. Even though we have not included the alignment of the Illumina yeast reads in this benchmark, it should be noted that this task also runs well on laptop computers. On the Lenovo laptop the alignment of the SRR453566 sample, consisting of 5,725,730 paired reads, has a peak RAM consumption of 34 MB and took 13:50 min, while the Surface Book is even faster at less than 8 min. This presumably can be explained due to different SSD speeds.

The results in Table 2 show that even (computational power-wise) lower-end computers can run bioinformatics tools. More interestingly these results show that the WSL allows the execution of interesting bioinformatic tools. It can be seen that WSL is slow for IO operations, but has a comparable speed for in-memory operations. Particularly tools requiring a lot of IO are considerably faster on the Linux Server (Assembly, RNAhybrid), while the computational expensive tools like MS-EmpiRe (Ammar et al., 2019) and featureCounts (Liao, Smyth & Shi, 2014) run within similar times.

Use-cases

bioGUI paper mockup

Extending templates with script nodes

Often it is required to perform string-manipulations (e.g., remove file extensions) for CL arguments. For instance, the example below takes as input a HISAT2 index file, and removes the file extension, such that the index will be accepted by HISAT2. For evaluation of this node, the evaluate-function is called with the argv-references as input parameters. The last return-value of the script’s call stack is taken as output value of the script -node.

Listing 1. bioGUI script node with LUA function example. Upon evaluating this node, the evaluate function will be called with the arguments listed in the argv attribute of the script node.

<script id=“hisat_index_rel” argv=“${hisat_index_rel_raw}”>

<![CDATA[

function evaluate(arg1)

    if (string.match(arg1, “.%d.ht2$”)) then

      return(string.sub(arg1, 0, arg1:find(“.%d.ht2$”)-1))

    end

   return(arg1)

  end

]]>

</script>

Evaluating a bioGUI template

In Fig. A2 the process of assembling a CL call from the shown bioGUI template is explained. First, the creation of the <window> model (dark gray) will be explained, followed by the creation of the CL arguments using the <execution> model (shaded).

The window component consists of four different components, which are grouped in a vertical layout (default for window component). A label describing the input file dialog is placed on the main window, followed by the actual file dialog with ID input. Then a group box with title and a checkable status is created, which contains an output file dialog. Finally, the action button, which starts the CL call assembly and the subprocess of this program, and the text output elements are created.

When the user has entered all desired data, and clicks the action button, the execution phase defined by the execution model will be launched. Therefore the program defined in the execute element is started. For this, the parameters (param) must be assembled. Any text within ${var} is interpreted as a reference to a variable var or the value of a GUI element with id var. Thus, the CL is successively assembled. At first the ${input} element is interpreted and retrieves the value from the input file dialog as this element matches the id. Next the ${output} is interpreted. The ${output} refers to an if construct in the execution part, which compares the value of the element with id os to the string TRUE (which is whether the groupbox is checked). If this value is true, this node evaluates to netcat 192.168.1.100 55025, otherwise to tee -a {output file path}. Finally, the program sh is executed with the created CL arguments. For instance, if the group box is checked, the sh -c “cat inputFile | netcat 192.168.1.100 55025” will be executed. A full reference of all input types as well as all execution nodes is available online.

In fact, the evaluation of the execution network resembles the simulation of a petri net (Fig. A3). Each node in the execution network is a place, and its modification/function is the transition, which requires values for all its input places, to generate the output token.

Running programs via bioGUI

Program execution via bioGUI can be accomplished via different paths, which are shown in Fig. A4. The easiest way is to execute a native program (one that runs natively on the operating system, e.g., Docker). Then all output can be piped to bioGUI to display this to the user. If the host is a Microsoft Windows 10 OS, bioGUI can also run Unix programs via WSL. Then the Unix program runs natively in a WSL bash. The resulting output can be transferred to bioGUI via pipes. Of course, for both native and WSL processes, the output can also be transferred via netcat to bioGUI. The transfer of the GUI template within install modules is an example. If a process runs on a remote computer, the output can be transferred to bioGUI also via network, for example, netcat. Such a process can, for instance, be started by calling ssh from bioGUI with appropriate parameters. Finally bioGUI can also send HTTP POST requests to web-services and accepts an HTTP response as answer. This output can also be displayed by bioGUI.

Since the Docker engine is a local, native process, bioGUI also supports the use of Docker containers. The Circlator template is an example of how this can be implemented.

Hardware specification for benchmarks

The relevant hardware for benchmarking bioGUI is summarized in Table A1.

Template access

User survey

A user survey among 10 participants (four bioinformaticians, six collaborators (consisting of two under-graduate bioinformatics students and four external collaborators)) has been performed. The derived results are shown in Table A2 and the raw data are shown in Table A3.