BioCompute Objects.

Establishing better conceptual descriptions of workflows is important for reproducibility [48]. Workflow languages are machine-readable files that should be able to provide all the details needed to re-execute workflows. However, usually there is insufficient information that is needed to fully reproduce a workflow. For this, the IEEE 2791-2020 BioCompute Objects (BCO) standard was developed [26]. BCO is a rigorously defined standard for bioinformatics analysis workflow documentation that is flexible enough to accommodate any pipeline, but rigid enough to define a structure for computable metadata to annotate workflows. There is an ecosystem of tools that are compatible with the BCO standard, including cloud genomics platforms like Seven Bridges Genomics, CAVATICA, and DNAnexus. The BioCompute Portal is part of this ecosystem, and acts as a repository of published BCOs, as well as a place to manually build BCOs. The portal is provided with several published examples [49,50]. Using the PWB interface, a BCO can be constructed from any workflow. The exported BCO contains full provenance about the workflow including description of the steps, data, versioning, and authorship. The serialized BCO specifications can be downloaded as well as submitted to the BioCompute Portal via API where they can be inspected, modified with additional annotations, or extended to other schemas, and ultimately published.

Common workflow language (CWL).

Common Workflow Language (CWL) is an open standard for describing how to run command line tools and connect them to create workflows [21]. A command line interface (CLI) was developed from the KRG to invoke any Process metanode, providing inputs in JSON-serialized files, and writing the output to a JSON-serialized file. Using this CLI, a CWL CommandLineTool specification can be constructed out of any Process metanode, and a CWL Workflow specification and input variables file can be constructed from the FPPRG database. License, authorship, versioning information, and descriptions for the individual metanodes in use, are included in the resulting CWL output. All Prompt data that may be captured via interactions between the user and the UI are instead specified in the input variable file. Hence, the PWB platform metanodes are fully compatible with CWL, and CWL workflows can be exported from the PWB interface.

Research object crate (RO-Crate).

RO-Crate is a community-based specification for research data packaging of Research Objects (RO) with rich metadata, based on open standards and vocabularies including JSON Linked Data (JSON-LD) and schema.org [51]. Adopting a similar structure to describing workflows as WorkflowHub [52], an RO-Crate can be created from entries in the FPPRG database containing complete workflow provenance including license, authorship, versioning information and descriptions for the individual metanodes in use. The RO-Crate can then be used for registering PWB workflows in WorkflowHub and for minting citable Digital Object Identifiers (DOIs) for published workflows.

External API-driven metanodes.

The PWB platform is agnostic to the actual implementation of each step but a significant portion of the currently implemented metanodes rely on external API services. These services can more efficiently provide the most up-to-date knowledge as the data behind these APIs evolves. The downside of this is that reproducibility of workflows in not guaranteed. To mitigate this risk, we annotate the relevant steps that rely on external services and save the outputs of those services with a timestamp when the execution occurred.

RNA-seq data analysis and visualization.

Beginning from a user-uploaded count matrix of gene expression, where each row represents a gene, and each column is a sample with associated metadata, data is uploaded to the PWB and encoded with AnnData [53]. From the gene expression matrix, several metanodes enable different normalization and data visualizations. For example, the contents of the data matrix can be visualized with PCA [54], UMAP [55], or t-SNE [56] plots. These metanodes are supported by the Scanpy Python package [57]. From the data matrix, users can compute differential expression to produce gene expression signatures. Differential expression analysis can be performed with methods such as the Characteristic Direction [58], limma-voom [59,60], or DESeq2 [61]. Differentially expressed genes can be used as input for downstream analysis such as enrichment analysis which is described next.

Enrichment analysis.

Enrichment analysis can be performed within the PWB using the Enrichr API [62]. The gene sets data type in the PWB can be submitted for enrichment analysis against the many gene set libraries available from Enrichr. For example, the GTEx [63] and ARCHS4 [64] tissue expression gene set libraries can be selected to obtain ranks of more relevant tissues. Similarly, the KEGG [65] and WikiPathways [66] gene set libraries can be used to prioritize relevant cell signaling pathways. Enrichr also provides an API to search for metadata terms across the Enrichr gene set libraries. For example, a disease term search can be used to construct a consensus gene set [67]. Another way to obtain gene sets is through literature search. By submitting to PubMed any search term, term-gene co-mentions in publications can be used to construct the most relevant gene set for any search term. This functionality is supported by PWB via the Geneshot API [68].

Gene set manipulation.

The gene matrix transpose (GMT) file format is commonly used to serialize gene set libraries. GMT files contain lists of terms followed by sets of genes associated with each term. GMT files can be uploaded and analyzed by the PWB platform. A common way to interrogate the overlap between several gene sets is through UpSet plots [69] or SuperVenn diagrams [70]. The PWB has metanodes to display interactive versions of such plots. Additionally, several operations are implemented to transform data types from one to another. For example, turning ranked lists of genes into gene sets by choosing a cutoff, turning multiple gene sets into a GMT file, or collapsing a GMT file into a single gene set by applying a consensus or a union set operation on it.

Healthy human tissue expression atlases.

GTEx has profiled postmortem tissues from healthy donors with RNA-seq to produce gene expression data matrices from 54 human tissues [63]. The GTEx API can be used to find median tissue expression levels for all human genes for each one of these 54 profiled tissues. Similarly, the ARCHS4 resource [64] was created by uniformly aligning approximately 2 million publicly available RNA-seq samples collected from human and mouse. The ARCHS4 API can also be used to find median tissue expression across over 200 tissues and cell types. The PWB enables users to obtain summary statistics from these APIs, which can be visualized as bar graphs. It is also possible to use these data resources as a baseline to identify novel drug targets. For example, gene expression data collected by RNA-seq from tumor samples, can be compared to all normal tissue to identify genes that are only highly expressed in the tumor using the TargetRanger API [71].

Metanodes created from LINCS resources.

The Library of Integrated Network-Based Cellular Signatures (LINCS) program [72] profiled the response of human cells to thousands of chemical and genetic perturbations followed by omics profiling. The PWB provides several metanodes related to prioritizing drugs and preclinical small molecules for targeting individual genes and gene expression signatures. For example, a metanode can be used to perform LINCS L1000 reverse search queries for a given gene, producing interactive visualizations and tables of significant LINCS L1000 chemical perturbagen signatures that may maximally increase or decrease the expression level of the single human gene. A similar metanode was implemented to provide search against the L1000 CRISPR KO signatures. Other metanodes enable users to query the SigCom LINCS database [73] with gene expression signatures or gene sets. Such signatures may be in the form of a vector of differential gene expression, or up- and down-regulated gene sets. Both types of input signature queries can yield ranked lists of chemical perturbations and CRISPR KOs.

Metanodes created from GlyGen resources.

GlyGen is an international initiative funded by the NIH to promote research about glycoscience [74]. The GlyGen consortium developed a web-based portal that brings together glycan and protein specific data from major resources such as UniProt [75], GlyConnect [76], Protein Data Bank (PDB) [77], UnicarbKB [78], ChEBI [79] and PubChem [80] and other resources [81]. These datasets are presented to users through a standardized data model [82] via the GlyGen data portal (https://data.glygen.org). The GlyGen API endpoints (https://api.glygen.org) facilitate the same functionality provided by the user interface, providing the PWB with several GlyGen metanodes that can be integrated in various workflows. The GlyGen metanodes also support data visualization and kinase enrichment analysis. Furthermore, the GlyGen metanodes operate several core data types such as, glycans, proteins, and glycoproteins. For other glycoconjugate species, such as glycolipids, GlyGen metanodes implement the passthrough search APIs to the GlySpace alliance [83] and other resources. In addition, uploaded mass spectrographic glycan files are analyzed with various GlyGen specific metanodes, and then knowledge is extended with other PWB metanodes.

Metanodes created from metabolomics resources.

The Metabolomics Workbench (MW) is another resource supported by the NIH CF [84]. MW contributed several metanodes to the PWB including those from the bioinformatics tools MetGENE [85], MetENP [86], and a gene ID conversion tool. These tools, originally designed to be stand-alone web applications, provide REST APIs to obtain relevant information for analyses related to profiled metabolites within the PWB. MetGENE is a hierarchical, knowledge-driven tool designed for gene-centered information retrieval. By entering a single gene, or a set of genes, users can access information related to the gene such as pathways, reactions, metabolites, and studies from metabolomics in MW. To refine searches, MetGENE incorporates filtering options based on organism, tissue or anatomy, and disease or phenotype. This feature provides tailored and context-specific search experience. Several metanodes using MetGENE are implemented that take as input either a gene, or a gene set, for downstream analyses. The relevant functionality from MetENP is provided via a REST API called MetNet. Briefly, given a list of metabolites, e.g., metabolites with significant change between two conditions such as disease/normal or treatment/control in a metabolomics study obtained by using MetENP or another tool, a researcher may want to find what are the pathways and functions affected. MetENP/MetNet facilitates metabolite name harmonization using RefMet [87], metabolite class enrichment, metabolic pathway enrichment and visualization, and identification of reactions related to the given metabolites and genes coding for enzymes catalyzing these reactions. In MetNet, the list of these genes can be used to develop their protein-protein interaction (PPI) subnetwork using the STRING database APIs [88]. Each of these metanodes has an associated table that renders the information obtained from the API.

The connect the dots (CTD) metanode.

The Connect the Dots (CTD) metanode takes as input a set of genes or proteins and identifies a subset of genes or proteins that are highly connected within either knowledge graphs or networks derived from gene expression, metabolomic or other omic datasets [89]. CTD algorithm has previously discovered multi-gene biomarkers of drug response to breast cancer therapies based on mouse PDX models [90], and metabolomic signatures of rare inborn errors of metabolism [89,91]. While CTD has been previously deployed as independent R and Python packages (https://github.com/BRL-BCM/CTD), its deployment on the Playbook will allow for its use by a wider scientific audience. The CTD workflow starts with an input set of genes. The user then has the option of identifying significant connections within this set in the STRING protein-protein interaction network [88], WikiPathways [66], or a network derived from user-supplied data. The networks represented as weighted graphs, can be derived from expression data, proteomic data, metabolomics, or any other normalized omic dataset. This allows for users to identify highly connected sets of genes within their specific disease, treatment, or condition of interest. Given a weighted graph and a set of graph nodes as an input, CTD identifies significant highly connected subsets. An optional “guilt by association” feature identifies neighboring nodes using probability diffusion. CTD also returns a visual display of the nodes and connections.

Metanodes created from ERCC resources.

The ExRNA Communication Consortium (ERCC) Common Fund (CF) Data Coordination Center created a framework and toolset for FAIR data, information, and knowledge that delineate the regulatory relationships between genes, regulatory elements, and variants, and made them available to PWB via metanodes. We have implemented the ClinGen Allele Registry (CAR) and Genomic Location (GL) Registry [92], variant and genomic region on demand naming services, respectively. The CAR canonical identifiers (CAid) or Genomic Location identifiers (GLid) provided are reference genome-agnostic, stable, and globally unique. The ERCC metanodes enable the retrieval and mapping of unique identifiers and other commonly used identifiers, such as dbSNP IDs [96], connected through the Allele Registry and GL Registry using the Allele Registry RESTful APIs. Moreover, we have created the CFDE Linked Data Hub (LDH) [93], a graph-based database, to extract and link tissue and cell type-specific regulatory information from SCREEN [94], GTEx [63], and other CF projects, including Roadmap Epigenome [95] and EN-TEx [96]. Each excerpt on the CFDE LDH is created in a machine-readable format and contains a link to the original data source for provenance tracking. The CFDE LDH RESTful APIs provide read and write capabilities for both accessing and contributing gene regulatory information. This enables the CFDE LDH to connect more than 800 million regulatory data and information documents, which can be quickly retrieved by PWB through the API endpoints given any variant, regulatory region, or gene as input.

Use case 13: Cell surface targets for individual cancer patients analyzed with Common Fund datasets.

The input to this workflow is a data matrix of gene expression that was collected from a pediatric tumor from the Kids First CF program [24]. The RNA-seq samples are the columns of the matrix, and the rows are the raw expression gene counts for all human coding genes. This data matrix is fed into TargetRanger [71] to screen for targets that are highly expressed in the tumor but lowly expressed across most healthy human tissues based on gene expression data collected from postmortem patients with RNA-seq by the GTEx CF program [63]. Based on this analysis, the gene Insulin-like growth factor II m-RNA-binding protein 3 (IMP3) was selected because it was the top candidate returned from the TargetRanger analysis (Table 2). Next, we leveraged unique knowledge from various other CF programs to examine knowledge related to IMP3. First, we queried the LINCS L1000 data [97] from the LINCS program [72] converted into RNA-seq-like LINCS L1000 Signatures [98] using the SigCom LINCS API [73] to identify mimickers or reversers small molecules and CRISPR KOs that maximally impact the expression of IMP3 in human cell lines. These potential drugs and targets were filtered using the CF IDG program’s list of understudied proteins [99] to produce a set of additional targets. Next, IMP3 was searched for knowledge provided by the Metabolomics Workbench MetGENE tool [85]. MetGENE aggregates knowledge about pathways, reactions, metabolites, and studies from the Metabolomics Workbench CF supported resource [84]. The Metabolomics Workbench was searched to find associated metabolites linked to IMP3. Furthermore, we leveraged the Linked Data Hub (LDH) API [93] to list knowledge about regulatory elements associated with IMP3. Finally, the GlyGen database [74] was queried to identify relevant sets of proteins that are the product of the IMP3 genes, as well as known post-translational modifications discovered on IMP3. The discovery of IMP3 is not completely novel, IMP3 has been previously reported to be aberrantly expressed in several cancer types and its high expression is associated with poor prognosis [100].

Download:

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

Table 2. Ranked list of targets identified by TargetRanger to be highly expressed in the tumor sample and lowly expression across normal tissues from GTEx.

https://doi.org/10.1371/journal.pcbi.1012901.t002

Use case 1: Explaining drug-drug interactions.

This workflow takes as input an adverse event term and two drugs. The adverse event is identified in several databases that contain gene sets already associated with the adverse events and mammalian phenotypes related to the adverse event. Namely, matching adverse events and mammalian phenotypes are identified from the GWAS Catalog [101], MGI Mammalian Phenotype Ontology [102], and from the Human Phenotype Ontology (HPO) [103]. A set of consensus genes associated with the matching terms is assembled. Then, the workflow queries the LINCS L1000 chemical perturbation signatures [73] with the two input drugs to find gene sets that are consistently up- or down-regulated by the treatment of human cell lines with these drugs. The consensus gene sets impacted by the drugs, and the gene set related to the adverse events are then compared and visualized using a SuperVenn diagram to highlight overlapping genes between these sets. Genes of interest are those affected by both drugs and are associated with the phenotype. Such overlapping genes can be further interrogated individually for evidence in the literature, or as a gene set using enrichment and network analyses.

To demonstrate the workflow for a specific instance, we start with the adverse event “bleeding” and the drugs warfarin and aspirin. It is known that these drugs interact to increase the risk of internal bleeding [104] but the exact intracellular mechanism of such interaction is still not fully understood. The workflow starts with selecting “bleeding” as the search term. Gene sets with set labels containing the word "bleeding" were queried from Enrichr [1]. Identified matching terms from the GWAS Catalog 2019 [2], MGI Mammalian Phenotype Level 4 2019 [3] and the Human Phenotype Ontology [4] libraries are then assembled into a collection of gene sets. A GMT file is extracted from the Enrichr results for all the identified gene sets from each library and then these are combined using the union set operation. Gene sets with set labels containing the terms warfarin and aspirin were next identified from the LINCS L1000 Chem Pert Consensus Sigs [5] library. The gene sets collected for each drug were combined into one gene set library. The collection of gene sets was then visualized with a SuperVenn diagram (Fig 6). This analysis identified 243 genes up-regulated and 245 genes down-regulated by warfarin; 249 genes up-regulated and 244 genes down-regulated by aspirin, 85 genes associated with bleeding from MGI, and 35 from HPO. Only one gene, namely THBS2, is up regulated by both drugs, and is also associated with bleeding related phenotype in MGI. While the gene SLC7A11 is downregulated by both drugs and is linked to an MGI bleeding phenotype. THBS2 is a member of the thrombospondin family, and as such it plays a critical role in coagulation. It was shown that knockout mice of THBS2 have an increased bleeding time phenotype (MP:0005606) [105] and THBS2 is a potent inhibitor of tumor growth and angiogenesis [106]. It is difficult to explain why both drugs are found to up-regulate this gene. The expected effect is that these drugs would reduce the expression of the genes to reduce coagulation. At the same time, both drugs are also found to down-regulate the expression of the amino acid transporter SLC7A11. SLC7A11 knockout mice also have an increased bleeding time phenotype (MP:0005606), and mutations in this gene have been implicated in many acute human diseases through induction of ferroptosis [107,108]. Hence, for SLC7A11 the direction of the impact of the drugs on its expression is consistent with other prior evidence.

Download:

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

Fig 6. SuperVenn diagram to visualize the overlap between sets of genes that are up and down regulated by aspirin and warfarin based on LINCS L1000 signatures, as well as knockout mouse, HPO, and GWAS phenotypes associated with the term “bleeding”.

The permanent URL for a description of this workflow is: https://doi.org/10.48546/WORKFLOWHUB.WORKFLOW.1237.3.

https://doi.org/10.1371/journal.pcbi.1012901.g006

Use case 11: Related proteins/metabolites across DCCs.

The enzyme ribulose-5-phosphate epimerase (RPE) participates in the catalysis of the interconversion of ribulose-5-phosphate (Ru5P) to xylulose-5-phosphate (Xu5P) in the pentose phosphate pathway. A recent study [109] focused on the biophysical and enzymatic characterization of RPE in several organisms. Interestingly, the study suggested that RPE may play a crucial role in protection against oxidative stress. Toward integrative analysis to further elucidate the roles of RPE in various pathways and mechanisms of human disease, we collected knowledge about PRE from various NIH CF programs and other sources. The collected information about RPE includes: 1) Associated metabolites from the Metabolomics Workbench [84]; 2) Expression across human tissues from GTEx [63]; 3) Small molecules and single gene knockouts that maximally induce the expression of RPE from LINCS [73]; 4) Associated variants from ClinGen via LDH [110]; 5) Protein-protein interactions from STRING [88]; and 6) Regulation of RPE by transcription factors from ChEA3 [111]. In addition, the use case converts RPE into a gene set using the Geneshot API [68]. The Geneshot API returns a set of 100 genes that mostly correlate with RPE based on thousands of human RNA-seq uniformly processed from GEO [112]. Co-expression correlations computed from the data processed by ARCHS4 [64]. The comprehensive approach to find knowledge about a single gene is also applied to the generated gene set with all six resources. The final report provides a mechanistic understanding of how RPE can affect various pathways and functions despite not being involved in the pathways and processes directly.

Use case 10: Identifying gene regulatory relationships between genes, regulatory elements, and variants.

This workflow takes as input one or more genes, regulatory elements, or variants. One may then query for regulatory relations of the selected entity type with other entity types. In one application, we may ask what genomic regions regulate a gene of interest and what evidence supports that regulatory relationship. We start the workflow by providing the gene of interest as input. We first focus on regulatory elements that are in the vicinity of the gene body identified using the epigenomic data from NIH Roadmap Epigenomics [95] and ENCODE projects stored in the ENCODE SCREEN database [94]. Regulatory evidence associated with the SCREEN regulatory elements was connected to genes and variants using CFDE LDH [93], a graph-based database that facilitates the linking of findable, accessible, interoperable, and reusable (FAIR) [113] information about genes, regulatory elements, and variants retrieved through well-documented RESTful APIs. The available regulatory information includes: 1) Variants associated with regulatory elements from the ClinGen Allele Registry [92]; 2) Allele-specific epigenomic signatures, such as DNA methylation, histone modifications, and transcription factor binding, from Roadmap Epigenomics [95] and EN-TEx [96] projects; 3) Quantitative trait loci information from GTEx [63] and other studies; and 4) Regulatory element activity, all presented in a tissue- and cell-type-specific manner. The workflow also provides users with commonly used identifiers for variants that fall within a regulatory element of interest, including those from dbSNP [114], ClinVar [115], and the ClinGen Allele Registry [92].