Summary
Autoimmune diseases are heterogeneous pathologies with difficult diagnosis and few therapeutic options. In the last decade, several omics studies have provided significant insights into the molecular mechanisms of these diseases. Nevertheless, data from different cohorts and pathologies are stored independently in public repositories and a unified resource is imperative to assist researchers in this field. Here, we present ADEx (https://adex.genyo.es), a database that integrates 82 curated transcriptomics and methylation studies covering 5609 samples for some of the most common autoimmune diseases. The database provides, in an easy-to-use environment, advanced data analysis and statistical methods for exploring omics datasets, including meta-analysis, differential expression or pathway analysis.
Introduction
Autoimmune diseases (ADs) are a group of complex and heterogeneous disorders characterized by immune responses to self-antigens leading to tissue damage and dysfunction in several organs. The pathogenesis of ADs is not fully understood, but both environmental and genetic factors have been linked to their development (Salaman, 2003). Although these disorders cause damage to different organs and their clinical outcomes vary between them, they share many risk factors and molecular mechanisms (Jörg et al., 2016). Some examples of ADs are systemic lupus erythematosus (SLE), rheumatoid arthritis (RA), Sjögren’s syndrome (SjS), systemic sclerosis (SSc), considered systemic autoimmune diseases (SADs) and type 1 diabetes (T1D), which is considered an organ-specific autoimmune disease. Most of these diseases are classified as rare given their prevalence, but altogether ADs affect up to 3 % of the population considering conservative estimates (Cooper and Stroehla, 2003).
In ADs patients, the pathology is developed during several years but it is only detected when tissue damage is significant. For that reason, early diagnosis is important and complicated. Additionally, some ADs often show a non-linear outcome that alternates between active and remission stages thus making their study even more difficult. Despite huge efforts have been made to develop ADs biomarkers and therapies, these do not fit for every patient and their clinical responses differ greatly (Barturen et al., 2018).
During the past decade, the use of omics technologies has provided new insights into the molecular mechanisms associated with the development of ADs, opening new scenarios for biomarkers and treatments discovery (Kim et al., 2014). In this context, it is remarkable the characterization of the type I interferon (IFN) gene expression signature as a key factor in the pathology of some SADs, especially in SLE and SjS (Thorlacius et al., 2018), which has improved our knowledge of the underlying molecular mechanisms and has opened new therapeutic strategies based on blocking the pathways related to this signature.
Regardless of the large amount of omics studies describing new biomarkers and therapeutic strategies in ADs (Arriens and Mohan, 2013; Ferreira et al., 2014; Teruel et al., 2017; Xie et al., 2018), in most cases these biomarkers are not consistent across different studies or have not fully accomplished their diagnostic goals. Indeed, the widely studied IFN signature is highly variable between patients (Rönnblom and Eloranta, 2013) and it is associated to differences in response to treatments which target it, as has been reported for example in the phase-II results of Sifalimumab clinical trial for SLE patients (Khamashta et al., 2016). In addition, in most of the cases, biomarkers are defined from the analysis of a single type of omic data (commonly gene expression), but multi-omics data integration can provide a more complete understanding of molecular mechanisms and more robust and biologically relevant biomarkers.
Most of the omics datasets generated from different cohorts and studies in ADs published to date have been deposited and are available in public repositories such as Gene Expression Omnibus (GEO) (Edgar et al., 2002) or ArrayExpress (Kolesnikov et al., 2015). Although all these valuable data can be used in retrospective analyses in order to generate new knowledge and accelerate drug discovery and diagnosis, it is not easy to compare neither to integrate available data because they are generated from different platforms and/or processed with different analytic pipelines. In this context, there are great efforts from the bioinformatics community to develop standardized data analysis workflows and resources that facilitate data integration and reproducible analysis. For example, Lachmann et al. (Lachmann et al., 2018) have recently reprocessed a large collection of raw human and mouse RNA-Seq data from GEO and Sequence Read Archive (SRA) using a unified pipeline and they have developed the ARCHS4 as a resource to provide direct access to these data through a web-based user interface. Other singular projects such as The Cancer Genome Atlas (TCGA) (Weinstein et al., 2013) or the Genotype-Tissue Expression project (GTEx) (Lonsdale et al., 2013) provide also large and homogeneously processed datasets for tumor samples and human tissues respectively. These unprecedented resources motivate the development of applications and data portals to help researchers gather information with the aim of improving diagnosis and treatment in multiple diseases, most notably in cancer research, where such information is actually being used in the clinical practice (Jang et al., 2018).
Despite such enormous potential, in the context of ADs there is a lack of a centralized and dedicated resource that facilitates the exploration, comparison and integration of available omics datasets. This is indeed an area in which this type of application would be tremendously beneficial, given that the low prevalence of each individual disease makes difficult the recruitment of large patients cohorts (Barturen et al., 2018).
To bridge this gap, in this work we have compiled and curated most of the publicly available gene expression and methylation datasets for five ADs: SLE, RA, SjS, SSc and T1D. To this end, we have developed and applied homogeneous pipelines from raw data and we developed ADEx (Autoimmune Disease Explorer), a data portal where these processed data can be downloaded and exploited through multiple exploratory and statistical analyses. ADEx facilitates data integration and analysis to potentially improve diagnosis and treatment of ADs.
Results
Data collection and processing
ADEx contains data from 5609 samples. We have processed 82 expression and methylation datasets from case-control studies for SLE, RA, SjS, SSc and T1D diseases (see Table 1 for a summary and Supplementary Table 1 for complete information about all included datasets). We have manually curated all metadata in order to standardize the nomenclature of phenotypes, cell types, etc. from different studies and discard samples or datasets that do not meet the selection criteria (see Methods section). In addition, we have prepared five different pipelines to process data for each platform (RNA-Seq, Affymetrix and Illumina gene expression microarrays, and Illumina methylation arrays 27K and 450K).
All these workflows are written in R language and are publicly available in GENyO Bioinformatics Unit GitHub (https://github.com/GENyO-BioInformatics/public/tree/master/ADEx). The processed datasets are available from the Download Data section in the application. Figure 1 contains an overview of the different steps performed to prepare the data for ADEx application.
The ADEx application
ADEx data portal can be used to download and analyze the processed data. ADEx is freely available at https://adex.genyo.es. The tool is divided in 6 different sections arranged in different tabs (Figure 2a).
Section 1: Data overview
Information about the available datasets can be found in both table or pie plot formats in this section. In tables, information about the sample phenotype and their data origin is provided. In pie plots quantitative information is provided regarding the clinical and phenotype information. All this information has been extracted from GEO or from the associated published articles whenever supplied. This information can be presented individually for each dataset or grouped by disease. While a single dataset is being explored, the experiment summary is shown. Users can use this section to identify datasets of their interest to be analyzed in the following sections.
Section 2: Gene Query
This section was created in order to explore the expression and methylation of a specific gene, or the correlation between them, within a single dataset. Users can explore the different gene expression values for each dataset comparing case and control samples with a boxplot. Meanwhile, methylation data is presented at CpG level, so that users can select a region of the gene (e.g. promoter) and the mean methylation value for cases and controls is plotted for every CpG probe contained in the selected region. It has been demonstrated the strong relationship of gene expression and methylation levels (Suzuki and Bird, 2008). That is why, in this section, users can also integrate both expression and methylation values to search for direct or inverse correlations. Finally, gene expression correlation analysis can be performed in order to get insight into the relationship between different genes and to find groups of coexpressed genes.
Section 3: Gene Set Query
Here users can select several datasets and genes in order to explore the Fold-Change (FC) between patients and controls across studies. All datasets from a disease can be automatically selected by clicking the right buttons, or individual studies can be selected by clicking directly on the table. Users can introduce a list of genes to explore their expression, although there are several preloaded gene lists covering the coexpression modules reported by Chaussabel et al. (Chaussabel et al., 2008). These modules consist of sets of coexpressed genes among hundreds of samples from different diseases. Each transcriptional module is associated to different pathways and cell types, most of them related to the immune system (Chaussabel et al., 2008). See Figures 2b and 2c for an example of this type of analysis.
Section 4: Analyze Dataset
In this section, we focus the analysis on whole datasets instead of individual genes. By default, a heatmap with the expression of the top 50 differentially expressed genes (DEGs) sorted by FDR is displayed. It is also possible to sort them by FC and FDR cutoffs can be applied to both statistics. Additionally, differential expression analysis results can be downloaded as an excel table.
Furthermore, users can also study the KEGG (Kanehisa and Goto, 2000) enriched pathways associated to the dataset selected. These results are precomputed using all the DEGs that have an FDR value below 0.05. A table gathers the significantly enriched KEGG pathways along with their associated hypergeometric test statistics and an interactive plot shows detailed information of the participant genes in the pathway colored according to their FC.
Beyond conventional pathway enrichment methods, we have implemented more sophisticated mechanistic models of cell signaling activity which have demonstrated to be very sensitive in deciphering disease mechanisms (Cubuk et al., 2018; Hidalgo et al., 2017) as well as the mechanisms of action of drugs (Amadoz et al., 2015; Esteban-Medina et al., 2019).
To offer this functionality we have applied HiPathia software (Hidalgo et al., 2017) to gene expression data. This method estimates changes in the activity of signaling circuits defined into different pathways. With this approach, it becomes possible to study in detail the specific signaling circuits altered in ADs within the different signaling pathways. We precomputed this analysis for each dataset and the results are available as tables and interactive reports.
Section 5: Meta-Analysis
ADEx also implements meta-analysis functionalities based on gene expression data to integrate and jointly analyze different and heterogeneous datasets. We implemented a meta-analysis approach to search for biomarkers and common gene signatures across different datasets from the same or different pathologies (Toro-Domínguez et al., 2014a) based on the FCs of each dataset and gene. Datasets have to be selected similarly to Section 3 to launch the meta-analysis.
Section 6: Download data
In this section, users can select one or several datasets and download them. Curated data is obtained with the aim of performing additional analyses externally to ADEx application.
Methods
Data collection
Collection of the datasets included in ADEx was carried out by searching in GEO web page with ADs names as key terms. We filtered the results by study type (expression profiling by array, expression profiling by high throughput sequencing and methylation profiling by array), organism (Homo sapiens) and platform manufacturer (Affymetrix or Illumina).
We downloaded the metadata for these initial datasets with GEOquery (Davis and Meltzer, 2007) R package in order to apply our inclusion criteria and exclude those studies and samples that do not meet them. We only included case-control studies from samples, which were not treated with drugs in vitro. Exclusively datasets with available raw data were considered. Studies whose controls and cases belong to different tissues were discarded. We only selected datasets with 10 samples at least. We divided the datasets containing samples from different diseases, platforms, tissues or cell types in subgroups so that these are constant and avoid possible batch effects.
82 datasets containing 5609 samples passed our filtering criteria (see Table 1 for a summary and Supplementary Table 1 for complete information about all included datasets). Then we downloaded their raw data with GEOquery (Davis and Meltzer, 2007). For expression microarrays, we downloaded CEL files and raw text files for Affymetrix and Illumina platforms respectively. For RNA-Seq, we downloaded the fastq files from the European Nucleotide Archive. For methylation microarrays, we downloaded raw methylation tables if they were available and idat files otherwise.
Metadata curation
GEO does not require submitters to use either a fixed structure or standard vocabulary to describe the samples of an experiment. For that reason, it was necessary to manually homogenize the information provided within all the selected datasets using standardized terms. There are some methods for automatic curation of GEO metadata, but manual curation is still necessary to get high-quality metadata (Wang et al., 2019). This metadata curation was an essential step for the following analyses and permits an easy datasets information exploration.
Platforms curation
We have used a total of 12 different gene expression platforms from microarray and RNA-Seq technologies. Microarray platforms quantify expression levels in probes. In order to match probe identifiers to gene names, platforms annotation files are available from GEO. However, we found that some of these annotation files match probes to inappropriate gene names. On the one hand, some platforms save gene names with errors due to the conversion of gene names such as MARCH1 or SEPT1 into dates, a common error that has been reported previously (Ziemann et al., 2016). In these cases, we fixed manually these genes in the annotation files. On the other hand, different platforms use obsolete or simply different synonyms or aliases to refer to the same genes. We used human genes’ information from NCBI repository in order to match aliases with actual official gene symbols and substituted them in the platform annotations.
Data processing
Raw data from Illumina expression microarrays were loaded by reading the plain text files. In order to remove background noise, we kept only the probes that had a Detection P-value lower than 0.05 in 10 % of the samples. Then we performed a background correction and quantile normalization (Shi et al., 2010) using neqc function from limma package (Ritchie et al., 2015).
CEL files from Affymetrix expression microarrays platforms were loaded to R environment with affy package (Gautier et al., 2004). To filter low intensity probes, we removed all probes with an intensity lower than 100 in at least 10 % of the samples. Normalization was carried out computing Robust Multichip Average (RMA) normalization (Irizarry et al., 2003) with affy package (Gautier et al., 2004).
For RNA-Seq datasets, fastq files were aligned to human transcriptome reference hg38 using STAR 2.4 (Dobin et al., 2013) and raw counts were obtained with RSEM v1.2.31 (Li and Dewey, 2011) with default parameters. Raw counts were filtered using NOISeq R package (Tarazona et al., 2015), removing those features that have an average expression per condition lower than 0,5 counts per million (CPM) and a coefficient of variation (CV) higher than 100 in all conditions. Counts normalization was carried out with TMM method (Robinson and Oshlack, 2010).
We translated microarrays probes identifiers to gene symbols using our curated annotation tables. For those genes targeted by two or more microarray probes, we calculated the median expression values of all their targeting probes. For RNA-Seq, we translated ENSEMBL identifiers to gene symbols using biomaRt package (Durinck et al., 2005, 2009).
Methylation raw data are available in GEO as idat or text files depending on the dataset. Idat files were read with minfi package (Aryee et al., 2014), while text files were read in the R environment. In both cases, poorly performing probes with a detection P-value above 0.05 in more than 10 % of samples were removed. Probes adjacent to SNPs, located in sexual chromosomes or reported to be cross-reactive (Chen et al., 2013) were also removed. We normalized the methylation signals using quantile normalization with lumi package (Du et al., 2008). Finally, for datasets generated with 450k platform, we applied BMIQ normalization (Teschendorff et al., 2013) using wateRmelon package (Pidsley et al., 2013) in order to correct for the two types of probes contained in this platform.
Differential expression analysis
We performed a differential expression analysis in all datasets independently towards the identification of differential patterns among disease samples and healthy controls. These analyses were performed in different ways depending on the source of data. Gene expression profiles from microarray platforms were carried out by the standard pipeline of limma package (Ritchie et al., 2015). We used lmFit function to fit a linear model to the gene expression values followed by the execution of a t-test by the empirical Bayes method for differential activity (eBayes function). On the other hand, gene expression profiles from RNA-Seq platforms were analyzed by the standard pipeline of DESeq2 package (Love et al., 2014). In both cases, differential expression analysis provided P-values, adjusted P-values by FDR and log2 FC.
Pathway analysis
Pathway enrichment analysis was precomputed for each expression dataset using differential expression analysis results. We considered DEGs those genes with a FDR lower than 0.05 and we performed hypergeometric tests to check if each pathway contains more DEGs as expected by chance. We used KEGGprofile 1.24.0 R package to perform this analysis but beforehand we manually updated its dependency, KEGG.db, the database used to perform the statistical test. The pathways were plotted using KEGG mapper tool, Search&Color Pathway, with the genes colored by their FC between case and control samples.
Signaling network analysis
We integrated signaling network analysis applying HiPathia software (Hidalgo et al., 2017) to gene expression data so that changes in the activity of the network from different pathways can be detected. We precomputed this analysis for each gene expression dataset. Firstly, we translated the gene expression matrix and scaled it. Then, we calculated the transduction signal and compared among conditions, cases and controls. Finally, the results were stored in interactive html reports.
Database architecture
Pursuing an optimal data organization and a quick access to all the data in ADEx, we have enabled an internal database with PostgreSQL. We chose this technology since it is open source and it is best suited to the huge dimensionality of omics datasets.
Webtool
ADEx user interface was designed with RStudio Shiny package. The application uses a set of external packages to perform analysis and graphics on demand. Most of the plots are generated with ggplot2 (Wickham, 2009). All the computations in Meta-Analysis section are performed whenever users request them. Biomarkers analysis is performed with Rank Products algorithm integrated in RankProd R package (Del Carratore et al., 2017). The tool runs in our own server with CentOS 7.0 operating system, 16 processors and 32 Gb of RAM memory.
Discussion
Despite that the heterogeneity of ADs is evident, there are common molecular mechanisms involved in the activation of immune responses. In this context, integrative analyses of multiple studies are crucial to discover shared and differential molecular signatures (Toro-Domínguez et al., 2014b). Nowadays there are many ADs datasets publicly available, but a strong computational knowledge is necessary in order to analyze them properly. With the aim of filling this gap between experimental research and computational biology, interactive easy-to-use software are valuable tools to perform exploratory and statistical analysis without strong computational expertise. This type of tool has been developed for other diseases and has helped to reuse public data and generate new knowledge and hypotheses (Cerami et al., 2012; Díez-Villanueva et al., 2015; Toro-Domínguez et al., 2019).
A resource of this type is urged in the field of ADs to: 1) Compile available ADs’ public data in a single data portal, 2) Access to integrable data processed with uniform pipelines, and 3) Perform both individual and integrated analysis interactively. We developed ADEx database to accomplish all those objectives.
As far as we know, ADEx is the first ADs omics database and we expect it to be a reference in this area. During the coming years, ADEx will be expanded including data from more ADs and other omics.
Supplementary Information
Acknowledgements
We would like to thank all the authors of the datasets included in ADEx. We also would like to thank Alberto Ramírez for his technical support during the implementation of ADEx in our server. This work is part of the JMM’s PhD thesis. JMM is enrolled in the PhD program in Biomedicine at the University of Granada, Spain.
JMM is partially funded by Ministerio de Economía, Industria y Competitividad. This work is partially funded by Consejería de Salud, Junta de Andalucía (Grant PI-0173-2017).