Onassis: Ontology Annotation and Semantic Similarity software

In Onassis: OnASSIs Ontology Annotation and Semantic SImilarity software

library(Onassis)
library(DT)
library(gplots)
library(org.Hs.eg.db)
library(kableExtra)
library(Rtsne)
library(dendextend)
library(clusteval)
library(ggplot2)
library(ggfortify)

Introduction to OnASSis

Public repositories of biological omics data contain thousands of experiments. While these resources are extrenely useful, those data are difficult to mine. The annotation of the associated metadata with controlled vocabularies or ontology terms can facilitate the retrieval of the datasets of interest [@galeota2016ontology]. OnASSiS (Ontology Annotations and Semantic Similarity software) is a package aimed at matching metadata associated with biological experiments with concepts from ontologies, allowing the construction of semantically structured omics datasets, possibly representing various data types from independent studies. The recognition of entities specific for a domain allows the retrieval of samples related to a given cell type or experimental condition, but also allows unravelling previously unanticipated relationships between experiments. Onassis applies Natural Language Processing tools to annotate sample's and experiments' descriptions, recognizing concepts from a multitude of biomedical ontologies and quantifying the similarities between pairs or groups of query studies. Moreover, it assists the semantically-driven analysis of the corresponding omics data. In particular the software includes modules to enable:

• the retrieval of samples’ metadata from repositories of large scale biologial data

• the annotation of these data with concepts belonging to Open Biomedical Ontologies (OBO)

• the organization of the annotated samples in structured groups based on semantic similarity measures

• the comparison of omics data (e.g. gene expression or ChIP enrichment) based on the entities associated to a set of samples and their relationship

Onassis relies on Conceptmapper, an Apache UIMA (Unstructured Information Management Architecture) dictionary lookup tool to retrieve dictionary terms in a given text. https://uima.apache.org/downloads/sandbox/ConceptMapperAnnotatorUserGuide/ConceptMapperAnnotatorUserGuide.html
In particular, the ccp-nlp Conceptmapper wrapper, specific for the biomedical domain, implements a pipeline through which it is possible to retrieve concepts from OBO ontologies in any given text with different adjustable options [@Verspoor:2009b].

Onassis features can be easily accessed through a main class named Onassis, having as slots 'dictionary', 'entities', 'similarity' and 'scores'. In the following sections we first show details on the usage of classes and methods constituting the building blocks of a semantically-driven integrative analysis workflow. Next, in Section 6 we show how the Onassis class wraps all these functions for a simplified access and usage. Regarding the input data, Onassis can handle any type of text, but is particularly well suited for the analysis of the metadata from Gene Expression Omnibus (GEO). Indeed, it allows associating concepts from any OBO ontology to GEO metadata retrieved using r BiocStyle::Biocpkg("GEOmetadb"). In general, any table or database (such as Sequence Read Archive (SRA) [@Zhu2013] or Cistrome [@Shengling2017]) containing textual descriptions that can be easily imported in R as a data frame can be used as input for Onassis. Regarding the dictionary module, gene/protein symbols or epigenetic modifications can also be recognized in the text, in addition to ontology concepts. This can be particularly important, especially when dealing with experiments directed to specific factors or marks (such as ChIP-seq experiments). The similarity module uses different semantic similarity measures to determine the semantic similarity of concepts in a given ontology. This module has been developed on the basis of the Java slib http://www.semantic-measures-library.org/sml. The score module applies statistical tests to determine if omics data from samples annotated with different concepts, belonging to one or more ontologies, are significantly different.

Installation

To run Onassis Java (>= 1.8) is needed. To install the package please run the following code

if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("Onassis")
````

Onassis can be loaded with the following code
```r
library(Onassis)

Some of the optional functions, which will be described in the following parts of the vignette, require additional libraries. These include:

• org.Hs.eg.db (Bioconductor)
• GenomicRanges (Bioconductor)
• gplots (CRAN)
• GEOmetadb (Bioconductor)
• SRAdb (Bioconductor)
• Rtsne
• dendextend
• clusteval
• ggplot2
• ggfortify

Retrieving metadata from public repositories

One of the most straightforward ways to retrieve metadata of samples provided in GEO is through the r BiocStyle::Biocpkg("GEOmetadb") package. In order to use GEOmetadb through Onassis, the corresponding SQLite database should be available. This can be downloaded by Onassis (see below), and this step has to be performed only once. As described below, Onassis provides functions to facilitate the retrieval of specific GEO metadata without the need of explicitly making SQL queries to the database. Additionally, an example on how to access r BiocStyle::Biocpkg("SRAdb") metadata, is also provided.

Handling GEO (Gene Expression Omnibus) metadata

First, it is necessary to obtain a connection to the GEOmetadb SQLite database. If this were already downloaded, connectToGEODB returns a connection to the database given the full path to the SQLite database file. Alternatively, by setting download to TRUE the database is downloaded. The getGEOmetadata function can be used to retrieve the metadata related to specific GEO samples, taking as minimal parameters the connection to the database and one of the experiment types available. Optionally it is possible to specify the organism and the platform. The following code illustrates how to download the metadata corresponding to expression arrays, or DNA methylation sequencing experiments. The meth_metadata object, containing the results for the latter, was stored within Onassis. Therefore, the queries illustrated here can be skipped.

require('GEOmetadb')

## Running this function might take some time if the database (6.8GB) has to be downloaded.
geo_con <- connectToGEODB(download=TRUE)

#Showing the experiment types available in GEO
experiments <- experiment_types(geo_con)

#Showing the organism types available in GEO
species <- organism_types(geo_con)

#Retrieving Human gene expression metadata, knowing the GEO platform identifier, e.g. the Affymetrix Human Genome U133 Plus 2.0 Array
expression <- getGEOMetadata(geo_con, experiment_type='Expression profiling by array', gpl='GPL570')

#Retrieving the metadata associated to experiment type "Methylation profiling by high througput sequencing"
meth_metadata <- getGEOMetadata(geo_con, experiment_type='Methylation profiling by high throughput sequencing', organism = 'Homo sapiens')

Some of the experiment types available are the following:

experiments <- readRDS(system.file('extdata', 'vignette_data', 'experiment_types.rds', package='Onassis'))

knitr::kable(as.data.frame(experiments[1:10]), col.names = c('experiments')) %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "300px", height = "200px")

Some of the organisms available are the following:

species <- readRDS(system.file('extdata', 'vignette_data', 'organisms.rds', package='Onassis'))
knitr::kable(as.data.frame(species[1:10]), col.names=c('species')) %>%
              kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "300px", height = "200px")

As specified above, meth_metadata was previously saved and can be loaded from the Onassis package external data (hover on the table to view additional rows and columns):

meth_metadata <- readRDS(system.file('extdata', 'vignette_data', 'GEOmethylation.rds', package='Onassis'))

methylation_tmp <- meth_metadata
methylation_tmp$experiment_summary <- sapply(methylation_tmp$experiment_summary, function(x) substr(x, 1, 50))
knitr::kable(methylation_tmp[1:10,], 
              caption = 'GEOmetadb metadata for Methylation profiling by high throughput sequencing (only the first 10 entries are shown).') %>%
              kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "300px")

Handling SRA (Sequence Read Archive) metadata

In this section we provide an example showing how it is possible to retrieve metadata from other sources such as SRA. This database is not directly supported by Onassis, since it is not available for Windows platforms. Hence, the code reported below is slightly more complicated, and exemplifies how to query the SRA database provided by the SRAdb package and store metadata of human ChIP-seq experiments within a data frame. Due to the size of the SRA database (2.4 GB for the compressed file, 36 GB for the .sqlite file), a sample of the results of the query is available within Onassis as external data (see below), and the example code illustrated here can be skipped.

# Optional download of SRAdb and connection to the corresponding sqlite file
require(SRAdb)
sqliteFileName <- '/pathto/SRAdb.sqlite'
sra_con <- dbConnect(SQLite(), sqliteFileName)

# Query for the ChIP-Seq experiments contained in GEO for human samples 
library_strategy <- 'ChIP-Seq' #ChIP-Seq data
library_source='GENOMIC' 
taxon_id=9606 #Human samples
center_name='GEO' #Data from GEO

# Query to the sample table 
samples_query <- paste0("select sample_accession, description, sample_attribute, sample_url_link from sample where taxon_id='", taxon_id, "' and sample_accession IS NOT NULL", " and center_name='", center_name, "'"  )

samples_df <- dbGetQuery(sra_con, samples_query)
samples <- unique(as.character(as.vector(samples_df[, 1])))

experiment_query <- paste0("select experiment_accession, center_name, title, sample_accession, sample_name, experiment_alias, library_strategy, library_layout, experiment_url_link, experiment_attribute from experiment where library_strategy='", library_strategy, "'" , " and library_source ='", library_source,"' ", " and center_name='", center_name, "'" )
experiment_df <- dbGetQuery(sra_con, experiment_query)

#Merging the columns from the sample and the experiment table
experiment_df <- merge(experiment_df, samples_df, by = "sample_accession")

# Replacing the '_' character with white spaces
experiment_df$sample_name <- sapply(experiment_df$sample_name, function(value) {gsub("_", " ", value)})
experiment_df$experiment_alias <- sapply(experiment_df$experiment_alias, function(value) {gsub("_", " ", value)})

sra_chip_seq <- experiment_df

The query returns a table with thousands of samples. Alternatively, as described above, a sample of this table useful for the subsequent examples can be retrieved in Onassis:

sra_chip_seq <- readRDS(system.file('extdata', 'vignette_data', 'GEO_human_chip.rds',  package='Onassis'))

knitr::kable(head(sra_chip_seq, 10), rownames=FALSE, 
  caption = 'Metadata of ChIP-seq human samples obtained from SRAdb (first 10 entries)') %>%
              kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "300px")

Annotating text with ontology concepts

The Onassis EntityFinder class has methods for annotating any text with dictionary terms. More specifically, Onassis can take advantage of the OBO dictionaries (http://www.obofoundry.org/).

Data preparation

The findEntities method supports input text in the form of:

• The path of a directory containing named documents.
• The path of a single file containing multiple documents. In this case each row contains the name/identifier of the document followed by a '|' separator and the text to annotate.

Alternatively, the annotateDF method supports input text in the form of a data frame. In this case each row represents a document; the first column has to be the document identifier; the remaining columns will be combined and contain the text to analyze. This option can be conveniently used with the metadata retrieved from r BiocStyle::Biocpkg("GEOmetadb") and r BiocStyle::Biocpkg("SRAdb"), possibly selecting a subset of the available columns.

Creation of a Conceptmapper Dictionary

Onassis handles the convertion of OBO dictionaries into a format suitable to Conceptmapper: XML files with a set of entries specified by the xml tag <token> with a canonical name (the name of the entry) and one or more variants (synonyms).

The constructor CMdictionary creates an instance of the class CMdictionary.

• If an XML file containing the Conceptmapper dictionary is already available, it can be uploaded into Onassis indicating its path and setting the dictType option to "CMDICT".
• If the dictionary has to be built from an OBO ontology (OBO or OWL formats are supported), the path or URL to the corresponding file has to be provided and dictType has to be set to "OBO". The synonymType argument can be set to EXACT_ONLY or ALL to consider only canonical concept names or also to include any synonym. The resulting XML file is written in the indicated outputdir.
• Additionally, to facilitate the named entity recognition of specific targets, such in the case of ChIP-seq experiments, these can be included within a specific dictionary, and dictType has to be set to ENTREZ. If a specific Org.xx.eg.db Bioconductor library is installed and loaded, it can be indicated in the inputFileOrDb parameter as a character string, and gene names will be derived from it. Instead, if inputFileOrDb is empty and a specific species is indicated in the taxID parameter, gene names will be derived from the corresponding gene_info.gz file downloaded from NCBI (300MB). Finally, if dictType is set to TARGET, known histone post-translational modifications and epigenetic marks are also included, in addition to gene names.

# If a Conceptmapper dictionary is already available the dictType CMDICT can be specified and the corresponding file loaded
sample_dict <- CMdictionary(inputFileOrDb=system.file('extdata', 'cmDict-sample.cs.xml', package = 'Onassis'), dictType = 'CMDICT')

#Creation of a dictionary from the file sample.cs.obo available in OnassisJavaLibs
obo <- system.file('extdata', 'sample.cs.obo', package='OnassisJavaLibs')

sample_dict <- CMdictionary(inputFileOrDb=obo, outputDir=getwd(), synonymType='ALL')

# Creation of a dictionary for human genes/proteins. This requires org.Hs.eg.db to be installed
require(org.Hs.eg.db)
targets <- CMdictionary(dictType='TARGET', inputFileOrDb = 'org.Hs.eg.db', synonymType='EXACT')

The following XML markup code illustrates a sample of the Conceptmapper dictionary corresponding to the Brenda tissue ontology.

   <?xml version="1.0" encoding="UTF-8" ?>
   <synonym>
      <token id="http://purl.obolibrary.org/obo/BTO_0005205" canonical="cerebral artery">
        <variant base="cerebral artery"/>
      </token>
      <token id="http://purl.obolibrary.org/obo/BTO_0002179" canonical="184A1N4 cell">
        <variant base="184A1N4 cell"/>
        <variant base="A1N4 cell"/>
      </token>
      <token id="http://purl.obolibrary.org/obo/BTO_0003871" canonical="uterine endometrial cancer cell">
        <variant base="uterine endometrial cancer cell"/>
        <variant base="endometrial cancer cell"/>
        <variant base="uterine endometrial carcinoma cell"/>
        <variant base="endometrial carcinoma cell"/>
      </token>
  </synonym>

Setting the options for the annotator

Conceptmapper includes 7 different options controlling the annotation step. These are documented in detail in the documentation of the CMoptions function. They can be listed through the listCMOptions function. The CMoptions constructor instantiates an object of class CMoptions with the different parameters that will be required for the subsequent step of annotation. We also provided getter and setter methods for each of the 7 parameters.

#Creating a CMoptions object and showing hte default parameters 
opts <- CMoptions()  
show(opts)

To obtain the list of all the possible combinations:

combinations <- listCMOptions()

To create a CMoptions object having has SynonymType 'EXACT_ONLY', that considers only exact synonyms, rather than 'ALL' other types included in OBO (RELATED, NARROW, BROAD)

myopts <- CMoptions(SynonymType='EXACT_ONLY')
myopts

To change a given parameter, for example to use a search strategy based on the Longest match of not-necessarily contiguous tokens where overlapping matches are allowed:

#Changing the SearchStrategy parameter
SearchStrategy(myopts) <- 'SKIP_ANY_MATCH_ALLOW_OVERLAP'
myopts

Running the entity finder

The class EntityFinder defines a type system and runs the Conceptmapper pipeline. It can search for concepts of any OBO ontology in a given text. The findEntities and annotateDF methods accept text within files or data.frame, respectively, as described in Section 4.1. The function EntityFinder automatically adapts to the provided input type, creates an instance of the EntityFinder class to initialize the type system and runs Conceptmapper with the provided options and dictionary. For example, to annotate the metadata derived from ChIP-seq experiments obtained from SRA with tissue and cell type concepts belonging to the sample ontology available in Onassis and containing tissues and cell names, the following code can be used:

sra_chip_seq <- readRDS(system.file('extdata', 'vignette_data', 'GEO_human_chip.rds',  package='Onassis'))
chipseq_dict_annot <- EntityFinder(sra_chip_seq[1:50, c('experiment_accession', 'title', 'experiment_attribute', 'sample_attribute', 'description')], dictionary=sample_dict, options=myopts)

The resulting data.frame contains, for each row, a match to the provided dictionary for the document/sample indicated in the first column. The annotation is reported with the id of the concept (term_id), its canonical name (term name), its URL in the obo format, and the matching sentence of the document.

knitr::kable(head(chipseq_dict_annot, 20), rownames=FALSE, 
  caption = 'Annotating the metadata of DNA methylation sequencing experiments with a dictionary including CL (Cell line) and UBERON ontologies') %>% kable_styling() %>%
  scroll_box(width = "80%", height = "400px")

The function filterTerms can be used to remove all the occurrences of unwanted terms, for example very generic terms.

chipseq_dict_annot <- filterTerms(chipseq_dict_annot, c('cell', 'tissue'))

knitr::kable(head(chipseq_dict_annot, 20), rownames=FALSE, 
  caption = 'Filtered Annotations') %>% kable_styling() %>%
  scroll_box(width = "80%", height = "400px")

\r \r The function EntityFinder can also be used to identify the targeted entity of each ChIP-seq experiment, by retrieving gene names and epigenetic marks in the ChIP-seq metadata.

#Finding the TARGET entities
target_entities <- EntityFinder(input=sra_chip_seq[1:50, c('experiment_accession', 'title', 'experiment_attribute', 'sample_attribute', 'description')], options = myopts, dictionary=targets)

knitr::kable(head(target_entities, 20), 
  caption = 'Annotations of ChIP-seq test metadata obtained from SRAdb and stored into files with the TARGETs (genes and histone variants)') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

Semantic similarity

Once a set of samples is annotated, i.e. associated to a set of ontology concepts, Onassis allows the quantification of the similarity among these samples based on the semantic similarity between the corresponding concepts. Similarity is an Onassis class applying methods of the Java library slib \url{http://www.semantic-measures-library.org/sml/} [@harispe2014semantic], which builds a semantic graph starting from OBO ontology concepts and their hierarchical relationships. The following methods are available and are automatically chosen depending on the settings of the Similarity function. The sim and groupsim methods allow the computation of semantic similarity between single terms (pairwise measures) and between group of terms (groupwise measures), respectively. Pairwise measures can be edge based, if they rely only on the structure of the ontology, or information-content based if they also consider the information that each term in the ontology carries. Rather, groupwise measures can be indirect, if they compute the pairwise similarity between each couple of terms, or direct if they consider each set of concepts as a whole. The samplesim method allows to determine the semantic similarity between two documents, each possibly associated to multiple concepts. Finally, the multisim method allows to determine the semantic similarity between documents annotated with two or more ontologies: first samplesim is run for each ontology, then a user defined function can be used to aggregate the resulting semantic similarities for each pair of documents.

The function listSimilarities shows all the measures supported by Onassis. For details about the measures run {?Similarity}.

#Instantiating the Similarity
similarities <- listSimilarities()

Semantic similarity between ontology terms

The following example shows pairwise similarities between the individual concepts of previously annotated ChIP-seq experiments metadata. The lin similarity measure is used by default, which relies on a ratio between the Information content (IC) of the terms most specific common ancestor, and the sum of their IC (based on the information content of their most informative common ancestor). In particular, the seco information content is used by default, which determines the specificity of each concept based on the number of concepts it subsumes.

#Retrieving URLS of concepts
found_terms <- as.character(unique(chipseq_dict_annot$term_url))

# Creating a dataframe with all possible couples of terms and adding a column to store the similarity
pairwise_results <- t(combn(found_terms, 2))
pairwise_results <- cbind(pairwise_results, rep(0, nrow(pairwise_results)))

# Similarity computation for each couple of terms
for(i in 1:nrow(pairwise_results)){
  pairwise_results[i, 3] <- Similarity(obo, pairwise_results[i,1], pairwise_results[i, 2])
}
colnames(pairwise_results) <- c('term1', 'term2', 'value')  

# Adding the term names from the annotation table to the comparison results 
pairwise_results <- merge(pairwise_results, chipseq_dict_annot[, c('term_url', 'term_name')], by.x='term2', by.y='term_url')
colnames(pairwise_results)[length(colnames(pairwise_results))] <- 'term2_name'
pairwise_results <- merge(pairwise_results, chipseq_dict_annot[, c('term_url', 'term_name')], by.x='term1', by.y='term_url')
colnames(pairwise_results)[length(colnames(pairwise_results))] <- 'term1_name'
pairwise_results <- unique(pairwise_results)
# Reordering the columns
pairwise_results <- pairwise_results[, c('term1', 'term1_name', 'term2', 'term2_name', "value")]

knitr::kable(pairwise_results,  
  caption = 'Pairwise similarities of cell type terms annotating the ChIP-seq metadata') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

Noteworthy, the terms 'B-cell' and 'lymphocyte' are closer (similarity 0.83) than 'B cell' and 'epithelial cell' (similarity 0.26). It is also possible to compute the semantic similarity between two groups of terms. For example, to determine a value of similarity for the combination of ('non-terminally differentiated cell', 'epithelial cell') and the combination of ('lymphocyte' , 'B cell') we can use the ui measure (set as default measure in Onassis), a groupwise direct measure combining the intersection and the union of the set of ancestors of the two groups of concepts.

oboprefix <- 'http://purl.obolibrary.org/obo/'
Similarity(obo, paste0(oboprefix, c('CL_0000055', 'CL_0000066')) , paste0(oboprefix, c('CL_0000542', 'CL_0000236')))

The similarity between these two groups of terms is low (in the interval [0, 1]), while the addition of the term 'lymphocyte of B lineage' to the first group the group similarity increases.

Similarity(obo, paste0(oboprefix, c('CL_0000055', 'CL_0000236' ,'CL_0000236')), paste0(oboprefix, c('CL_0000542', 'CL_0000066')))

Semantic similarity between annotated samples

Similarity also supports the computation of the similarity between annotated samples. Since each sample is typically associated tu multiple terms, Similarity runs in the groupwise mode when applied to samples. To this end, samples identifiers and a data frame with previously annotated concepts returned by EntityFinder are required.

# Extracting all the couples of samples 
annotated_samples <- as.character(as.vector(unique(chipseq_dict_annot$sample_id)))

samples_couples <- t(combn(annotated_samples, 2))

# Computing the samples semantic similarity 
similarity_results <- apply(samples_couples, 1, function(couple_of_samples){
  Similarity(obo, couple_of_samples[1], couple_of_samples[2], chipseq_dict_annot)
})


#Creating a matrix to store the results of the similarity between samples
similarity_matrix <- matrix(0, nrow=length(annotated_samples), ncol=length(annotated_samples))
rownames(similarity_matrix) <- colnames(similarity_matrix) <- annotated_samples

# Filling the matrix with similarity values 
similarity_matrix[lower.tri(similarity_matrix, diag=FALSE)] <- similarity_results
similarity_matrix <- t(similarity_matrix)
similarity_matrix[lower.tri(similarity_matrix, diag=FALSE)] <- similarity_results 
# Setting the diagonal to 1
diag(similarity_matrix) <- 1

# Pasting the annotations to the sample identifiers
samples_legend <- aggregate(term_name ~ sample_id, chipseq_dict_annot, function(aggregation) paste(unique(aggregation), collapse=',' ))
rownames(similarity_matrix) <- paste0(rownames(similarity_matrix), ' (', samples_legend[match(rownames(similarity_matrix), samples_legend$sample_id), c('term_name')], ')')

# Showing a heatmap of the similarity between samples
heatmap.2(similarity_matrix, density.info = "none", trace="none", main='Samples\n semantic\n similarity', margins=c(12,50), cexRow = 2, cexCol = 2, keysize = 0.5)

Onassis class

The class Onassis was built to wrap the main functionalities of the package in a single class. It consists of 4 slots:

• dictionary: stores the source dictionary used to find entities.

• entities: a table containing the annotations of documents (samples). The list of unique concepts belonging to the dictionary and found in the metadata representing a given sample is defined as a that sample semantic set

• similarity: a matrix of the similarities between the unique semantic sets identified in the entities table

• scores: a dataset of quantitative measurements (e.g. gene expression) associated to the same samples annotated in the entities slot.

Metadata annotation

In this section we illustrate the use of the Onassis class to annotate the metadata of ChIP-seq samples having as target H3K27ac. The dataset used for the following examples was obtained by annotating the all the ChIP-seq samples from SRA with target entities (as described in Section 4.2) and selecting the sample identifiers having H3K27ac as target. To load it from the vignette data:

h3k27ac_chip <- readRDS(system.file('extdata', 'vignette_data', 'h3k27ac_metadata.rds',  package='Onassis'))

The method annotate takes as input a data frame of metadata to annotate, the type of dictionary and the path of an ontology file and returns an instance of class Onassis. The input data frame should have unique identifiers in the first column (sample identifiers or generic document identifiers) and for each row one or more columns containing the metadata related to the identifier. Importantly, to subsequently compute the semantic similarities, the dictionary given to the method needs to be an 'OBO' ontology. To annotate tissue and cell types we used the previously loaded dictionary available in OnassisJavaLibs containing a sample of the Cell Line ontology CL merged with UBERON terms to identify also tissues.

cell_annotations <- Onassis::annotate(h3k27ac_chip, 'OBO', obo, FindAllMatches='YES' )

To slot entities of an Onassis object reports for each sample the unique list of concepts found in the corresponding metadata. Specifically, term ids, term urls and term names are reported, and multiple entries per sample are comma separated. We usually refer to these unique lists as semantic sets.

To retrieve the semantic sets in an object of class Onassis we provided the accessor method entities

cell_entities <- entities(cell_annotations) 

knitr::kable(cell_entities[sample(nrow(cell_entities), 10),], 
  caption = ' Semantic sets of ontology concepts (entities) associated to each sample, stored in the entities slot of the Onassis object') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

The filterconcepts method can be used to filter out unwanted annotations, for example terms that we consider redundant or too generic. The method modifies the entities slot of the Onassis object and returns a new Onassis object with filtered semantic sets.

filtered_cells <- filterconcepts(cell_annotations, c('cell', 'tissue'))

knitr::kable(entities(filtered_cells), 
  caption = 'Entities in filtered Onassis object') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

Semantic similarity of semantic sets

The method sim populates the similarity slot within an Onassis object. Specifically, it generates a matrix containing semantic similarities between the semantic sets for each pair of samples annotated in the entities slot.

filtered_cells <- sim(filtered_cells)

The matrix of similarities can be accessed using the method simil(filtered_cells).

Semantic sets with semantic similarities above a given threshold can be combined using the method collapse. This method, based on hierarchical clustering, unifies the similar semantic sets by concatenating their unique concepts. Term names and term urls in the entities slot will be updated accordingly. For each concept, the number of samples associated is also reported in bracket squares, while the total number of samples associated to a given semantic set is indicated in parentheses. After the collapse, the similarity matrix in the similarity slot is consequently updated with the similarities of the new semantic sets.

collapsed_cells <- Onassis::collapse(filtered_cells, 0.9)

knitr::kable(entities(collapsed_cells), 
  caption = ' Collapsed Entities in filtered Onassis object') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

In the following heatmap the similarity values of collapsed cell/tissue semantic sets is reported.

heatmap.2(simil(collapsed_cells), density.info = "none", trace="none", margins=c(36, 36), cexRow = 1.5, cexCol = 1.5, keysize=0.5)

Integrating the annotations from different ontologies

In typical integrative analyses scenarios, users could be interested in annotating concepts from different domains of interest. In this case, one possibility could be building a tailored application ontology including the concepts from different ontologies and their relationships. However this is complicated, since it requires to match and integrate the relationships from one ontology to the other. Rather, Onassis allows integrating two ontologies by repeating the annotation process with another ontology, while keeping separate the semantic sets from the two ontologies. In the following example ChIP-seq samples will be annotated with information about disease conditions. For this particular semantic type, Onassis provides also a boolean variable disease that can be set to TRUE to recognize samples metadata explicitly annotated as Healthy conditions, to be differentiated from metadata where disease terms are simply lacking.

obo2 <- system.file('extdata', 'sample.do.obo', package='OnassisJavaLibs')
disease_annotations <- Onassis::annotate(h3k27ac_chip, 'OBO',obo2, disease=TRUE )

knitr::kable(entities(disease_annotations), 
  caption = ' Disease entities') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

The method mergeonassis can be used to combine two Onassis objects in which the same set of samples was annotated with two different ontologies. This is useful to perform a nested analysis driven by the annotation provided by the two ontologies. The first object associates the samples metadata to semantic sets from a primary domain of interest. For each semantic set, the associated samples will be further separated based on the semantic sets belonging to a secondary domain. For example, users could be interested in comparing different diseases (secondary domain) within each cell type (primary domain), for a set of cell type.

cell_disease_onassis <- mergeonassis(collapsed_cells, disease_annotations)

The following table shows merged entities:

knitr::kable(entities(cell_disease_onassis), 
  caption = ' Cell and disease entities') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

Semantically-driven analysis of omics data

The method compare exploits the Onassis object to analyze the actual omics data. To this end, a score matrix containing measurements for any arbitrary genomic unit (rows) within each sample (columns) is necessary. For example, genomics units can be genes, for expression analyses, or genomic regions (such as promoters) for enrichment analyses (such as ChIP-seq experiments). We suggest to take advantage of data repositories where omics data were reanalyzed with standardazied analysis pipelines. This would minimize confunding issues related to the differences due to alternative analysis and normalization procedures. For example, CistromeDB provides ChIP-seq data from GEO re-analyzed with a standardized pipeline [@Shengling2017]. To illustrate the compare method, we obtained from Cistrome the genomic positions of H3K27ac peaks for the previously annotated ChIP-seq samples. Cistrome samples can be retrieved using GSMs (GEO) sample identifiers, which were matched with those reported in the SRA ChIP-seq metadata table. We precomputed a score matrix having as rows human promoter regions in chromosome 20, and as columns the sample identifiers. Each entry of the matrix contains the peak score value reported by Cistrome if there was a peak for a given promoter region in a given sample, or 0 if there was no peak overlapping that promoter region. Promoters for the human genome were obtained using the Bioconductor annotation package TxDb.Hsapiens.UCSC.hg19.knownGene considering 2000 bases upstream and 2000 bases downstream transcription start sites of genes. The score table can be retrieved from the vignette_data:

score_matrix <- readRDS(system.file('extdata', 'vignette_data', 'score_matrix.rds', package='Onassis'))

The method compare can be used with any test function to compare scores across different semantic sets. When by = 'col' samples will be compared, meaning that the distribution of all the scores of the samples associated to a given semantic state will be compared with those of any other semantic state. This process will be then repeated for any pair of semantic states. Hence, test statistic and p-values are reported. In this example, these measure global differences in the distribution of H3K27ac at promoters among variuos cell types. The method returns a matrix whose each elements includes test statistic and p-value.

cell_comparisons_by_col <- compare(collapsed_cells, score_matrix=as.matrix(score_matrix), by='col', fun_name='wilcox.test')

matrix_of_p_values <- matrix(NA, nrow=nrow(cell_comparisons_by_col), ncol=ncol(cell_comparisons_by_col))
for(i in 1:nrow(cell_comparisons_by_col)){
  for(j in 1:nrow(cell_comparisons_by_col)){
   result_list <- cell_comparisons_by_col[i,j][[1]]
   matrix_of_p_values[i, j] <- result_list[2]
  }
}
colnames(matrix_of_p_values) <- rownames(matrix_of_p_values) <- colnames(cell_comparisons_by_col)

In the following heatmap -log10 p-values of the test are shown:

heatmap.2(-log10(matrix_of_p_values), density.info = "none", trace="none", main='Changes in\n H3K27ac signal \nin promoter regions', margins=c(36,36), cexRow = 1.5, cexCol = 1.5, keysize=1)

By setting by='row', we can use compare to test for differences for each genomic unit in the different tissue/cell type semantic sets. In this example, for each promoter, the distribution of the scores within samples associated to a given semantic state, will be compared with those of any other semantic state. This process will be then repeated for any pair of semantic states. This allows measuring differences in the distribution of H3K27ac among variuos cell types within each specific promoter. Indeed, compare returns a matrix whose elements are lists with genomic regions including test statistics and p-values. After the application of the wilcoxon test we extracted for each couple of semantic sets the number of regions having a p.value <= 0.1.

cell_comparisons <- compare(collapsed_cells, score_matrix=as.matrix(score_matrix), by='row', fun_name='wilcox.test', fun_args=list(exact=FALSE))

# Extraction of p-values less than 0.1
significant_features <- matrix(0, nrow=nrow(cell_comparisons), ncol=ncol(cell_comparisons))
colnames(significant_features) <- rownames(significant_features) <- rownames(cell_comparisons)
for(i in 1:nrow(cell_comparisons)){
  for(j in 1:nrow(cell_comparisons)){
   result_list <- cell_comparisons[i,j][[1]]
   significant_features[i, j] <- length(which(result_list[,2]<=0.1))
  }
}

In the following table we report, for each pair of semantic states, the number of promoter regions with different H3K27ac binding patterns

knitr::kable(significant_features, 
  caption = ' Number of promoter regions with p.value <=0.1') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

If annotations from two ontologies are provided, for example cell types and diseases, compare returns a nested analysis: the method iterates for each cell type, and performs all pair-wise comparisons (of the scores) between the diseases associated to a given cell type. Therefore, a named list with the semantic sets of the first level ontology (cell type) is returned. Each entry of the list contains a matrix with the results of the test for each pair of second level semantic sets (diseases). For example, to use a wilcoxon test to compare H3K27ac between different diseases within each tissue/cell type, promoter by promoter, the following code can be used:

disease_comparisons <- compare(cell_disease_onassis, score_matrix=as.matrix(score_matrix), by='row', fun_name='wilcox.test', fun_args=list(exact=FALSE))

To visualize the diseases associated with the tissue semantic set "breast [8], mammary gland epithelial cell [2], gland [1] (11)" the following code can be used:

rownames(disease_comparisons$`breast [8], mammary gland epithelial cell [2], gland [1] (11)`)

To access to the test results for the the comparison of "Healthy"" with "breast cancer, cancer"

selprom <- (disease_comparisons$`breast [8], mammary gland epithelial cell [2], gland [1] (11)`[2,1][[1]])
selprom <- selprom[is.finite(selprom[,2]),]
head(selprom)

To determine the number of promoter regions having p.value <= 0.1 within each comparison

disease_matrix <- disease_comparisons[[1]]

# Extracting significant p-values
significant_features <- matrix(0, nrow=nrow(disease_matrix), ncol=ncol(disease_matrix))
colnames(significant_features) <- rownames(significant_features) <- rownames(disease_matrix)
for(i in 1:nrow(disease_matrix)){
  for(j in 1:nrow(disease_matrix)){
   result_list <- disease_matrix[i,j][[1]]
   significant_features[i, j] <- length(which(result_list[,2]<=0.1))
  }
}

knitr::kable(significant_features, 
  caption = ' Number of promoter regions with p.value <= 0.1 ') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

Noteworthy, 5 of the 19 promoters that were differential for H3K27ac in breast cancer, refer to genes found to be important in that disease (according to GeneCards). In particular, these include CDC25B, which was recently identifed as therapeutic target for triple-negative breast cancer [@LIU2018112].

Alternatively, when not available within R libraries, personalized test functions can be applied. The code below for example implements a function named signal to noise statistic that could be applied as an alternative to the wilcoxon test.

personal_t <- function(x, y){
    if(is.matrix(x))
      x <- apply(x, 1, mean)
    if(is.matrix(y))
      y <- apply(y, 1, mean)
    signal_to_noise_statistic <- abs(mean(x) - mean(y)) / (sd(x) + sd(y))
    return(list(statistic=signal_to_noise_statistic, p.value=NA))
}

disease_comparisons <- compare(cell_disease_onassis, score_matrix=as.matrix(score_matrix), by='col', fun_name='personal_t')

Further details and examples about compare are available in the help page of the method.

Correspondence between semantically-driven and data-driven analyses

We reasoned that the aggregation of H3K27ac samples in groups defined by semantic similarities should reflect, at least in part, the similarity of the actual ChIP-seq data. To this end, we identified semantic groups at different cutoff values of semantic similarity (from 0.1 to 1). We used t-SNE dimensionality reduction method to identify an equal number of data-driven groups, based on promoter H3K27ac densities. The overlap between semantic and data-driven groups can be quantified with the Jaccard index. The overlap analysis was then repeated using a randomly reshuffled score matrix.

onassis_annotations <- entities(filtered_cells)
# Sorting the annotations based on the order of the score matrix labels
onassis_annotations <- onassis_annotations[match(colnames(score_matrix), onassis_annotations$sample_id),]

# Creating a score matrix with scrambled annotations 
random_score_matrix <- score_matrix[,sample(1:ncol(score_matrix))]

# Running t-SNE on the score matrix and on the random matrix
set.seed(456)
tsne_out <- Rtsne(t(score_matrix), perplexity = 10, dims=3)$Y 
random_tsne_out = Rtsne(t(random_score_matrix), perplexity = 10, dims=3)$Y 
tsne_hclus <- hclust(dist(tsne_out[,c(1,2)]),  method='ward.D')
random_tsne_hclus <- hclust(dist(random_tsne_out[,c(1,2)]) , method='ward.D')


# Vetors to store the similarities of data driven and semantic driven clusters
real_jaccard_similarity_vector <- c()
random_jaccard_similarity_vector <- c()


sequence_of_similarities <- rev(seq(0.1, 1, 0.1))
sequence_of_similarities[1] <- 0.99

for(i in sequence_of_similarities){
  collapsed_cells <- Onassis::collapse(filtered_cells,  as.numeric(i)) 
  onassis_annotations <- entities(collapsed_cells)
  onassis_annotations <- onassis_annotations[match(colnames(score_matrix), onassis_annotations$sample_id),]

  # Clustering based on the tSNE results
  tsne_clus <- factor(cutree(tsne_hclus, length(unique(onassis_annotations$short_label))))

  # Similarity of the tSNE driven vs semantic set driven clusters
  real_jaccard_similarity_vector <- c(real_jaccard_similarity_vector, cluster_similarity(tsne_clus, as.numeric(factor(onassis_annotations$short_label)), similarity='jaccard'))  

  # Clustering of the random matrix
  random_tsne_clus <- factor(cutree(random_tsne_hclus, length(unique(onassis_annotations$short_label))))

  # Similarity of the tSNE driven vs semantic set driven clusters
  random_jaccard_similarity_vector <- c(random_jaccard_similarity_vector, cluster_similarity(random_tsne_clus, as.numeric(factor(onassis_annotations$short_label)), similarity='jaccard')) 
}
names(real_jaccard_similarity_vector) <- rev(seq(0.1, 1, 0.1))

We plot the Jaccard similarities for real and random datasets

# Plotting the Jaccard similarities for real and random datasets 
myx <- rev(seq(0.1, 1, 0.1))
myylim <- max(random_jaccard_similarity_vector, real_jaccard_similarity_vector) + 0.05
mydata = cbind(myx, random_jaccard_similarity_vector, real_jaccard_similarity_vector)

plot1 <- ggplot(mydata, aes(x=myx)) + 
  geom_line(aes(y = random_jaccard_similarity_vector, color = "darkred")) + 
  geom_line(aes(y = real_jaccard_similarity_vector, color="steelblue" )) +
  geom_point(aes(y = random_jaccard_similarity_vector, color = "darkred")) + 
  geom_point(aes(y = real_jaccard_similarity_vector, color = "steelblue"))  +
  scale_x_continuous(breaks=seq(0,1, 0.1)) + 
  scale_color_manual(labels=c('Random dataset', 'Real dataset'), values=c('darkred', 'steelblue'), name='') + 
  theme_bw() +
  xlab('Semantic similarity threshold') + 
  ylab('Jaccard Similarity')

plot1

From the previous image we could observe lower Jaccard similarities for the randomized dataset. We also identified the semantic similarity cutoff at 0.7 as the one maximizing the similarity between data driven and semantically driven clusters.

best_similarity <- names(real_jaccard_similarity_vector)[which(real_jaccard_similarity_vector==max(real_jaccard_similarity_vector))]

In the following plot we show the tSNE positioning of the samples with samples colored by semantic set. The proximity of samples associated with the same semantic group idicates that they have similar patterns of H3K27ac promoters.

# Collapsing semantic sets at the best similarity value
collapsed_cells <- Onassis::collapse(filtered_cells, as.numeric(best_similarity))
onassis_annotations <- entities(collapsed_cells)

# Sorting the annotations based on the score matrix columns
onassis_annotations <- onassis_annotations[match(colnames(score_matrix), onassis_annotations$sample_id),]

tsne_clus <- cutree(hclust(dist(tsne_out[,c(1,2)]), method='ward.D'), length(unique(onassis_annotations$short_label)))

my_tsne_out <- data.frame(tsne_out)
colnames(my_tsne_out) <- c('tSNE1', 'tSNE2', 'tSNE3')
my_tsne_out$annotations <- onassis_annotations$short_label
my_tsne_out$cluster <- as.numeric(as.vector(tsne_clus))

my_random_tsne_clus <- factor(cutree(hclust(dist(random_tsne_out[,c(1,2)]), method='ward.D'), length(unique(onassis_annotations$short_label))))

my_colors <- c("#7FC97F", "#386CB0", "#F0027F", "#666666", "#FF7F00", "#6A3D9A", "#F4CAE4")

plot2 <- ggplot(my_tsne_out[, c(1, 2, 4, 5)], aes(x=tSNE1, y=tSNE2,colour= annotations)) +
  geom_point(size=5) + 
  theme_bw()  +
  scale_colour_manual(name = "Semantic set", values = my_colors) 

plot2 

In the following plot instead we show how reshuffling the data matrix the association between semantic groups and data driven groups is lost

my_random_tsne_out <- data.frame(random_tsne_out)
colnames(my_random_tsne_out) <- c('tSNE1', 'tSNE2', 'tSNE3')
my_random_tsne_out$annotations <- onassis_annotations$short_label
my_random_tsne_out$cluster <- as.numeric(as.vector(random_tsne_clus))

plot3 <- ggplot(my_random_tsne_out[, c(1, 2, 4, 5)], aes(x=tSNE1, y=tSNE2,colour= annotations)) +
  geom_point(size=5) +
  theme_bw()  +
  scale_colour_manual(name = "Semantic set", values = my_colors) 

plot3 

We also show the correspondence between tSNE clusters and semantic sets

plot4 <- ggplot(my_tsne_out[, c(1, 2, 4, 5)], aes(x=tSNE1, y=tSNE2,colour= annotations)) +
  stat_ellipse(type = "t", level=0.9, data=my_tsne_out[, c(1, 2, 4,5)], aes(x=tSNE1, y=tSNE2, colour=factor(cluster))) +
  geom_point(size=5) +
  theme_bw()  +
  scale_colour_manual(name = "Semantic set",values = c(rep('black', 7), my_colors)) +

  theme(legend.position = "none")
plot4

Performances of the tool

In this section we provide a table with details on the performance of Onassis in running the annotation, semantic similarity, and data integrative analysis tasks. Different input sizes and dictionary sizes were tested and run time and memory footprint are reported. Tests were run on a desktop computer having a 3,5 GHz Intel Core i7 processor and 24 GB RAM. The table reports the name of the function, the elapsed time in seconds as reported by the R system.time function, the size of the object in bytes as reported by object_size function, a description and the used code. We additionaly report the performances of the queries to SRA used to obtain the ChIP-seq dataset used in the vignette.

performances_table <- readRDS(system.file('extdata', 'vignette_data', 'performances.rds', package='Onassis'))

knitr::kable(performances_table, 
  caption = ' Onassis performances ') %>% kable_styling(bootstrap_options = c("striped"), position="center") %>%
  scroll_box(width = "80%", height = "400px")

Session Info

Here is the output of sessionInfo() on the system on which this document was compiled through kintr:

sessionInfo()

References

Onassis documentation built on Nov. 8, 2020, 8:18 p.m.