1 Introduction to OnASSis

Public repositories contain thousands of experiments and samples that are difficult to mine. Annotating the description of this data with controlled vocabularies or ontology terms could improve the retrieval of data of interest both programmatically or manually (Galeota and Pelizzola 2016). OnASSiS (Ontology Annotations and Semantic Similarity software) is a package aimed at matching metadata associated with biological experiments with concepts from ontologies, thus aiming at obtaining semantically coherent omics datasets, possibly representing various data types as derived from independent studies. The recognition of domain specific entities not only allows users to retrieve samples related to a given cell type or experimental condition, but also to discover different and not immediately obvious 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/divergences between pairs or groups of query studies. In particular the software includes modules to assist on:

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

• the annotation of these data with concepts belonging to OBO biomedical ontologies

• the organization of available samples in comparable and coherent groups based on semantic similarity measures

• the comparison of biological signal stored in a matrix of scores (e.g. gene expression in different samples) based on the annotated entities associated to each sample

Onassis uses 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 et al. 2009).

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 the classes and methods that constitute the building blocks of typical metadata integration workflows and than we show how the Onassis class wraps all these functions for a simplified access and usage.

Onassis can handle any type of text as input, but is particularly well suited for the analysis of the metadata from Gene Expression Omnibus (GEO).Indeed it provides the possibility to associate concepts from any OBO ontology to GEO metadata retrieved using GEOmetadb. This represents a fundamental first step in the integrative analysis of the data from large scale repositories (Galeota and Pelizzola 2016). In general, any table or database (Sequence Read Archive (SRA) (Zhu et al. 2013) or Cistrome (Mei et al. 2017) ) containing textual descriptions that can be easily imported in R as a data frame can be used as input for Onassis. In addition to the ontological concepts, the recognition of gene/protein symbols or epigenetic modifications can be highly relevant, especially for experiments directed to those specific factors or marks (such as ChIP-seq experiments). The semantic 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 simple statistical tests to determine if biological signals from samples annotated with different concepts are significantly different.

2 Installing additional libraries to run the examples in the vignette

To run Onassis Java (>= 1.8) is needed. Some of the optional functions, which will be described in the followign parts of the vignette, require additional libraries. For the analyses in this vignette please install following libraries:

source("https://bioconductor.org/biocLite.R")
biocLite('org.Hs.eg.db')
biocLite("GenomicRanges")
install.packages('data.table')
install.packages('DT')
install.packages('gplots')

3 Retrieving public repositories metadata

One of the most straightforward ways to retrieve metadata of samples provided in GEO is through GEOmetadb package. In order to use GEOmetadb through Onassis, users should download the corresponding SQLite database file, by following the instructions provided in the package vignette. Onassis provides functions to help GEO metadata retrieval without the need of explicitly making SQL queries to the database. While GEOmetadb can be accessed on any platform, another important database SRAdb, is not available for Windows users. In the following sections we show how to query GEOmetadb through Onassis, but we also provide an example accessing SRAdb metadata.

## Running this function might take long time if the database 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 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')

#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')

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

# Connection to the SRAmetadb and potential download of the sqlite file
sqliteFileName <- './data/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])))

# Query to the experiment table
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,
                           "' " )

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 field separators with white spaces
experiment_df$experiment_attribute <- sapply(experiment_df$experiment_attribute, 
                                             function(value) {
                                               gsub("||", "  ", value)
                                             })
experiment_df$sample_attribute <- sapply(experiment_df$sample_attribute, 
                                         function(value) {
                                           gsub("||", "  ", value)
                                         })
# 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

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

4 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/), as shown in the next section where we will define relationships between different samples annotated with different ontology concepts thanks to the structure of the 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>

# 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
require(org.Hs.eg.db)
targets <- CMdictionary(dictType='TARGET', inputFileOrDb = 'org.Hs.eg.db')

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

## CMoptions object to set ConceptMapper Options

## SearchStrategy: CONTIGUOUS_MATCH
## CaseMatch: CASE_INSENSITIVE
## Stemmer: NONE
## StopWords: NONE
## OrderIndependentLookup: ON
## FindAllMatches: YES
## SynonymType: ALL

combinations <- listCMOptions()

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

## CMoptions object to set ConceptMapper Options

## SearchStrategy: CONTIGUOUS_MATCH
## CaseMatch: CASE_INSENSITIVE
## Stemmer: NONE
## StopWords: NONE
## OrderIndependentLookup: ON
## FindAllMatches: YES
## SynonymType: EXACT_ONLY

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

## CMoptions object to set ConceptMapper Options

## SearchStrategy: SKIP_ANY_MATCH_ALLOW_OVERLAP
## CaseMatch: CASE_INSENSITIVE
## Stemmer: NONE
## StopWords: NONE
## OrderIndependentLookup: ON
## FindAllMatches: YES
## SynonymType: EXACT_ONLY

chipseq_dict_annot <- EntityFinder(sra_chip_seq[1:20,c('sample_accession', 'title', 'experiment_attribute', 'sample_attribute', 'description')], dictionary=sample_dict, options=myopts)

#Finding the TARGET entities
target_entities <- EntityFinder(input=sra_chip_seq[1:20,c('sample_accession', 'title', 'experiment_attribute', 'sample_attribute', 'description')], options = myopts, dictionary=targets) 

5 Semantic similarity

With Onassis it is possible to quantify the semantic similarity between the concepts of a given ontology using different semantic similarity measures. Similarity is an Onassis class applying methods of the Java library slib (Harispe et al. 2014), 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, using groupwise measures. 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()

The following example shows the pairwise similarities between sample cell concepts obtained annotating the ChIP-seq metadata. The lin similarity measure is used by default, which relies on a ratio between the Information content (IC) of the Most Specific Common Ancestor of the compared concepts and the sum of their IC based on the information content of the most informative common ancestor of the considered concepts. 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.

found_terms <- unique(chipseq_dict_annot$term_url)
n <- length(found_terms)

ontologyfile <- obo
pairwise_results <- data.frame(term1 = character(0), term2= character(0), value = double(0L))
for(i in 1:(n-1)){
  term1 <- as.character(found_terms[i])
  j = i + 1 
  for(k in j:n){
    term2 <- as.character(found_terms[k])
    two_term_similarity <- Similarity(ontologyfile,  term1, term2 )
    new_row <- cbind(term1, term2, two_term_similarity)
    pairwise_results <- rbind(pairwise_results, new_row )
  }
}
pairwise_results <- unique(pairwise_results)
pairwise_results <- merge(pairwise_results, chipseq_dict_annot[, c('term_url', 'term_name')], by.x='term2', by.y='term_url', all.x=TRUE)
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', all.x=TRUE)
colnames(pairwise_results)[length(colnames(pairwise_results))] <- 'term1_name'
pairwise_results <- unique(pairwise_results)

In the following code the semantic similarity between two groups of terms is computed using the ui measure, a groupwise direct measure combining the intersection and the union of the set of ancestors of the two groups of concepts.

Similarity(obo, found_terms[1:2], found_terms[3])

## [1] 0.1875

Lastly, the pariwise semantic similarity between ChIP-seq samples is illustrated.

annotated_samples <- as.character(as.vector(unique(chipseq_dict_annot$sample_id)))
n <- length(annotated_samples)


samples_results <- data.frame(sample1 = character(0), sample2= character(0), value = double(0L))
samples_results <- matrix(0, nrow=n, ncol=n)
rownames(samples_results) <- colnames(samples_results) <- annotated_samples
for(i in 1:(n-1)){
  sample1 <- as.character(annotated_samples[i])
  j = i + 1 
  for(k in j:n){
    sample2 <- as.character(annotated_samples[k])
    two_samples_similarity <- Similarity(ontologyfile, sample1, sample2, chipseq_dict_annot)
    samples_results[i, k] <- samples_results[k, i] <- two_samples_similarity
  }
}
diag(samples_results) <- 1
heatmap.2(samples_results, density.info = "none", trace="none", main='Semantic similarity of annotated samples', margins=c(5,5))

6 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 can be defined as a 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 samples annotated in entities and separated in the different semantic sets identified in the annotation process.

In this section we illustrate the use of the Onassis class to annotate the previously retrieved metadata. 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.

onassis_annotations <- annotate(sra_chip_seq, 'OBO',obo )

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

onassis_entities <- entities(onassis_annotations) 

The filterconcepts method can be used to filter out unwanted annotations. It takes the Onassis object and removes from its entities the undesired concepts.

filtered_onassis <- filterconcepts(onassis_annotations, c('cell'))

The method sim cretes a matrix of the semantic similarities between the annotations of each couple of samples annotated in the entities slot of an Onassis object.

filtered_onassis <- sim(filtered_onassis)

Annotations with semantic similarities above a given threshold can be unified using the method collapse. This method unifies the similar annotations by concatenating their unique concepts. Entities are replaced with the new concatenated annotations. For each concept in the concatenated annotations the number of samples associated is also reported, together with the total number of samples annotated with the new annotations. The similarity slot will be consequently updated.

collapsed_onassis <- Onassis::collapse(filtered_onassis, 0.8)
head(entities(collapsed_onassis))

heatmap.2(simil(collapsed_onassis), margins=c(15,15), cexRow = 1, cexCol = 1)

7 Session Info

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

## R version 3.5.1 Patched (2018-07-12 r74967)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 16.04.5 LTS
## 
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.7-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.7-bioc/R/lib/libRlapack.so
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8          LC_NUMERIC=C                 
##  [3] LC_TIME=en_US.UTF-8           LC_COLLATE=C                 
##  [5] LC_MONETARY=en_US.UTF-8       LC_MESSAGES=en_US.UTF-8      
##  [7] LC_PAPER=en_US.UTF-8          LC_NAME=en_US.UTF-8          
##  [9] LC_ADDRESS=en_US.UTF-8        LC_TELEPHONE=en_US.UTF-8     
## [11] LC_MEASUREMENT=en_US.UTF-8    LC_IDENTIFICATION=en_US.UTF-8
## 
## attached base packages:
## [1] parallel  stats4    stats     graphics  grDevices utils     datasets 
## [8] methods   base     
## 
## other attached packages:
##  [1] kableExtra_0.9.0      org.Hs.eg.db_3.6.0    AnnotationDbi_1.42.1 
##  [4] IRanges_2.14.11       S4Vectors_0.18.3      Biobase_2.40.0       
##  [7] BiocGenerics_0.26.0   gplots_3.0.1          DT_0.4               
## [10] Onassis_1.2.7         OnassisJavaLibs_1.2.0 rJava_0.9-10         
## [13] BiocStyle_2.8.2      
## 
## loaded via a namespace (and not attached):
##  [1] Rcpp_0.12.18       tidyr_0.8.1        gtools_3.8.1      
##  [4] assertthat_0.2.0   rprojroot_1.3-2    digest_0.6.17     
##  [7] R6_2.2.2           backports_1.1.2    RSQLite_2.1.1     
## [10] evaluate_0.11      highr_0.7          httr_1.3.1        
## [13] pillar_1.3.0       rlang_0.2.2        rstudioapi_0.7    
## [16] data.table_1.11.4  gdata_2.18.0       blob_1.1.1        
## [19] rmarkdown_1.10     readr_1.1.1        stringr_1.3.1     
## [22] htmlwidgets_1.2    RCurl_1.95-4.11    bit_1.1-14        
## [25] munsell_0.5.0      compiler_3.5.1     xfun_0.3          
## [28] pkgconfig_2.0.2    htmltools_0.3.6    tidyselect_0.2.4  
## [31] tibble_1.4.2       GEOquery_2.48.0    bookdown_0.7      
## [34] viridisLite_0.3.0  crayon_1.3.4       dplyr_0.7.6       
## [37] bitops_1.0-6       DBI_1.0.0          magrittr_1.5      
## [40] scales_1.0.0       KernSmooth_2.23-15 stringi_1.2.4     
## [43] bindrcpp_0.2.2     limma_3.36.3       xml2_1.2.0        
## [46] tools_3.5.1        bit64_0.9-7        glue_1.3.0        
## [49] purrr_0.2.5        hms_0.4.2          yaml_2.2.0        
## [52] colorspace_1.3-2   caTools_1.17.1.1   rvest_0.3.2       
## [55] memoise_1.1.0      GEOmetadb_1.42.0   knitr_1.20        
## [58] bindr_0.1.1