Facilities for Filtering Bioconductor Annotation Resources
In AnnotationFilter: Facilities for Filtering Bioconductor Annotation Resources

BiocStyle::markdown()

Package: r Biocpkg("AnnotationFilter")
Authors: r packageDescription("AnnotationFilter")[["Author"]]
Last modified: r file.info("AnnotationFilter.Rmd")$mtime
Compiled: r date()

Introduction

A large variety of annotation resources are available in Bioconductor. Accessing the full content of these databases or even of single tables is computationally expensive and in many instances not required, as users may want to extract only sub-sets of the data e.g. genomic coordinates of a single gene. In that respect, filtering annotation resources before data extraction has a major impact on performance and increases the usability of such genome-scale databases.

The r Biocpkg("AnnotationFilter") package was thus developed to provide basic filter classes to enable a common filtering framework for Bioconductor annotation resources. r Biocpkg("AnnotationFilter") defines filter classes for some of the most commonly used features in annotation databases, such as symbol or genename. Each filter class is supposed to work on a single database table column and to facilitate filtering on the provided values. Such filter classes enable the user to build complex queries to retrieve specific annotations without needing to know column or table names or the layout of the underlying databases. While initially being developed to be used in the r Biocpkg("Organism.dplyr") and r Biocpkg("ensembldb") packages, the filter classes and the related filtering concept can be easily added to other annotation packages too.

Filter classes

All filter classes extend the basic AnnotationFilter class and take one or more values and a condition to allow filtering on a single database table column. Based on the type of the input value, filter classes are divided into:

CharacterFilter: takes a character value of length >= 1 and supports conditions ==, !=, startsWith and endsWith. An example would be a GeneIdFilter that allows to filter on gene IDs.
IntegerFilter: takes a single integer as input and supports the conditions ==, !=, >, <, >= and <=. An example would be a GeneStartFilter that filters results on the (chromosomal) start coordinates of genes.
DoubleFilter: takes a single numeric as input and supports the conditions ==, !=, >, <, >= and <=.
GRangesFilter: is a special filter, as it takes a GRanges as value and performs the filtering on a combination of columns (i.e. start and end coordinate as well as sequence name and strand). To be consistent with the findOverlaps method from the r Biocpkg("IRanges") package, the constructor of the GRangesFilter filter takes a type argument to define its condition. Supported values are "any" (the default) that retrieves all entries overlapping the GRanges, "start" and "end" matching all features with the same start and end coordinate respectively, "within" that matches all features that are within the range defined by the GRanges and "equal" that returns features that are equal to the GRanges.

The names of the filter classes are intuitive, the first part corresponding to the database column name with each character following a _ being capitalized, followed by the key word Filter. The name of a filter for a database table column gene_id is thus called GeneIdFilter. The default database column for a filter is stored in its field slot (accessible via the field method).

The supportedFilters method can be used to get an overview of all available filter objects defined in AnnotationFilter.

library(AnnotationFilter)
supportedFilters()

Note that the AnnotationFilter package does provides only the filter classes but not the functionality to apply the filtering. Such functionality is annotation resource and database layout dependent and needs thus to be implemented in the packages providing access to annotation resources.

Usage

Filters are created via their dedicated constructor functions, such as the GeneIdFilter function for the GeneIdFilter class. Because of this simple and cheap creation, filter classes are thought to be read-only and thus don't provide setter methods to change their slot values. In addition to the constructor functions, AnnotationFilter provides the functionality to translate query expressions into filter classes (see further below for an example).

Below we create a SymbolFilter that could be used to filter an annotation resource to retrieve all entries associated with the specified symbol value(s).

library(AnnotationFilter)

smbl <- SymbolFilter("BCL2")
smbl

Such a filter is supposed to be used to retrieve all entries associated to features with a value in a database table column called symbol matching the filter's value "BCL2".

Using the "startsWith" condition we could define a filter to retrieve all entries for genes with a gene name/symbol starting with the specified value (e.g. "BCL2" and "BCL2L11" for the example below.

smbl <- SymbolFilter("BCL2", condition = "startsWith")
smbl

In addition to the constructor functions, AnnotationFilter provides a functionality to create filter instances in a more natural and intuitive way by translating filter expressions (written as a formula, i.e. starting with a ~).

smbl <- AnnotationFilter(~ symbol == "BCL2")
smbl

Individual AnnotationFilter objects can be combined in an AnnotationFilterList. This class extends list and provides an additional logicOp() that defines how its individual filters are supposed to be combined. The length of logicOp() has to be 1 less than the number of filter objects. Each element in logicOp() defines how two consecutive filters should be combined. Below we create a AnnotationFilterList containing two filter objects to be combined with a logical AND.

flt <- AnnotationFilter(~ symbol == "BCL2" &
                            tx_biotype == "protein_coding")
flt

Note that the AnnotationFilter function does not (yet) support translation of nested expressions, such as (symbol == "BCL2L11" & tx_biotype == "nonsense_mediated_decay") | (symbol == "BCL2" & tx_biotype == "protein_coding"). Such queries can however be build by nesting AnnotationFilterList classes.

## Define the filter query for the first pair of filters.
afl1 <- AnnotationFilterList(SymbolFilter("BCL2L11"),
                             TxBiotypeFilter("nonsense_mediated_decay"))
## Define the second filter pair in ( brackets should be combined.
afl2 <- AnnotationFilterList(SymbolFilter("BCL2"),
                             TxBiotypeFilter("protein_coding"))
## Now combine both with a logical OR
afl <- AnnotationFilterList(afl1, afl2, logicOp = "|")

afl

This AnnotationFilterList would now select all entries for all transcripts of the gene BCL2L11 with the biotype nonsense_mediated_decay or for all protein coding transcripts of the gene BCL2.

Using `AnnotationFilter` in other packages

The AnnotationFilter package does only provide filter classes, but no filtering functionality. This has to be implemented in the package using the filters. In this section we first show in a very simple example how AnnotationFilter classes could be used to filter a data.frame and subsequently explore how a simple filter framework could be implemented for a SQL based annotation resources.

Let's first define a simple data.frame containing the data we want to filter. Note that subsetting this data.frame using AnnotationFilter is obviously not the best solution, but it should help to understand the basic concept.

## Define a simple gene table
gene <- data.frame(gene_id = 1:10,
                   symbol = c(letters[1:9], "b"),
                   seq_name = paste0("chr", c(1, 4, 4, 8, 1, 2, 5, 3, "X", 4)),
                   stringsAsFactors = FALSE)
gene

Next we generate a SymbolFilter and inspect what information we can extract from it.

smbl <- SymbolFilter("b")

We can access the filter condition using the condition method

condition(smbl)

The value of the filter using the value method

value(smbl)

And finally the field (i.e. column in the data table) using the field method.

field(smbl)

With this information we can define a simple function that takes the data table and the filter as input and returns a logical with length equal to the number of rows of the table, TRUE for rows matching the filter.

doMatch <- function(x, filter) {
    do.call(condition(filter), list(x[, field(filter)], value(filter)))
}

## Apply this function
doMatch(gene, smbl)

Note that this simple function does not support multiple filters and also not conditions "startsWith" or "endsWith". Next we define a second function that extracts the relevant data from the data resource.

doExtract <- function(x, filter) {
    x[doMatch(x, filter), ]
}

## Apply it on the data
doExtract(gene, smbl)

We could even modify the doMatch function to enable filter expressions.

doMatch <- function(x, filter) {
    if (is(filter, "formula"))
        filter <- AnnotationFilter(filter)
    do.call(condition(filter), list(x[, field(filter)], value(filter)))
}

doExtract(gene, ~ gene_id == '2')

For such simple examples AnnotationFilter might be an overkill as the same could be achieved (much simpler) using standard R operations. A real case scenario in which AnnotationFilter becomes useful are SQL-based annotation resources. We will thus explore next how SQL resources could be filtered using AnnotationFilter.

We use the SQLite database from the r Biocpkg("org.Hs.eg.db") package that provides a variety of annotations for all human genes. Using the packages' connection to the database we inspect first what database tables are available and then select one for our simple filtering example.

We use an EnsDb SQLite database used by the r Biocpkg("ensembldb") package and implement simple filter functions to extract specific data from one of its database tables. We thus load below the EnsDb.Hsapiens.v75 package that provides access to human gene, transcript, exon and protein annotations. Using its connection to the database we inspect first what database tables are available and then what fields (i.e. columns) the gene table has.

## Load the required packages
library(org.Hs.eg.db)
library(RSQLite)
## Get the database connection
dbcon <- org.Hs.eg_dbconn()

## What tables do we have?
dbListTables(dbcon)

org.Hs.eg.db provides many different tables, one for each identifier or annotation resource. We will use the gene_info table and determine which fields (i.e. columns) the table provides.

## What fields are there in the gene_info table?
dbListFields(dbcon, "gene_info")

The gene_info table provides the official gene symbol and the gene name. The column symbol matches the default field value of the SymbolFilter as does the column gene_name for the GeneNameFilter. If the column in the database would not match the field of an AnnotationFilter, we would have to implement a function that maps the default field of the filter object to the database column. See the end of the section for an example.

We next implement a simple doExtractGene function that retrieves data from the gene_info table and re-uses the doFilter function to extract specific data. The parameter x is now the database connection object.

doExtractGene <- function(x, filter) {
    gene <- dbGetQuery(x, "select * from gene_info")
    doExtract(gene, filter)
}

## Extract all entries for BCL2
bcl2 <- doExtractGene(dbcon, SymbolFilter("BCL2"))

bcl2

This works, but is not really efficient, since the function first fetches the full database table and subsets it only afterwards. A much more efficient solution is to translate the AnnotationFilter class(es) to an SQL where condition and hence perform the filtering on the database level. Here we have to do some small modifications, since not all condition values can be used 1:1 in SQL calls. The condition "==" has for example to be converted into "=" and the "startsWith" into a SQL "like" by adding also a "%" wildcard to the value of the filter. We would also have to deal with filters that have a value of length > 1. A SymbolFilter with a value being c("BCL2", "BCL2L11") would for example have to be converted to a SQL call "symbol in ('BCL2','BCL2L11')". Here we skip these special cases and define a simple function that translates an AnnotationFilter to a where condition to be included into the SQL call. Depending on whether the filter extends CharacterFilter or IntegerFilter the value has also to be quoted.

## Define a simple function that covers some condition conversion
conditionForSQL <- function(x) {
    switch(x,
           "==" = "=",
           x)
}

## Define a function to translate a filter into an SQL where condition.
## Character values have to be quoted.
where <- function(x) {
    if (is(x, "CharacterFilter"))
        value <- paste0("'", value(x), "'")
    else value <- value(x)
    paste0(field(x), conditionForSQL(condition(x)), value)
}

## Now "translate" a filter using this function
where(SeqNameFilter("Y"))

Next we implement a new function which integrates the filter into the SQL call to let the database server take care of the filtering.

## Define a function that 
doExtractGene2 <- function(x, filter) {
    if (is(filter, "formula"))
        filter <- AnnotationFilter(filter)
    query <- paste0("select * from gene_info where ", where(filter))
    dbGetQuery(x, query)
}

bcl2 <- doExtractGene2(dbcon, ~ symbol == "BCL2")
bcl2

Below we compare the performance of both approaches.

system.time(doExtractGene(dbcon, ~ symbol == "BCL2"))

system.time(doExtractGene2(dbcon, ~ symbol == "BCL2"))

Not surprisingly, the second approach is much faster.

Be aware that the examples shown here are only for illustration purposes. In a real world situation additional factors, like combinations of filters, which database tables to join, which columns to be returned etc would have to be considered too.

What if the database column on which we want to filter does not match the field of an AnnotatioFilter? If for example the database column is named hgnc_symbol instead of symbol we could for example package-internally overwrite the default field method for SymbolFilter to return the correct field for the database column.

## Default method from AnnotationFilter:
field(SymbolFilter("a"))

## Overwrite the default method.
setMethod("field", "SymbolFilter", function(object, ...) "hgnc_symbol")

## Call to field returns now the "correct" database column
field(SymbolFilter("a"))