ssrch: selectize-based search engine for corpora of modest size
In ssrch: a simple search engine

suppressPackageStartupMessages({
library(ssrch)
library(DT)
})

Introduction

Metadata are crucial to all forms of statistical analysis. Metadata are used to define formal steps of upstream data preprocessing, to annotate outcomes and covariates, and to interpret inferential results.

The ssrch package was developed to provide a lightweight approach to searching a metadata corpus with R. There are three basic steps:

tokenize corpus elements syntactically,
filter tokens into hashed environments,
use selectize.js to achieve a predictive, autocompleted search interface.

We illustrate the system using a collection of tables of metadata about cancer transcriptomics.

Illustration

To provide a sense of what is at stake, we work with 68 CSV files derived from the NCBI Sequence Read Archive metadata.
The files consist of contents of the sample.attribute subtable of study metadata retrieved using the SRAdbV2 package at github.com:seandavi/SRAdbV2. The CSV files are lodged in a zip file canc68.zip with the ssrch package in inst/cancer_corpus.

The CSV files have been indexed in an S4 object of class DocSet.

data(docset_cancer68)
docset_cancer68

A single document can be retrieved with the retrieve_doc function, given the study accession number.

retrieve_doc("ERP010142", docset_cancer68)[1:3,1:5]

Diversity of field names

There is partial standardization of field names in this corpus, but there is considerable variation among studies in the number and types of fields used.

docids = ls(envir=docs2kw(docset_cancer68))
allc = lapply(docids, 
   function(x) try(retrieve_doc(x, docset_cancer68)))
summary(sapply(allc,ncol))
lapply(allc[c(11,55)], names)

The ssrch package provides tools for identifying studies in which a given field is used, and in which given values are recorded.

A motivating example is determining what studies involve measurements on normal tissue adjacent to tumor. In our collection, this can be assessed via:

library(ssrch)
searchDocs("Adjacent", docset_cancer68, ignore.case=TRUE)

Managing tokenized metadata

Before getting into the details of search engine construction, we review some high-level concepts and methods.

First, we have the container for managing our search resource.

getClass(class(docset_cancer68))

docset_cancer68 is the result of using the ssrch function parseDoc in conjunction with the CSV files zipped in ssrch/inst/cancer_corpus. It is an S4 class instance that manages environments mapping from keywords to documents, documents to records, documents to keywords,

Furthermore, DocSet instances can have a component that retrieves parsed documents from the corpus.

Querying the corpus

We'll search for the phrase Non Small Cell with an optional hyphen, ignoring character case of possible hits, including as well the abbreviation for Non-small cell lung cancer.

NSChits = searchDocs("Non.Small Cell|NSCLC$", docset_cancer68, ignore.case=TRUE)
NSChits

Three studies, three different patterns matched by the query string. We can use retrieve_doc to look at the contents of the metadata tables for these studies.

NSCdocs = lapply(unique(NSChits$docs), 
   function(x) retrieve_doc(x, docset_cancer68))
names(NSCdocs) = NSChits$docs
datatable(NSCdocs[[1]], options=list(lengthMenu=c(3,5,10)))
datatable(NSCdocs[[2]], options=list(lengthMenu=c(3,5,10)))
datatable(NSCdocs[[3]], options=list(lengthMenu=c(3,5,10)))

A prototypical app

The ctxsearch function starts a shiny app that provides access to full attribute data via a selectize input that uses all tokens in the corpus (filtered).

ctxsearch illustration