vjcitn / metametrics

measuring content and utility of sample-level annotation in genomic experiments

title	author	date	vignette	output
Metametrics: an R package with metadata metrics for annotation of genomic compendia	Vincent J. Carey, stvjc at channing.harvard.edu	`r format(Sys.time(), '%B %d, %Y')`	%\VignetteEngine{knitr::rmarkdown} %\VignetteIndexEntry{Semantic metrics for cancer corpus} %\VignetteEncoding{UTF-8}	BiocStyle::html_document highlight number_sections theme toc pygments true united true

suppressPackageStartupMessages({
library(ggplot2)
library(plotly)
library(metametrics)
library(ssrch)
})

Basic observations on a corpus of human RNA-seq studies in cancer

Using the Omicidx system, we harvested metadata about human samples for which RNA-seq data was deposited in NCBI SRA.

We work with a subset of 1009 studies for which a cancer-related term was present in study title as recorded at NCBI SRA.

library(ggplot2)
library(plotly)
library(metametrics)
data(study_publ_dates) # harvesting omicidx early 2019
library(lubridate)
ds_ca = DocSet_ca1009()
ds_ca

We accumulate (over dates of study submissions) the set of fields used in the sample annotation of the 1009 cancer studies.

study_publ_dates = na.omit(study_publ_dates)
studs1009 = ls(docs2kw(ds_ca))  # in cancer corpus
stud_dates = as_datetime(study_publ_dates[,2])
names(stud_dates) = study_publ_dates[,1]
stud_dates = stud_dates[studs1009]  # limit to corpus
stud_dates = sort(stud_dates)
ofields = lapply(names(stud_dates), 
    function(x) names(retrieve_doc(x, ds_ca)))
freqs = table(unlist(ofields))
#sort(freqs,decreasing=TRUE)[1:20]
cumfields = ofields
for (i in 2:length(cumfields)) cumfields[[i]] = 
    union(cumfields[[i]], cumfields[[i-1]])
csiz = sapply(cumfields,length)
bag_fields_ca1009 = unique(unlist(cumfields))
nfields = length(bag_fields_ca1009)
mydf = data.frame(date_published=stud_dates, nfields=csiz)

The growth in size of the set of fields in use over time is displayed here:

ggplot(mydf, aes(x=date_published, y=nfields)) + geom_point()

library(plotly)
ddf = data.frame(date=stud_dates[-1], newly_introduced_fields=diff(csiz),
    study=paste0(names(stud_dates[-1]), "\na"))

The next display is interactive -- hover over points to see study accession number and newly introduced field names.

incrs = lapply(2:length(cumfields), function(x) setdiff(cumfields[[x]],
   cumfields[[x-1]]))
incrs = unlist(lapply(incrs, function(x) paste0(x, collapse="\n")))
sn = names(stud_dates[-1])
incrs = paste(sn, incrs, sep="\n")
dddf = cbind(ddf, incrs)
g2 = ggplot(dddf, aes(x=date, y=newly_introduced_fields, text=incrs)) + geom_point()
ggplotly(g2)

Reference resources for reducing metadata isolation and variability

Use of common data elements is promoted by various initiatives. Dictionaries, thesauri, and ontologies are all relevant. We have examples of each in the metametrics package.

A snapshot of the Genomic Data Commons gdcdictionary, with fields and values related to diagnosis and sample characteristics is provided in gdc_dx_sam.

gdc_dx_sam

A table with all entries from several ontologies and the NCI Thesaurus is provided by load_ontolookup:

olook = load_ontolookup()
olook

Statistics on field use

We use robust linear modeling to estimate growth in vocabulary of fields employed over time. The data.frame mydf includes a variable nfields taking a value for each study publication date. The value of nfields associated with date $d$ records the the number of fields used to annotate all studies up to date $d$.

library(MASS)
nsecpy = 3600*24*365
summary( mm <- rlm(nfields~I(as.numeric(date_published)/nsecpy), data=mydf))
plot(nfields~I(as.numeric(date_published)/nsecpy), data=mydf)
abline(mm)

Proximity of terms in use to endorsed terminologies