The sim_reg function uses MCMC to sample parameters from the ‘Phenotype similarity regression’ model conditional on supplied binary genotypes and ontological term sets.
The vignette ‘Similarity Regression - Introduction’ shows a simple application based on simulated data. This vignette demonstrates how to include prior phenotypic information in the inference. The information is supplied to the inference procedure as a parameter called lit_sims and should be a numeric vector of relative weights for terms included in the sample space of phi (by default the set of all terms present amongst the terms in x and their ancestors). Terms not specified in this vector default to having a weight of 1.
library(ontologyIndex)
library(ontologySimilarity)
library(SimReg)
data(hpo)
set.seed(1)
terms <- get_ancestors(hpo, c(hpo$id[match(c("Abnormality of thrombocytes","Hearing abnormality"),
hpo$name)], sample(hpo$id, size=50)))To help illustrate the ideas, we’ll consider a scenario where there is some evidence for an association between a phenotype - which in this example we’ll set as ‘Hearing abnormality’ - and a binary genotype, and where there is also a prior expectation that the ‘characteristic phenotype’ of the genotype would involve hearing abnormality. For instance, the genotype might depend on variants in a gene orthologous to one known to harbour ‘Hearing abnormality’ variants in a model organism. We apply the inference procedure to the data with and without the prior and observe the effect on the inferred ‘probability of association’, gamma.
hearing_abnormality <- hpo$id[match("Hearing abnormality", hpo$name)]
genotypes <- c(rep(TRUE, 4), rep(FALSE, 96))
#give all subjects 5 random terms and add 'hearing abnormality' for those with y_i=TRUE
phenotypes <- lapply(genotypes, function(y_i) minimal_set(hpo, c(
if (y_i) hearing_abnormality else character(0), sample(terms, size=5))))So there are three cases with the rare variant (i.e. having y_i = TRUE) and all of them have the ‘Hearing abnormality’ HPO term.
An application of yields:
samples <- sim_reg(ontology=hpo, x=phenotypes, y=genotypes)
print(summary(samples), ontology=hpo)## ---------------------------------------------------------------------------
## P(gamma=1|y) = 0.339
## ---------------------------------------------------------------------------
## Phi:
## t Name P
## HP:0000364 Hearing abnormality 0.59
## HP:0000598 Abnormality of the ear 0.33
## HP:0011277 Abnormality of the urinary system physiology 0.13
## HP:0003110 Abnormality of urine homeostasis 0.09
## HP:0000079 Abnormality of the urinary system 0.09
## HP:0010343 Aplasia/Hypoplasia of the 5th toe 0.06
## HP:0001939 Abnormality of metabolism/homeostasis 0.06
## HP:0040064 Abnormality of limbs 0.05
## HP:0003040 Arthropathy 0.05
## HP:0002813 Abnormality of limb bone morphology 0.05
## ---------------------------------------------------------------------------We now consider constructing the lit_sims parameter to capture our knowledge about the gene from the model organism. We can either explicitly create the lit_sims vector of prior weights, i.e. by assigning higher weights to terms which involve some kind of hearing problem. For example, we could set the prior weight of all terms which have the word ‘hearing’ in to ten times that of terms which don’t.
lit_sims <- ifelse(grepl(x=hpo$name, ignore=TRUE, pattern="hearing"), 10, 1)
names(lit_sims) <- hpo$nameNote: one must set the names of the lit_sims vector, as sim_reg will use it.
If the prior knowledge of the phenotype/phenotype of the model organism has been ontologically encoded (for example, it may be available as MPO terms from the Mouse Genome Informatics (MGI) website, http://www.informatics.jax.org/, [1]), another option is to use a phenotypic similarity function to obtain the numeric vector of weights for inclusion of terms in phi [2]. This may be more convenient, particularly when dealing with large numbers of genes. In the SimReg paper [2], the vector is set by exponentiating the Resnik-based [3] similarities of terms to the terms in the ‘literature phenotype’. In order to calculate the similarities based on Resnik’s similarity measure, we must first compute an ‘information content’ for the terms, equal to the negative log frequency. The frequencies can be calculated with respect to different collections of phenotypes. Here, we will calculate it with respect to the frequencies of terms within our collection, phenotypes, by calling the function exported by the ontologyIndex package, get_term_info_content. Note, it could also be calculated with respect to the frequency of the term amongst the ontological annotation of OMIM diseases (available from the HPO website, http://human-phenotype-ontology.github.io [4]). The function get_term_set_to_term_sims in the package ontologySimilarity can then be used to calculate the similarities between the terms in the sample space of phi and the literature_phenotype. It calculates a matrix of similarities between the individual terms in the literature phenotype and terms in the sample space. Let’s say the phenotype of the model organism in our example contains abnormalities of the thrombocytes and hearing abnormality.
thrombocytes <- hpo$id[match("Abnormality of thrombocytes", hpo$name)]
literature_phenotype <- c(hearing_abnormality, thrombocytes)
info <- get_term_info_content(hpo, phenotypes)
lit_sims_resnik <- apply(exp(get_term_set_to_term_sims(
ontology=hpo, information_content=info, terms=names(info),
term_sim_method="resnik", term_sets=list(literature_phenotype))), 2, mean)This can then be passed to sim_reg through the lit_sims parameter.
with_prior_samples <- sim_reg(
ontology=hpo,
x=phenotypes,
y=genotypes,
lit_sims=lit_sims_resnik
)
print(summary(with_prior_samples), ontology=hpo)## ---------------------------------------------------------------------------
## P(gamma=1|y) = 0.6375
## ---------------------------------------------------------------------------
## Phi:
## t Name P
## HP:0000364 Hearing abnormality 0.66
## HP:0000598 Abnormality of the ear 0.32
## HP:0011277 Abnormality of the urinary system physiology 0.13
## HP:0003110 Abnormality of urine homeostasis 0.09
## HP:0000079 Abnormality of the urinary system 0.09
## HP:0040064 Abnormality of limbs 0.09
## HP:0002813 Abnormality of limb bone morphology 0.09
## HP:0000346 Whistling appearance 0.07
## HP:0006768 Localized neuroblastoma 0.06
## HP:0001939 Abnormality of metabolism/homeostasis 0.05
## ---------------------------------------------------------------------------Note that including the lit_sims parameter has increased the mean posterior value of gamma.
Often the binary genotype relates to a particular gene, and for many genes ontologically encoded phenotypes are available either in the form of HPO encoded OMIM annotations [4] or MPO annotations [1]. For a given set of subjects with HPO-coded phenotypes, it may be useful to apply the inference gene-by-gene, taking the binary genotype y to indicate the presence of a rare variant in each particular gene for each case. Thus, we may wish to systematically create informative prior distributions for phi for all genes. This can be done by downloading the file called ‘ALL_SOURCES_TYPICAL_FEATURES_genes_to_phenotype.txt’ from the HPO website, and running the following code yielding a list of term sets (i.e. character vectors of HPO term IDs).
annotation_df <- read.table(header=FALSE, skip=1, sep="\t",
file="ALL_SOURCES_TYPICAL_FEATURES_genes_to_phenotype.txt", stringsAsFactors=FALSE, quote="")
hpo_by_gene <- lapply(split(f=annotation_df[,2], x=annotation_df[,4]),
function(trms) minimal_set(hpo, intersect(trms, hpo$id)))