In teff-package/teff: Profiles and targets subpopulations with high Treatment EEFects

Introduction

teff is a software package to predict the effect of treating an individual on an outcome given the individual's profile in some feature data. The package focuses on transcriptomic features for which surrogate covariates need to be estimated. The estimation of treatment effects is based on inferences using random causal forest as implemented in the package grf by Tibshirani et al.

Here, we show how to use the package to estimate transcriptomic profiles with positive and negative sexual dimorphism on immune cell count. We use sex as a treatment and therefore search for the transcription profiles for which the effect of sex is high on the amount of immune cells in blood.

We have inferred the amount of T cells in blood in the GTEx individuals and have preselected the transcriptomic features with a differential expression analysis of the interaction between T cell count and sex (more details in Caceres et al. 2021, under preparation).

Data

library(teff)

Sys.setenv(VROOM_CONNECTION_SIZE=5000072)

The tcell data set is a list that contains the
teffdata with variables: eff for the outcome on which the effect is measured, namely the infered T-cell count in blood, t for treatment; 1:male, 2:female, and covariates such as age, bmi and cov5, which is a surrogate variable estimated for the transcription data, and associated to latent technical differences in their measurement.

names(tcell)
head(tcell$teffdata)

The transcriptomic feature data contains the expression levels of 14 genes, found significant in the differential expression analysis of the interaction between T cell count and sex.

dim(tcell$features)
head(tcell$features)

Analysis

We then use predictteff to estimate the effect of sex on T cell count in a subsample of test individuals. The function randomly selects 80\% of individuals to grow the forest using the feature data, adjusted by covariates. The estimated effect of sex is the estimated difference of T cell count between sexes when the transcriptomic data across these 14 genes are kept constant. We refer to the estimated effect of sex of an individuals as the individual's associated sexual dimorphism.

The predictor of sexual dimorphism is applied on the \20% of left-out individuals, who are used to estimate the effect of sex on each of them, given their gene expression levels across the genes. For the immune sexual dimorphism, the genes within the features belong to sex chromosomes. For those that are coupled with homologous genes between the sexual chromosomes, we estimate the average levels between homologs and use those as the features in the random forest. The matrix homologous couples the homologs between rows.

homologous<- matrix(c("DDX3Y","DDX3X","KDM5D","KDM5C",
                      "PRKY","PRKX","RPS4Y1","RPS4X",
                      "TXLNGY", "TXLNG", "USP9Y", "USP9X",
                      "XIST", "XIST", "TSIX", "TSIX"), nrow=2)

pred <- predicteff(tcell, featuresinf=homologous, profile=TRUE)

pred

We can plot the prediction of the associated sex effect with confidence intervals. Treated refers to females, and not treated to males.

plotPredict(pred, 
            ctrl.plot=list(lb=c("Male", "Female"),
                           wht="topleft", whs = "bottomright"))

For each individual predicteff predicts a sex effect with 95\% confidence intervals. Therefore, individuals that have a significantly negative effect of sex are those for which their specific transcription levels would result in lower T cell count when observed in females than in males. Those with significantly positive effects of sex are those for which their transcription profiles would result in higher T cell count when observed in females than males.

predicteff allows for a parameter profile that when it is set to TRUE it attempts to create a profile of gene expression levels across all individuals with significantly negative sex effect. It also creates a profile for the group of individuals with significantly positive sex effects.

pred$profile

The profile for positive dimorphism means, for instance, that the group of individuals who at the same time up-regulate XIST and TSIX and down-regulate PRKY-PRKX, TSIX-TSIX, KDM5D-KDM5C, USP9Y-USP9X, RPS4Y1-RPS4X, and DDX3Y-DDX3X presents more T cell counts in females than in men.

The sexual dimorphic profiles of immune cell count can be used to target individuals from other studies and test whether the profiles present different sex-associated risks or survival of different diseases. The following code shows how to set up data for the transcriptomic study in arthritis GSE17755 from GEO.

library(GEOquery)
gsm <- getGEO("GSE17755", AnnotGPL = TRUE)
gsm <- gsm[[1]]

data4teff <- feateff(gsm, tname="gender:ch1",
                     reft=c("male", "female"),
                     effname="disease:ch1",
                     refeff=c("healthy","arthritis"),
                     covnames="age:ch1", covtype="n",
                     sva=TRUE, UsegeneSymbol=TRUE)

The result has been saved in the data structure data4teff within the teff package, with only the genes within the sexual dimorphism profiles. We see how the data structure is a list with fields features and teffdata

data(data4teff)
names(data4teff)
names(data4teff$teffdata)

with variables: eff for the outcome on which the effect is measured, namely artrhitis diagnose, t for treatment; 1:male, 2:female, and covariates such as age, and surrogate variables.

The target function classifies individuals into their sexually dimorphic groups and produces a plot on the targeting, based on gene expression data available. It also tests the significance of the association of the outcome (arthritis disease) with the interaction between sex and the groups of sexual dimorphism.

res <- target(data4teff, pred, plot=TRUE, effect="positive", featuresinf=homologous, nmcov="age.ch1", model="binomial")

res

When the outcome is continuous a plot of the interaction can also be obtained. We now show the targeting of the sexual dimorphism on Tcell count on the entire tcell dataset.

data(tcell)
homologous<- matrix(c("DDX3Y","DDX3X","KDM5D","KDM5C","PRKY","PRKX","RPS4Y1","RPS4X","TXLNGY", "TXLNG", "USP9Y", "USP9X", "XIST", "XIST", "TSIX", "TSIX"), nrow=2)
pf <- predicteff(tcell, featuresinf=homologous, profile=TRUE)
res <- target(tcell, pf, effect="positiveandnegative", featuresinf=homologous, nmcov="age", model="log2")
res

Clearly, for this case, the association of the effect with the interaction between sex and the groups of sexual dimorphism is high because this data and this effect were used to infer the treatment effect groups. The highly significant interaction is illustrated by the interaction plot of mean estimates

plotTarget(res,  labs=c("Sexual dimorphism", "T cell count", "Condition", "Male", "Female"))

or by the boxplot of T cell count quartiles

boxPlot(res, 
        labs=c("Sexual dimorphism", "T cell count", "Condition", "Male", "Female"), 
        lg="bottomleft")

teff-package/teff documentation built on March 20, 2022, 8:25 p.m.