In netZoo/netZooR: Unified methods for the inference and analysis of gene regulatory networks

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Introduction

EGRET is a method for constructing individual-specific gene regulatory networks (GRNs), taking into account the underlying genotype of the individual in question. EGRET combines multiple lines of evidence (See Figure 1 below) in order to predict the effect of an individual's mutations on TF-to-gene edges and construct a complete, individual-specific bipartite GRN. TF motifs are used to construct a prior bipartite network of the presence or absence of TFs in the promoter regions of genes. This prior serves as an initial "guess" as to which TFs bind within the promoter regions of, and thus potentially regulate the expression of which genes. This prior is then modified to account for individual-specific genetic information using the individual's genotype combined with publicly available eQTL data as well as computational predictions of the effects of variants on TF binding using QBiC [1].

For a given individual and a given prior edge connecting TF i to gene j, the edge weight is penalized if the individual has a genetic variant meeting 3 conditions, namely, the individual must have (1) an alternate allele at a location within a TF binding motif in the promoter region of a gene, which (2) is an eQTL affecting the expression of the gene adjacent to the promoter and (3) must be predicted by QBiC to affect the binding of the TF corresponding to the motif at that location. Each of these data types is essential to the accurate capturing of variant-derived regulatory disruptions. The altered prior is then integrated with gene expression data and protein-protein interaction information to refine the edge weights using the PANDA message-passing framework [2]. The message-passing algorithm uses the logic that if two genes are co-expressed, they are more likely to be co-regulated and thus are more likely to be regulated by a similar set of TFs; conversely, if two proteins physically interact, they are more likely to bind promoter regions as a complex and thus are more likely to regulate the expression of a similar set of genes. The result is a individual-and-tissue-specific GRN taking into account the genotype information of the individual in question.

EGRET has been integrated into the netZooR package.

Install/load netZooR

If you do not have netZooR installed, you can install it from the development branch as follows:

#install.packages("devtools")
#devtools::install_github("netZoo/netZooR@devel")

Load the netZooR package:

library(netZooR)

EGRET input data

Get the example data sets

First download the example datasets:

system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_qbic.txt")
system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_genotype.vcf")
system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_motif_prior.txt")
system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_expr.txt")
system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_ppi_prior.txt")
system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_eQTL.txt")
system("curl -O  https://netzoo.s3.us-east-2.amazonaws.com/netZooR/unittest_datasets/EGRET/toy_map.txt")

Read in each of the data types.

qbic <- read.table("toy_qbic.txt", header = FALSE)
vcf <- read.table("toy_genotype.vcf", header = FALSE, sep = "\t", stringsAsFactors = FALSE, colClasses = c("character", "numeric", "character", "character", "character", "character", "character", "character", "character", "character"))
motif <- read.table("toy_motif_prior.txt", sep = "\t", header = FALSE)
expr <- read.table("toy_expr.txt", header = FALSE, sep = "\t", row.names = 1)
ppi <- read.table("toy_ppi_prior.txt", header = FALSE, sep = "\t")
qtl <- read.table("toy_eQTL.txt", header = FALSE)
nameGeneMap <- read.table("toy_map.txt", header = FALSE)

Let's take a look at each of the inputs for EGRET:

Motif prior

The motif prior is a bipartite network represented as a 3 column data frame. Each row represents an edge in the bipartite graph, with column 1 representing source TFs, column 2 representing target genes and column 3 representing the edge weight. The edge weight represents the presence (edge weight = 1) or absence (edge weight = 0) of the motif corresponding to the TF in column 1 in the promoter region of the gene in column 2. Note that, for ease of differentiating TF nodes from gene nodes, we name TFs with the TF name, and we name genes with their ensembl id.

head(motif)

Gene expression

The gene expression data represents gene expression measurements (in this case as TPMs from GTEx https://gtexportal.org/home/datasets) across several individuals. These are represented in a data frame with rows corresponding to genes and columns corresponding to samples/individuals. Row names of the data frame should be assigned gene names.

head(expr)

Protein-protein interaction (PPI) data

The PPI prior can be obtained from interaction databases such as String (https://string-db.org/). EGRET takes in a PPI network of TFs as a data frame in which each row represents an edge, with columns one and two corresponding to TF nodes and column 3 representing the interaction weight.

head(ppi)

eQTL data

The eQTL data consists of eQTL variants where the eQTL variant lies within a motif within the promoter region of the eGene. These are passed to EGRET as a data frame with the following columns: (1) TF corresponding to the motif in which the eQTL variant resides, (2) eGene adjacent to the promoter, (3) position of the eQTL variant, (4) chromosome on which the eQTL variant and eGene reside, and (5) beta value for the eQTL association. The eQTL data should be from the same cell type/tissue as the gene expression data and can be obtained from databases such as GTEx (https://gtexportal.org/home/datasets).

head(qtl)

Individual genotype

The genotype data for the individual in question should be loaded as a VCF file. Columns of the VCF used include column 1 (chromosome), column 2 (variant position), column 4 (reference allele), column 5 (alternate allele) and column 10 (genotype).

head(vcf)

QBiC predictions

EGRET requires QBiC [1] to be run on the eQTL variants occurring in the individual(s) in question in order to determine which transcription factor's binding is potentially disrupted due to the variant, at the location of the variant. QBiC makes use of models trained on protein binding microarray (PBM) data to predict the impact of a given variant on TF binding at that location. Some of QBiC's models are trained on non-human PBMs. We thus require a more stringent filtering (p < 1e-20) of resulting QBiC predictions from non-human models. We also require the predicted effect on binding to be negative (i.e. disruption of binding). QBiC predictions are passed to EGRET in a dataframe with the following columns: (1) variant as chr[num]_position which occurs within a motif in a promoter, (2) TF predicted to be impacted by QBiC, (3) gene adjacent to the promoter, (4) QBiC effect on binding. Note that multiple TFs can be predicted to have disrupted binding at a given variant.

head(qbic)

Run EGRET

Set a tag for the EGRET run. The EGRET outputs will be labeled with this tag.

tag <- "my_toy_egret_run"

Call the runEgret function to

runEgret(qtl,vcf,qbic,motif,expr,ppi,nameGeneMap,tag)

EGRET output

EGRET produces two output GRNs - a genotype specific "EGRET" network, and a genotype-agnostic baseline network (equivalent to a PANDA network).

load("my_toy_egret_run_egret.RData")
load("my_toy_egret_run_panda.RData")

head(regnetE)

head(regnetP)

References

1. Martin, V., Zhao, J., Afek, A., Mielko, Z. and Gordân, R., 2019. QBiC-Pred: quantitative predictions of transcription factor binding changes due to sequence variants. Nucleic acids research, 47(W1), pp.W127-W135. [https://doi.org/10.1093/nar/gkz363]
2. Glass, K., Huttenhower, C., Quackenbush, J. and Yuan, G.C., 2013. Passing messages between biological networks to refine predicted interactions. PloS one, 8(5), p.e64832. [https://doi.org/10.1371/journal.pone.0064832]

netZoo/netZooR documentation built on July 28, 2024, 6 p.m.