In isglobal-brge/DiabPred: Type 2 diabetes prediction based on genomic data with deep learning algorithms

knitr::opts_chunk$set(collapse = TRUE, comment = "", 
                      cache=TRUE, message = FALSE, 
                      warning = FALSE
                      )

Purpose of this package

The main objective of this package is to offer an accurate way of predicting type 2 diabetes based on genomic data using deep learning neural networks. Said neural networks was trained on SNPs found significant when tested for association with diabetes. For the purpose of being able to make accurate predictions, all the SNPs that were used to train the neural network must be present in the GDS file used as input.

The list of SNPs can be accessed to using the following code:

# Loading list of needed SNPs
snp <- data(snps)

head(snps)

As it can be seen it returns a vector consisting of all the SNPs that must be present in the GDS file.

Getting started

For the purpose of this vignette we will need the packages $SNPRelate$ and $h2o$. Former one is needed in order to preprocess the data that will serve as input for the prediction function, while the later provides the tools to predict the outcome.

$h2o$ depends on the packages $jsonlite$ and $RCurl$ which should also be installed:

if (!"BiocManager" %in% rownames(installed.packages()))
     install.packages("BiocManager")
BiocManager::install("SNPRelate")

pkgs <- c("RCurl","jsonlite")
for (pkg in pkgs) {
if (! (pkg %in% rownames(installed.packages()))) { install.packages(pkg) }
}

install.packages("h2o", type="source", repos="http://h2o-release.s3.amazonaws.com/h2o/rel-yates/3/R")

Link provided for the installation of the $h2o$ package is the latest version as of r Sys.Date(). Please do consult the following website to check for new releases h2o.

Create GDS from PLINK data

A GDS-class object should always be used as input for the predictions. GDS object can be obtained from PLINK data as it will be demonstrated, or otherwise from VCF data which will also be included as code despite only plink files being available as test files in this package, regardless of this the output ought to be the same.

library(SNPRelate)

The $SNPRelate$ package provides a function snpgdsBED2GDS() for converting a PLINK text/binary file to a GDS file:

# PLINK files included in the DiabPred package
bed <- system.file("extdata","test.bed.gz", package = "DiabPred")
fam <- system.file("extdata","test.fam.gz", package = "DiabPred")
bim <- system.file("extdata","test.bim.gz", package = "DiabPred")

Or use your own PLINK files:

bed <- "C:/path/to/your_plink_file.bed"
fam <- "C:/path/to/your_plink_file.fam"
bim <- "C:/path/to/your_plink_file.bim"

#Convert
snpgdsBED2GDS(bed, fam, bim, "test.gds")

# Summary
snpgdsSummary("test.gds")

If not specified, the test.gds file will be created in the actual working directory.

If the files needed to be converted to GDS file are VCF files, the function snpgdsVCF2GDS() is provided by the same package. There are two options for extracting markers from the VCF file for posterior analysis: 1. Extract and store the reference allele only for biallelic SNPs; 2. Extract and store the reference allele for all variant sites, including biallelic SNPs, multiallelic SNPs, structural variants and indels.

vcf <- "C:/path/to/your_vcf_file.vcf"

snpgdsVCF2GDS(vcf, "test.gds", method = "biallelic.only")

Type 2 diabetes risk prediction

Our function, which receives the same name as the package as it's its main component DiabPred makes use of $SNPRelate$ for the filtering and preparation of the data for the prediction, while $h2o$ is used to return them.

# GDS object load
genofile <- snpgdsOpen("test.gds")

# Prediction
predictions <- DiabPred(genofile)

# Results
head(predictions)

As it can be seen, the DiabPred function has a single output which is a data.frame consisting of 3 columns and as many rows as individuals were present in the GDS file. The 3 columns correspond in their respective order to:

• Diabetes prediction, 0 for non-diabetic, 1 for diabetic
• Probability of the individual being non-diabetic
• Probability of the individual being diabetic.

Session Information

sessionInfo()

isglobal-brge/DiabPred documentation built on June 14, 2019, 11:02 p.m.