As explained in the class, the primary goal of this project is to build machine learning algorithms for predicting gene expression levels using genotypes of cis-SNPs. This is an active research area in genetics and genomics, the following papers provide necessary background on this topic.
- Gamazon et al. A gene-based association method for mapping traits using reference transcriptome data
- Gusev et al. Integrative approaches for large-scale transcriptome-wide association studies
Download a single zipped tar file project_data.tgz from here. Note that the file size is about 571Mb, make sure you have enough disk space to save and operate on the downloaded file.
To decompress the file, run the following command in a UNIX/Linux/Mac terminal
tar zxf project_data.tgz
The whole data set contains text formated data files for 68 genes. There are 3 separate files for each gene:
- genotype data for training samples (named as
ENSG###.training_geno.dat): genotype data matrix (Nxp) for a given gene. For all genes, training data sample size N=570. The number of cis-SNPs (i.e., the predictors) varies by genes. The summary of distribution of p across 68 genes is given by
Min. 1st Qu. Median Mean 3rd Qu. Max.
2178 5200 5736 6618 7056 17060
Note that genotype data file contains a header of SNP names.
-
phenotype (i.e. expression level) data for training samples (named as
ENSG###.training_pheno.dat): gene expression data matrix (Nx1) for a given gene. -
genotype data for testing samples (named as
ENSG###.testing_geno.dat): genoype data matrix (Mxp) for a given gene. For all genes, testing data sample size M=100. This genotype data file also contains a header of SNP names.
You may use/modify the following R code to test the downloaded training data by performing single SNP association analysis
# load in training genotype data
d = read.table("ENSG00000211955.2.training_geno.dat",head=T)
dim(d)
# load in gene expression data
y = as.matrix(read.table("ENSG00000211955.2.training_pheno.dat"))
# get all p-values by single SNP association testing
pval = apply(d,2,function(x) summary(lm(y~x))$coef[2,4])
summary(pval)
You can choose one of the following two tasks to work on.
This is a prediction problem. You may choose and implement any algorithm to predict the gene expression levels of the testing samples. The accuracy of the prediction will be judged by the squared rank correlation (R^2) between the predicted values and the actual values (not accessible to you)
This is a variable selection problem. You may use any algorithm of your choice to identify potentially multiple association signals. (In this case, the testing genotype data may not be useful to you.) Importantly, some measurement/quantification/justification of the type I errors of your identified signals is required.