Integration: Liver cohort

Recent high-throughput technologies have caused an explosion in the number of cohort studies which monitor and record medical (phenotype, genetic), social (phenotype), and biological (phenotype, genotype, transcriptional) information of large number of human participants for a long period of time. This allows to identify the variables that, for example, correlate with occurrence of diseases. Researchers hope that the causative events that have led to the disease are among the highest correlating variables. Given the limited number of validation experiments that a scientist can perform during his/her career, being able to narrow down the number of candidate variables without loosing the causative variable is of utmost importance. From population health perpective, earlier diagnosis of disease or slightly better prediction of disease progress (higher accuracy and precision) can have significant effects on population health and its costs. Machine learning algorithms can integrate information from multiple sources and are, therefor, most widely used.

This workflow uses data acquired in a cohort study by Schadt EE et al., 2008 on over 400 human liver samples. First gene expression, mutation and phenotype data from are cleaned up by removing variables containing missing data and patients which lack expression, mutation, or phenotype data. Multiple machine leaning algorithms such as Support Vector Machines, Native Bayesian, and Robust regression are than used to create predictive models.

Support Vector Machine (‘e1071’ package) is used to train a model (classification and regression) using random sample of 2/3 of samples (training set) and tested with the remaining 1/3 of samples (test set). Model is trained based on independent variables such as age and liver triglyceride levels, and dependent variables such as activity of nine liver enzymes.

Using enzyme activity information, heatmaps (‘stats’ package) are generated to visualize correlation between enzymes and correlation between patients. For clustering of patients based on enzyme activity data, heatmap are grouped based on Principal Component Analysis results (prcomp from ‘stats’ package). Naive Bayesian classifier (‘e1071’ package) is used to cluster samples based on gene expression profile with aldehyde oxydase levels or liver enzyme activity as class vector (independent variable).

Furthermore, Robust linear model (‘MASS’ package) is used to train model using liver triglyceride levels and gene expression levels of genes with highest variance. A random sample of 25 genes are selected from 1000 genes with the highest variance and used in combination with triglyceride level phenotype. This is done in 50 iteration and the iteration with highest correlation in training and test sets is recorded.

Diagram for integration livercohort benchmark.

Packages and Dependencies

There are 3 packages used in this workflow, which depend on 1 additional package from CRAN (dependency)

Used packages:

• CRAN: stats, e1071, MASS

Package dependencies:

• CRAN: class

Data

from Human Liver Cohort (Synapse ID: syn4499)

License

• Copyright (c) 2015 Ieuan Clay based on code from genbench
• Copyright (c) 2015-2016 BeDataDriven B.V. License: GPL version 2 or higher