bootNet ('strapping the elastic-net)

bootNet is a wrapper for the fantastic glmnet R package - it brings bootstrapping and parallel processing to the elastic-net framework.

This script was originally designed to analyse methylation data in the form of beta matrices. The beta matrix must have CpG sites as rows and samples as columns for bootNet to work.

With a recent update bootNet is now data agnostic and has been tested with SNP and expression data (as well as methylation).

Updates

2020-07-30

Added a directory (here) which includes code and documentation contributed by Sean Burnard which is used in a recent manuscript exploring genetic associations in multiple sclerosis. DOI and citation will be added shortly.

Version: 0.1.2.0

• code base overhauled to be data agnostic (i.e. it can be used for methylation, SNP, expression data, but also any matrix of data if set up correctly in normal glmnet format)
• added an example data set in /data, this is a collection of blood and buccal cell samples

Version: 0.1.1.1

• This update adds a parallel version of the bootNet function to utilise multiple cores if available.

WARNING: be aware of the amount of available system RAM when using bootNet.parallel(), if the data set is large even running across 4-8 cores will quickly utalise many GB of RAM - you have been warned! Some real-world usage metrics across different Linux systems are provided below.

• Added functionality to detect NAs in data/outcome and terminate the function

• Created a gh-pages branch that will eventually house a website for bootNet: here

Example usage:

What do you need to provide the bootNet script?

• data - data matrix (formatted predictors as rows, samples as columns)
• outcome - either as numeric (quantitative trait) or as factor (qualitative trait)
• Alpha - sets the mixture between ridge-regression and lasso
• iter - the number of iterations to bootstrap
• sub_sample - percentage of sample/case-control groups to sub-sample for bootstrapping
• sampleID - a list of sample IDs (usually the column names of the beta matrix). These MUST be in the same order as the samples in the beta matrix
• method - optional string specifying one of two methods ('JACKKNIFE', 'LOOCV')
• cores - number of cores to use for bootNet.parallel

bootNet()

bootNet(data = x, outcome = y, Alpha = 0.1, iter = 1000, sub_sample = 0.666, sampleID = sampleID, method = method)

bootNet.parallel()

bootNet.parallel(data = x, outcome = y, Alpha = 0.1, iter = 1000, sub_sample = 0.666, cores = 4, sampleID = sampleID)

To do list

• add functionality to automatically 'tally' the results and rank them by number of iterations passed
• merge development script back into the main code base
• add BOOTNET, JACKKNIFE and LOOCV methods to main script
• add BOOTNET, JACKKNIFE and LOOCV methods to the parallel method
• add example data set and a decent tutorial
• include additional quality checks
  • ensure outcome is either numeric or factor
  • check for missing data (NAs) in both data and outcome
• explore bias estimation and performance of selected markers
• implement a double-bootstrap method as an additional way to explore performance metrics
  • separate training and test data
  • bootstrap both data sets
  • compare selected sites
• look into AUC and ROC as another form of marker selection/validation
• add an argument to perform "leave one out" analysis (alternative to sub-sampling)
• write up as manuscript
• add option for paired samples in LOO-CV method
  • bootNet: identifying robust classifiers in methylation data

Performance expectations

A few test examples showing systems and performance metrics.

Experiment 1: 24 samples run on Illumina 450K methylation array (24 columns, 446280 rows). Phenotype was quantitative (age).

System used (Dell laptop: Precision M4800):

• Intel Core i7-4900MQ (4 core, 8 threads)
• 32GB DDR3
• 256GB SSD
• Linux - 4.3.0-towo.3-siduction-amd64 x86_64 (64 bit)
  • Running 100 iterations and using 4 cores the maximum observed RAM usage was 20GB, taking 85.6 seconds.

Experiment 2: 75 samples run on Illumina 450K methylation array (75 columns, 445998 rows). Phenotype was quantitative (glucose).

System used (Dell workstation):

• Intel Xeon E5-2620 (6 cores, 12 threads)
• 128GB DDR3
• 256GB SSD
• Linux - Debian Sid (64 bit)
  • Running 100 iterations and using 10 cores the maximum observed RAM usage was 30GB, taking 120.9 seconds.
  • Running the above at 1000 iterations and using 10 cores the maximum observed RAM usage was 30GB, taking 1152.8 seconds.
  • Running the above at 5000 iterations and using 10 cores the maximum observed RAM usage was 30GB, taking 5718.8 seconds.

Experiment 3: 105 samples run on Illumina 450k methylation array (105 columns, 380777 rows). Phenotype was qualitative (BMI separated into two categories)

System used (Z170-D3H)

• Intel Core i7-6700 3.40GH (4 cores, 8 threads)
• 16GB DDR4 (2 x 8GB @ 2133 MHz)
• 2TB Western Digital Harddrive
• Linux Ubuntu 14.04 LTS (64 bit)
  • 10 iterations and using 1 core the maximum observed RAM usage was 5GB, taking 34.42089 secs.
  • 10 iterations and using 2 cores the maximum observed RAM usage was 9GB, taking 20.0074 secs.
  • 10 iterations and using 3 cores the maximum observed RAM usage was 12.5GB, taking 18.31608 secs.
  • 20 iterations and using 3 cores the maximum observed RAM usage was 13.8GB, taking 33.594 secs.
  • 30 iterations and using 3 cores the maximum observed RAM usage was 12.4GB, taking 48.210 secs.
  • 50 iterations and using 3 cores the maximum observed RAM usage was 13GB, taking 83.090 secs.
  • 100 iterations and using 3 cores the maximum observed RAM usage was 13.6GB, taking 161.983 secs.
  • 1000 iterations and using 3 cores the maximum observed RAM usage was 13 GB, taking 1628.748 secs.

Dependencies

There are 3 R packages required by bootNet:

• glmnet
• foreach (bootNet.parallel)
• doParallel (bootNet.parallel)