In vitkl/NetFeaturePval: Find Features Enriched among Nodes that Interact with Each Single Node

knitr::opts_chunk$set(echo = TRUE)
library(NetFeaturePval)

Load example dataset

These are interactions of viral and human proteins along with protein domain annotations. The aim of the analysis is to predict domains in human proteins likely to bind viral proteins. More details and be found in a paper.

data = fread("../data/viral_human_net_w_domains", sep = "\t", stringsAsFactors = F)

Find enriched domains

First, let's use multi-core parallelisation. Correct analysis would require many more premutations but this is just an example of how to use this package.

res = permutationPval(interactions2permute = IDs_interactor_viral ~ IDs_interactor_human,
                      associations2test = IDs_interactor_viral ~ IDs_domain_human,
                      # node_attr gives a way to add columns needed to compute statistic and filter the data,
                      # in this case only domain_count is needed to filter 
                      # and node columns to compute statistic
                      node_attr = list(IDs_interactor_viral ~ IDs_interactor_viral_degree,
                                       IDs_domain_human ~ domain_count,
                                       IDs_interactor_viral + IDs_domain_human ~ domain_frequency_per_IDs_interactor_viral), 
                      data = data,
                      # in this example statistic is just count of 
                      # IDs_interactor_human for each IDs_interactor_viral / IDs_domain_human pair
                      statistic = IDs_interactor_viral + IDs_domain_human ~ .N,
                      select_nodes = IDs_domain_human ~ domain_count >= 1,
                      N = 1, # number of permutations
                      cores = 1, # how many cores to use on a local machine
                      # computations are split into inner and outer replicate 
                      # to help manage memory load, here it is 50*2
                      # hint: you can use microbenchmark package to find optimal value
                      split_comp_inner_N = 1, 
                      seed = 2) # seed for reproducible sampling (permutations)

Next, let's try multi-node parallelisation (on a computing cluster). Default installation of clustermq will still work but run multiple processes on a local machine.

res = permutationPval(interactions2permute = IDs_interactor_viral ~ IDs_interactor_human,
                      associations2test = IDs_interactor_viral ~ IDs_domain_human,
                      node_attr = list(IDs_domain_human ~ domain_count),
                      data = data,
                      statistic = IDs_interactor_viral + IDs_domain_human ~ .N,
                      select_nodes = IDs_domain_human ~ domain_count >= 1,
                      N = 30, 
                      clustermq = T, # use clustermq
                      clustermq_jobs = 3, # how many cluster jobs to start 
                      # how much memory each job needs
                      clustermq_mem = 2000, 
                      # caution: allocating more memory than needed is a waste of resources,
                      # but allocating not enough will result in you jobs being killed,
                      # in turn, this will freese R requiring restart
                      # hint: try 2-3 jobs with realistic load and excess of memory
                      split_comp_inner_N = 2, seed = 2)

Look at the results and p-value distribution

res
plot(res)

Date and packages used

Sys.Date. = Sys.Date()
Sys.Date.
session_info. = devtools::session_info()
session_info.

vitkl/NetFeaturePval documentation built on May 19, 2019, 9:39 p.m.