In armenabnousi/naddaR: Prediction of Protein Conserved Regions Using NADDA Algorithm

---

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Overview of NADDA

NADDA[@abnousi2016fast] uses k-mer frequeuncies of the protein sequences in an input dataset to generate a vectorized representation of each protein. Then it utilizes the generated representative vectors, called k-mer ferquency profiles, as features for prediction of conserved regions, i.e., domains and motifs in the sequences. This package (naddaR) provides the functions required for generation of instances that can be used to train a new model for conserved index prediction and for generation of test instances that can be used as input to a trained model to predict the conserved indices on them.

Generating Training and Test Instances

To generate a training or test set, use the function generate_instances(). The labeled parameter for this function decides is to indicate whether a training set should be generated or a test set. For generation of a training set, the current version of the package uses Pfam or InterPro annotations. These annotations should be provided as a tsv (tabulated) file using the truth_filename parameter.

In the following example we generate a dataset of three protein sequences, calling them "seq1", "seq2", and "seq3". Then we use the generate_instances method to create a test set.

library(naddaR)
print("now printing that")
methods("generate_instances")
print("passed test print statement")
library(Biostrings)
library(data.table)

## Generate a set of three example protein sequences
seqs <- AAStringSet(c("seq1"="MLVVD",
                      "seq2"="PVVRA",
                      "seq3"="LVVR"))
## Generate a test set using the dataset seqs
ins <- generate_instances(seqs, labeled = FALSE, parallel = FALSE, klen = 3, impute = TRUE, winlen = 5, normalize = FALSE)
head(ins)

# To generate a training set using Pfam annotation located at user/seq_pfam.tsv the following command can be used.
# ins <- generate_instances(seqs, labeled = FALSE, parallel = FALSE, groundtruth = "Pfam", 
#                               truth_filename = "user/seq_pfam.tsv",  klen = 3, impute = TRUE, 
#                               winlen = 5, normalize = FALSE)

Note that when the functions are run in parallel, each processor will hold only a chunk of the data. In this case, to perform the training on the complete set of data, one will need to aggregate all chunks and input them to a training method. For this purpose we suggest writing all instances on a single file on the shared memory, using pbdMPI::comm.write method from the pbdMPI package. After writing all instances to file, then one can use a single processor to read the data and perform the training using a package of choice, either in parallel or in serial.

Other functionalities

In addition, one can use more low-level functions to generate k-mer frequency profiles or to count the number of kmers (create a histogram of the kmers in the dataset). For more information on how to utilize these functionalities, please refer to the documentation of the generate_profiles and count_kmers methods.

When the count_kmers method is run in parallel (by setting parallel or distributed logical parameters), it computes the local k-mer counts for a chunk of data available to each processor. Then allgather method from pbdMPI is used to communicate these frequencies between different processors. Finally each processor sums up all acquired frequencies, thus generating a dataframe that is equal in between all processors and holds the number of times each k-mer appears in the dataset (Note: if a k-mer appears more than once in one sequence, it is counted only once.)

References:

armenabnousi/naddaR documentation built on May 24, 2019, 8:47 p.m.