#PIM

The Pairwise Interaction Model (PIM) describes the binding of a Transcription Factor (TF) to its DNA binding sites (BS) using the statistics of both single nulceotides and pairs of nucleotides. In the physics jargon, this model is known as an inhomogeneous Potts models with 4 colors.

##DESCRIPTION

The src folder contains scripts to estimate a variety of binding models from ChIPseq or ChIP-chip data, namely:

• The Position Weight Matrix (PWM) model, where the TF binds to each of the nucleotides independently.

• The PWM mixture model, where the TF is assumed to adopt several binding conformations, each of which is represented by a PWM and has a certain probability to occur.

• The Pairwise Interactions Model (PIM), where pairwise contributions from pairs of nucleotides are taken into account.

• The Nearest Neighbor Model (NNM), which is a PIM with pairwise contributions restricted to nearest-neighbor nucleotides.

The data folder contains an example ChIP-chip dataset for the Drosophila TF Twist that is used as default for the models estimation. It contains two folders:

• The data/nmers-chips folder contains a file giving the N-mers (here N=12) to be searched for binding sites. For each ChIP peak, a unique identifier was assigned (first column) and all N-mers from both strands were considered (second column). When learning a model, only the 50% best peaks having one or several high scoring BS are taken into account. The identifier allows to keep track of the peak on which the BS resides. Note that nothing restricts the models to be learnt from ChIP-chip or ChIPseq data, and such an N-mers file could be generated from other kinds of experiments.

• The data/motinits folder contains a file giving an initial frequency matrix for the TF of interest, where rows correspond to A,T,C,G frequencies and columns to different positions of the site.

For a given TF, both files should have the same name, which correspond the the MAT variable in the matlab scripts. For example, in the default case, MAT='twi' and the files are named twi.dat.

##FIRST STEPS

Most of the code is written in matlab. It has been executed with matlab versions greater than R2011b but has not been tested with older versions.

Before using the code, two things need to be taken care of.

1. First, the exhaust.c file has to be compiled using the mex function. To do that, open matlab in the src directory and execute

   mex exhaust.c

   This will produce a matlab compatible file that lies at the heart of the PIM estimation. The exhaust function loops across al possible N-mers (in our case, N=12) to compute several quantities of interest, such as the single and pairwise nucleotides probabilities estimated by the model or the Partition Function used for probability normalization. The computation time therefore grows exponentially with N, which prohibits from using large values (N=12 has proven to be a good compromise).

2. Second, the runMefirst.sh script should be executed:

   ./runMeFirst.sh

   This will produce the src/load_paths.m file that defines the absolute paths of the data and src folders, as well as the results directory that will be created when running the matlab scripts. These paths can be adapted to your need by modifying the definitions of the SRCDIR, DATADIR and RESDIR variables in the runMeFirst.sh script. All these directories are located at the root of the PIM folder per default. If the name or the location of any of the parent folders changes, this script should be executed again.

##USAGE

The src/compAllModels.m script can be executed to learn the different models on the sample Twist data. To do so, open matlab in the src folder and execute:

compAllModels

The different models can also be estimated separately. For example, if one is solely interested in learning the PWM model, then the src/PWM_model.m script should be executed.

The important parameters, such as the length of the binding sites N or the motif name, are all located in the file src/load_params.m, which is loaded in the beginning of all scripts. This is were any change should be made if one wants to conduct a custom study (look at a different TF, change N, etc).

##CITATION

Santolini M, Mora T, Hakim V (2014) A General Pairwise Interaction Model Provides an Accurate Description of In Vivo Transcription Factor Binding Sites. PLoS ONE 9(6): e99015. doi:10.1371/journal.pone.0099015