# rPWMDmm-methods: rPWMDmm method In TFBSTools: Software Package for Transcription Factor Binding Site (TFBS) Analysis

## Description

This function samples matrices from trainned Dirichlet mixture model based on selected matrices.

## Usage

 `1` ``` rPWMDmm(x, alpha0, pmix, N=1, W=6) ```

## Arguments

 `x` x can be a `matrix`, `PFMatrixList`. The count matrix on which the sampling is based. `alpha0` The trained Dirichlet mixture parameters. `pmix` The trained mixing proportions of the components. `N` The number of matrices to sample. `W` The desired width of matrice from the sampling.

## Details

This feature enables the users to generate random Position Frequency Matrices (PFMs) from selected profiles.

We assume that each column in the profile is independent and described by a mixture of Dirichlet multinomials in which the letters are drawn from a multinomial and the multinomial parameters are drawn from a mixture of Dirichlets. Within this model each column has its own set of multinomial parameters but the higher level parameters – those of the mixture prior is assumed to be common to all Jaspar matrices. We can therefore use a maximum likelihood approach to learn these from the observed column counts of all Jaspar matrices. The maximum likelihood approach automatically ensures that matrices receive a weight relative to the number of counts it contains.

Drawing samples from the prior distribution will generate PWMs with the same statistical properties as the Jaspar matrices as a whole. PWMs with statistical properties like those of the selected profiles can be obtained by drawing from a posterior distribution which is proportional to the prior times a multinomial likelihood term with counts taken from one of the columns of the selected profiles.

Each 4-dimensional column is sampled by the following three-step procedure: 1. draw the mixture component according to the distribution of mixing proportions, 2. draw an input column randomly from the concatenated selected profiles and 3. draw the probability vector over nucleotides from a 4-dimensional Dirichlet distribution. The parameter vector alpha of the Dirichlet is equal to the sum of the count (of the drawn input) and the parameters of the Dirichlet prior (of the drawn component).

Draws from a Dirichlet can be obtained in the following way from Gamma distributed samples: (X1,X2,X3,X4) = (Y1/V,Y2/V,Y3/V,Y4/V) ~ Dir(a1,a2,a3,a4) where V = sum(Yi) ~ Gamma(shape = sum(ai), scale = 1).

## Value

A list of matrices from the sampling.

## Methods

`signature(x = "PFMatrix")`
`signature(x = "matrix")`
`signature(x = "PFMatrixList")`

## Note

This code is based on the Matlab code original written by Ole Winther, binf.ku.dk, June 2006.

Ge Tan

## References

L. Devroye, "Non-Uniform Random Variate Generation", Springer-Verlag, 1986

Kimura, T., Tokuda, T., Nakada, Y., Nokajima, T., Matsumoto, T., & Doucet, A. (2011). Expectation-maximization algorithms for inference in Dirichlet processes mixture. Pattern Analysis and Applications, 16(1), 55-67. doi:10.1007/s10044-011-0256-4

`dmmEM`
 ```1 2 3 4 5 6 7``` ``` data(MA0003.2) data(MA0004.1) pfmList <- PFMatrixList(pfm1=MA0003.2, pfm2=MA0004.1, use.names=TRUE) dmmParameters <- dmmEM(pfmList, 6) rPWMDmm(MA0003.2, dmmParameters\$alpha0, dmmParameters\$pmix, N=1, W=6) ```