T.fit: Makes a stepwise regression fit for time series gene...
In maSigPro: Significant Gene Expression Profile Differences in Time Course Gene Expression Data

Description Usage Arguments Details Value Author(s) References See Also Examples

T.fit selects the best regression model for each gene using stepwise regression.

1 2	T.fit(data, design = data$dis, step.method = "backward", min.obs = data$min.obs, alfa = data$Q, nvar.correction = FALSE, family = gaussian(), epsilon=0.00001, item="gene")

`data`	can either be a `p.vector` object or a matrix containing expression data with the same requirements as for the `p.vector` function
`design`	design matrix for the regression fit such as that generated by the `make.design.matrix` function. If data is a `p.vector` object, the same design matrix is used by default
`step.method`	argument to be passed to the step function. Can be either `"backward"`, `"forward"`, `"two.ways.backward"` or `"two.ways.forward"`
`min.obs`	genes with less than this number of true numerical values will be excluded from the analysis
`alfa`	significance level used for variable selection in the stepwise regression
`nvar.correction`	argument for correcting T.fit significance level. See details
`family`	the distribution function to be used in the glm model. It must be the same used in p.vector
`epsilon`	argument to pass to `glm.control`, convergence tolerance in the iterative process to estimate de glm model
`item`	Name of the analysed item to show in the screen while T.fit is in process

In the maSigPro approach p.vector and T.fit are subsequent steps, meaning that significant genes are first selected on the basis of a general model and then the significant variables for each gene are found by step-wise regression.

The step regression can be "backward" or "forward" indicating whether the step procedure starts from the model with all or none variables. With the "two.ways.backward" or "two.ways.forward" options the variables are both allowed to get in and out. At each step the p-value of each variable is computed and variables get in/out the model when this p-value is lower or higher than given threshold alfa. When nva.correction is TRUE the given significance level is corrected by the number of variables in the model

`sol`	matrix for summary results of the stepwise regression. For each selected gene the following values are given: p-value of the regression ANOVA R-squared of the model p-value of the regression coefficients of the selected variables
`sig.profiles`	expression values for the genes contained in `sol`
`coefficients`	matrix containing regression coefficients for the adjusted models
`groups.coeffs`	matrix containing the coefficients of the impiclit models of each experimental group
`variables`	variables in the complete regression model
`G`	total number of input genes
`g`	number of genes taken in the regression fit
`dat`	input analysis data matrix
`dis`	regression design matrix
`step.method`	imputed step method for stepwise regression
`edesign`	matrix of experimental design
`influ.info`	data frame of genes containing influencial data

Ana Conesa, aconesa@cipf.es; Maria Jose Nueda, mj.nueda@ua.es

Conesa, A., Nueda M.J., Alberto Ferrer, A., Talon, T. 2006. maSigPro: a Method to Identify Significant Differential Expression Profiles in Time-Course Microarray Experiments. Bioinformatics 22, 1096-1102

p.vector, step

#### GENERATE TIME COURSE DATA
## generate n random gene expression profiles of a data set with 
## one control plus 3 treatments, 3 time points and r replicates per time point.

tc.GENE <- function(n, r,
             var11 = 0.01, var12 = 0.01,var13 = 0.01,
             var21 = 0.01, var22 = 0.01, var23 =0.01,
             var31 = 0.01, var32 = 0.01, var33 = 0.01,
             var41 = 0.01, var42 = 0.01, var43 = 0.01,
             a1 = 0, a2 = 0, a3 = 0, a4 = 0,
             b1 = 0, b2 = 0, b3 = 0, b4 = 0,
             c1 = 0, c2 = 0, c3 = 0, c4 = 0)
{

  tc.dat <- NULL
  for (i in 1:n) {
    Ctl <- c(rnorm(r, a1, var11), rnorm(r, b1, var12), rnorm(r, c1, var13))  # Ctl group
    Tr1 <- c(rnorm(r, a2, var21), rnorm(r, b2, var22), rnorm(r, c2, var23))  # Tr1 group
    Tr2 <- c(rnorm(r, a3, var31), rnorm(r, b3, var32), rnorm(r, c3, var33))  # Tr2 group
    Tr3 <- c(rnorm(r, a4, var41), rnorm(r, b4, var42), rnorm(r, c4, var43))  # Tr3 group
    gene <- c(Ctl, Tr1, Tr2, Tr3)
    tc.dat <- rbind(tc.dat, gene)
  }
  tc.dat
}

## Create 270 flat profiles
flat <- tc.GENE(n = 270, r = 3)
## Create 10 genes with profile differences between Ctl and Tr1 groups
twodiff <- tc.GENE (n = 10, r = 3, b2 = 0.5, c2 = 1.3)
## Create 10 genes with profile differences between Ctl, Tr2, and Tr3 groups
threediff <- tc.GENE(n = 10, r = 3, b3 = 0.8, c3 = -1, a4 = -0.1, b4 = -0.8, c4 = -1.2)
## Create 10 genes with profile differences between Ctl and Tr2 and different variance
vardiff <- tc.GENE(n = 10, r = 3, a3 = 0.7, b3 = 1, c3 = 1.2, var32 = 0.03, var33 = 0.03)
## Create dataset
tc.DATA <- rbind(flat, twodiff, threediff, vardiff)
rownames(tc.DATA) <- paste("feature", c(1:300), sep = "")
colnames(tc.DATA) <- paste("Array", c(1:36), sep = "")
tc.DATA [sample(c(1:(300*36)), 300)] <- NA  # introduce missing values

#### CREATE EXPERIMENTAL DESIGN
Time <- rep(c(rep(c(1:3), each = 3)), 4)
Replicates <- rep(c(1:12), each = 3)
Control <- c(rep(1, 9), rep(0, 27))
Treat1 <- c(rep(0, 9), rep(1, 9), rep(0, 18))
Treat2 <- c(rep(0, 18), rep(1, 9), rep(0,9))
Treat3 <- c(rep(0, 27), rep(1, 9))
edesign <- cbind(Time, Replicates, Control, Treat1, Treat2, Treat3)
rownames(edesign) <- paste("Array", c(1:36), sep = "")

## run T.fit from a p.vector object
tc.p <- p.vector(tc.DATA, design = make.design.matrix(edesign), Q = 0.01) 
tc.tstep <- T.fit(data = tc.p , alfa = 0.05)

## run T.fit from a data matrix and a design matrix
dise <- make.design.matrix(edesign)
tc.tstep <- T.fit (data = tc.DATA[271:300,], design = dise$dis, 
                   step.method = "two.ways.backward", min.obs = 10, alfa = 0.05)
tc.tstep$sol # gives the p.values of the significant 
             # regression coefficients of the optimized models