Initial Generic Functions {cubfits}

Initial Generic Functions of Codon Usage Bias Fits

Description

Initial generic functions for model fitting/approximation/prediction of cubfits.

Usage

  init.function(model = .CF.CT$model[1],
                type.p = .CF.CT$type.p[1],
                type.Phi = .CF.CT$type.Phi[1],
                model.Phi = .CF.CT$model.Phi[1],
                init.Phi = .CF.CT$init.Phi[1],
                init.fit = .CF.CT$init.fit[1],
                parallel = .CF.CT$parallel[1],
                adaptive = .CF.CT$adaptive[1])

Arguments

model

main fitted model.

type.p

proposal method for hyper-parameters.

type.Phi

proposal method for Phi (true expression values).

model.Phi

prior of Phi.

init.Phi

initial methods for Phi.

init.fit

how is coefficient initialed in vglm() of VGAM.

parallel

parallel functions.

adaptive

method for adaptive MCMC.

Details

This function mainly takes the options, find the according generic functions, and assign those functions to .cubfitsEnv. Those generic functions can be executed accordingly later within functions for MCMC or multinomial logistic regression such as cubfits(), cubappr(), and cubpred(). By default, those options are provided by .CF.CT which also leaves rooms for extensions of more complicated models and further optimizations.

It is supposed to call this function before running any MCMC or multinomial logistic regression. This function may affect cubfits(), cubpred(), cubappr(), estimatePhi(), and fitMultinom().

• model is the main fitting model, currently only roc is fully supported.

• type.p is for proposing hyper-parameters in Gibb sampler. Currently, lognormal_fix is suggested where mean 1 is fixed for log normal distribution. Conjugated prior and flat prior exist and are easily available in this step

• type.Phi is for proposing Phi (expression values) in the random walk chain updates. Only, RW_Norm is supported. Usually, the acceptance ratio can be adapted within 25% and 50% controlled by .CF.AC if adaptive = simple.

• model.Phi is for the distribution of Phi. Typically, log normal distribution lognormal is assumed.

• init.Phi is a way to initial Phi. Posterior mean PM is recommended which avoid boundary values.

• init.fit is a way of initial coefficients to fit mutation and selection coefficients (\log\mu and \Delta t or \omega) in vglm(). Option current means the b (log(mu) and Delta.t) of current MCMC iteration is the initial values, while random means vglm() provides the initial values.

• parallel is a way of parallel methods to speed up code. lapply means lapply() is used and no parallel; mclapply means mclapply() of parallel is used and good for shared memory machines; task.pull means task.pull() of pbdMPI is used and good for heterogeneous machines; pbdLapply means pbdLapply() of pbdMPI is used and good for homogeneous machines. Among those, task.pull is tested thoroughly and is the most reliable and efficient method.

• adaptive is a way for adaptive MCMC that propose better mixing distributions for random walks of Phi. The simple method is suggested and only the proposal distribution of Phi (type.Phi = RW_Norm) is adjusted gradually.

Value

Return an invisible object which is a list contain all generic functions according to the input options. All functions are also assigned in the .cubfitsEnv for later evaluations called by MCMC or multinomial logistic regression.

Note

Note that all options are taken default values from the global control object .CF.CT, so one can utilize/alter the object's values to adjust those affected functions.

Note that phi.Obs should be scaled to mean 1 before applying to MCMC.

Author(s)

Wei-Chen Chen [e-mail].

References

https://github.com/snoweye/cubfits/

See Also

.CF.CT, .CF.CT, cubfits(), cubpred(), and cubappr().

Examples

## Not run: 
suppressMessages(library(cubfits, quietly = TRUE))
set.seed(1234)

# Convert data.
reu13.list <- convert.reu13.df.to.list(ex.test$reu13.df)
y.list <- convert.y.to.list(ex.test$y)
n.list <- convert.n.to.list(ex.test$n)

# Get phi.pred.Init
init.function(model = "roc")
fitlist <- fitMultinom(ex.train$reu13.df, ex.train$phi.Obs, ex.train$y,
                       ex.train$n)
phi.pred.Init <- estimatePhi(fitlist, reu13.list, y.list, n.list,
                         E.Phi = median(ex.test$phi.Obs),
                         lower.optim = min(ex.test$phi.Obs) * 0.9,
                         upper.optim = max(ex.test$phi.Obs) * 1.1)

## End(Not run)