EBHMMNBfunForMulti: Baum-Welch algorithm for multiple hidden markov chains

In EBSeqHMM: Bayesian analysis for identifying gene or isoform expression changes in ordered RNA-seq experiments

Description

Baum-Welch algorithm for multiple hidden markov chains

Usage

EBHMMNBfunForMulti(Data,PPIn,
	NgVector=NULL,Conditions, sizeFactors,
	PriorFC=1.5,homo=TRUE, maxround=5,
	Pi0=NULL, Tran=NULL, NumTranStage=3,
	FCParam=NULL, AlphaIn=NULL,BetaIn=NULL,
	StateNames=c("Up","NC","Down"),
	EM=TRUE, UpdateParam=TRUE, Print=TRUE,WithinCondR=TRUE,
	PenalizeLowMed=TRUE, PenalizeLowMedQt=.2,PenalizeLowMedVal=10)

Arguments

Data	input data, rows are genes/isoforms and columns are samples
PPIn	PPDE for all adjacent comparisons
NgVector	Ng vector; NULL for gene level data
Conditions	A factor indicates the condition (time/spatial point) which each sample belongs to.
sizeFactors	a vector indicates library size factors
Tran	initial values for transition matrices
Pi0	initial values for starting probabilities
NumTranStage	number of states in two chains
PriorFC	target FC for gridient change
StateNames	name of the hidden states
homo	whether the chain is assumed to be homogenious
maxround	max number of iteration
AlphaIn,BetaIn	If the parameters are known and the user doesn't want to estimate them from the data, user may specify them here.
FCParam	not in use
EM	Whether estimate the prior parameters of the beta distribution by EM
UpdateParam	Whether update starting probabilities and transition probabilities
WithinCondR	By defining WithinCondR=TRUE, estimation of r's are obtained within each condition. (WithinCondR=FALSE is not suggested here)
Print	Whether print the elapsed-time while running the test.
PenalizeLowMed,PenalizeLowMedQt,PenalizeLowMedVal	Transcripts with median quantile < = PenalizeLowMedQt will be penalized

Details

EBHMMNBfunForMulti() function implements the Balm-Welch algorithm that estimates the starting probabilities and transition probabilities of a hidden Markov model with multiple chains. Here the emission distribution of each gene is assumed to be a Beta-Negative Binomial distribution with parameters (r_g, alpha, beta) , in which alpha and beta are shared by all the genes and r_g is gene specific. If not specified, r_g, alpha and beta will be estimated using method of moments. For isoform data, we assume isoforms from the same Ig group share the same beta^Ig. alpha is shared by all the isoforms and r_gi is isoform specific. The user also needs to specify an expected FC.

Value

MAPTerm: the most likely path of each gene/isoform.

MAPTermNum: numeric version of MAPTerm.

AllTerm: all possible expression paths considered in the model.

PP: posterior probability of being each expression path.

WhichMax: index of the most likely path.

Allf: prior probability of being each path.

Pi0Track: estimated starting probabilities of each iteration.

TranTrack: estimated transition probabilities of each iteration.

AlphaTrack, BetaTrack: estimated alpha and beta(s).

LLAll=PostSumForLL.Sum: log likelihood of the model.

Author(s)

Ning Leng

Examples

data(GeneExampleData)
CondVector <- rep(paste("t",1:5,sep=""),each=3)
Conditions <- factor(CondVector, levels=c("t1","t2","t3","t4","t5"))
Sizes <- MedianNorm(GeneExampleData)
tmp <- EBHMMNBfunForMulti(Data=GeneExampleData, PPIn=matrix(1,ncol=15, nrow=100),sizeFactors=Sizes, Conditions=Conditions,
          maxround=2)

EBSeqHMM documentation built on Nov. 8, 2020, 5:22 p.m.