hmm介绍.pptVIP

Hidden Markov Models (HMMs) Motivation: the CpG island problem Methylation in human genome “CG” - “TG” happens in most place except “start regions” of genes CpG islands = 100-1,000 bases before a gene starts Questions Q1: Given a short stretch of genomic sequence, how would we decide if it comes from a CpG island or not? Q2: Given a long sequence, how would we find the CpG islands in it? Answer to Q1: Bayes Classifier A Simple Model for Sequences:p(X) Answer to Q2: Hidden Markov Model HMM is just one way of modeling p(X,S)… A simple HMM How to “Generate” a Sequence? How to “Generate” a Sequence? HMM as a Probabilistic Model Three Problems Three Problems (cont.) Training – estimating parameters Supervised Have: model, marked data( data+states sequence) Want: parameters Unsupervised Have: model, data Want: parameters Problem I: Decoding/ParsingFinding the most likely path You can think of this as classification with all the paths as class labels… What’s the most likely path? Viterbi Algorithm: An Example Viterbi Algorithm Problem II: EvaluationComputing the data likelihood Another use of an HMM, e.g., as a generative model for discrimination Also related to Problem III – parameter estimation Data Likelihood: p(O|?) The Forward Algorithm Forward Algorithm: Example The Backward Algorithm Backward Algorithm: Example Problem III: TrainingEstimating Parameters Where do we get the probability values for all parameters? Supervised vs. Unsupervised Supervised Training Unsupervised Training Intuition Baum-Welch Algorithm Baum-Welch Algorithm (cont.) * (Lecture for CS397-CXZ Algorithms in Bioinformatics) Feb. 20, 2004 ChengXiang Zhai Department of Computer Science University of Illinois, Urbana-Champaign Hypothesis space: H={HCpG, HOther} Evidence: X=“ATCGTTC” Likelihood of evidence (Generative Model) Prior probability We need two generative models for sequences: p(X| HCpG), p(X|HOther) Probability rule Assume independence Capture som