剖面隐马氏模型训练算法的改进  

An Improvement on Training Algorithm of Profile Hidden Markov Model

作　　者：王斌宾[1] 史定华[1] 

机构地区：[1]上海大学理学院数学系,上海200444

出　　处：《应用数学与计算数学学报》2006年第1期26-32,共7页Communication on Applied Mathematics and Computation

基　　金：国家高技术研究发展计划(863计划)专项经费资助(课题编号:2002AA234021)

摘　　要：利用剖面隐马氏模型获得多序列联配,一般需要经过初始化、训练、联配三个过程．然而,目前广泛采用的Baum-Welch训练算法假设各条可观察序列互相独立,这与实际情况有所不符．本文对剖面隐马氏模型,给出可观察序列在互相不独立情况下的改进Baum-Welch算法,在可观察序列两种特殊情况下(互相独立和一致依赖),得到了改进算法的具体表达式,讨论了一般情况下权重的选取方法．最后通过一个具体的蛋白质家族的多序列联配来说明改进算法的效果．When using Profile Hidden Markov Model （PHMM） to obtain multiple sequence align- ment, we usually need initialization, training and alignment. However, the well-known Baum-Welch training algorithm assumes that all observable sequences are mutually independent. It may not hold in many cases. This paper presents an improving training algorithm of PHMM without the assumption of sequence independence. We obtain the whole expression of improved algorithm in two special cases of mutually independence and uniform dependence, and discuss choosing the weights in a general case. Finally we use multiple sequence alignment of a protein family to show the effect of the improved algorithm.

关 键 词：剖面隐马氏模型  Baum—Welch算法 多序列联配 可观察序列的相依性 

分 类 号：O211.62[理学—概率论与数理统计]