Abstract: Traditional method for training HMMs is Baum-Welch algorithm which is noted for the rapid convergence.However，this method can only lead to local optimization and effects quantity of sequence alignment.In response to these problems for training HMMs using Baum-Welch algorithm and integration of laws in biological genetics and evolution，a new HMMs training method based on BW-GA that combines traditional method and genetic algorithm is proposed.According to the need of genetic alignment and structure of HMMs，three genetic operations and coding methods are designed.HMMs is trained with 19 nuclear 5sRNA sequences by applying BW-GA，which makes the sequences generated have a higher log-likelihood of probability than just applying traditional methods，and a better multi-sequence alignment is generated.

Key words: multi-sequence alignment, hidden markov models, genetic algorithm, log likelihood probability

摘要： Baum-Welch算法是训练HMMs的传统方法，该方法虽然收敛速度快，但容易陷入局部最优，影响了序列比对的质量。针对该算法存在的问题，结合生物遗传与进化的规律，设计了一种将传统方法与遗传算法相结合训练HMMs的BW-GA方法。根据序列比对的需要和HMMs的结构，定义了3种遗传操作和编码方式。用19条原核5sRNA序列对模型进行了训练，用BW-GA训练模型产生序列的对数似然概率比单独用传统方法训练的要高，产生序列比对的质量较好。

关键词: 多序列比对, 隐马尔可夫模型, 遗传算法, 对数似然概率