蛋白质二级结构预测概率图模型的改进

蛋白质二级结构与蛋白质三级结构及蛋白质功能密切相关,是生物信息学研究的热点,其中概率图模型隐马尔可夫算法(HMM)是该领域研究的重要工具。但是在实际应用中,存在着HMM训练下溢、不同训练集的效果差异较大及参数优化困难等问题。对预测蛋白质二级结构时HMM遇到的训练下溢问题提出了改进方案;首次提出8-状态HMM来预测蛋白质二级结构,并且将参数B改进成为包含状态转移信息的三维参数;为了改进最优HMM模型的确定方法,用每个样本分别对初始HMM模型进行训练,得到一系列新的模型,然后对这些新模型的参数求均值,将求得的均值作为最优模型的参数。这些改进方法提高了HMM预测蛋白质二级结构的准确率,为HMM的进一步优化打下良好的基础。

Improved probability graph model for protein secondary structure prediction

Protein secondary structure is closely related to protein tertiary structure and function, and became a hot topic in bioinformatics. The probability graph model HMM (Hidden Markov model) is an important tool in this field. In practice, there exist problems such as: HMM training underflow, significant result differences derived from different training set, and hard process of parameter optimization. In this paper, aiming at HMM training underflow problem when predicting protein secondary structure, we put forward a method for solving the underflow problem; propose an 8-state HMM model to predict protein secondary structure for the first time; and modify parameter to be a three-dimensional parameter containing the state transition information. In order to improve the method drilling the optimal HMM, we train the initial HMM model with each sample, and get a series of new models; then average the parameters of the new models, and the obtained average parameter values are used to construct the optimal HMM model. The improved method increases the accuracy of protein secondary structure prediction, hence it is a good foundation for further improvement of HMM.