基于GRNN人工神经网络的寡核苷酸解链温度预测

The prediction of oligonucleotide melting temperature by generalized regression of neural network

解链温度预测在引物和探针设计中具有重要的作用,本研究以384条寡核苷酸的解链温度数据为材料,随机分为训练集(279条)和测试集(69条)样本,利用训练集样本对建立的GRNN人工神经网络进行训练;再利用训练好的人工神经网络对测试集样本的解链温度进行预测,发现本研究所建立的GRNN人工神经网络的平均预测误差为2.44±0.98℃,最大误差为5.77℃,说明本研究建立的GRNN人工神经网络具有较好的预测性能,完全可以用于寡核苷酸解链温度的预测。同时比较了GRNN人工神经网络与目前常用的3种邻近法在预测寡核苷酸解链温度上的差异,发现Breslauer(1986)建立的预测方法误差较大,其平均误差为6.81±3.90℃,Santalucia(1996)建立的预测方法次之,平均误差为2.41±1.96℃,Sugimoto(1996)建立的预测方法最准确,其平均误差为1.57±0.96℃,分析了各种预测方法产生误差的原因,为今后开发新的寡核苷酸解链温度预测工具提供了新的思路和方法。

The prediction of melting temperature is important in designing the primer and probe.Total 384 oligonucleotides were random divided into two groups and the first group includes 279 oligonucleotides which were used to train the GRNN,another group includes 69 oligonucleotides which were used to test the performance of trained GRNN in this study.The average absolute error value of predictive melting temperature used GRNN was 2.44±0.98℃and the maximum error value was 5.77℃.This indicated that the GRNN had satisfactory predictability and could be used to predict the melting temperature of oligonucleotides.The performance of GRNN was compared with three common nearest neighbor methods in predicting the melting temperature of oligonucleotides and results showed the method constructed by Breslauer（1986） had larger error value and the average absolute error value was 6.81±3.90℃; the method constructed by Santalucia（1996） had moderate error value and the average absolute error value was 2.41±1.96℃;the method constructed by Sugimoto（1996） had less error value and the average absolute error value was 1.57±0.96℃.The main reasons of errors in these methods were analyzed and new potential method and way were provided to design new tool to predict melting temperature of oligonucleotides.