Semi-supervised Learning of Promoter Sequences Based on EM Algorithm

The eukaryotic promoter prediction is one of the most important problems in DNA sequence analysis. Promoter is a short sub-sequence before a transcriptional start site(TSS) in a DNA sequence. The prediction of the position of a promoter may approximately describe the position of a TSS, and gives help to biology experiments. Most proposed prediction algorithms are based on some search strategies, such as search by signal, search by content or search by CpG island, their performances are still limited by low sensitivities and high false positives. The promoter classification algorithm based on Markov chain has been proved to be effective in promoter prediction, where parameters such as transition probabilities are calculated by statistics on the labeled samples. In this paper, semi-supervised learning is introduced in promoter sequence analysis to improve classification accuracy with a combination of labeled and unlabeled sequences, and the maximum likelihood estimation formulas for transition probabilities are deduced. In simulating experiments, each long genomic sequence is truncated to short segments, which are mixed with labeled data, and classified according to the calculated probabilities. Comparison with some known prediction algorithms show that semi-supervised learning of promoter sequences based on EM algorithm is efficient when the number of labeled data is small, and the value of F\-1 is higher than that of predictions based on labeled samples.