Predicting Eukaryotic Promoter using Both Interpolated Markov Chains and Time-Delay Neural Networks

To improve eukaryotic polymerase II promoter recognition, in this paper we present a new approach by using methods from two already existing promoter prediction programs. Our approach is mainly based on interpolated Markov chains (IMC), stochastic segment models (SSM) and time-delay neural networks (TDNN). The former two are used by the promoter recognition system McPromoter and the last one is used by NNPP. The outputs of these methods were then used as inputs to a neural network, which established our new prediction model. We trained and tested our model separately on the human and drosophila promoter datasets collected by Martin Reese. The final predictor shows a 5-fold cross-validation true positive rate of 76% with false positive rate 2% on human dataset. The average improvement of true positive rates is above 5% with varying false positive rates on both data sets as compared to McPromoter and NNPP. Our study demonstrates that these three methods: IMC, SSM and TDNN, can contribute simultaneously to the promoter prediction problem in a single algorithm