A Hybrid Machine Learning Approach To the Promoter Prediction Problem.
- Author:
Doo Heon SONG
1
Author Information
1. Department of Computer Software, Yongin Song Dam College. dsong@ysc.ac.kr
- Publication Type:Original Article
- Keywords:
Promoter;
Hybrid Learning Model;
False Positive;
Weight Matrix;
Hidden Markov Model
- MeSH:
Machine Learning*;
Bias (Epidemiology);
Computational Biology;
Data Mining;
Learning
- From:Journal of Korean Society of Medical Informatics
2003;9(4):391-399
- CountryRepublic of Korea
- Language:Korean
-
Abstract:
With recent explosive growth of bulky biological data available, there are great needs of developing rapid autonomous algorithms in bioinformatics. In result, there has be en a great deal of attempts to apply various data mining techniques and learning algorithms to various fields of bioinformatics and a good example of this trend is the promoter and motif search area to which NN (Neural Network), HMM (Hidden Markov Model), and clustering algorithms have been applied and several good public software programs are available. Learning algorithms explore a part of big learning space effectively by their own biases. Thus, in many occasions, different learning algorithms have radically different results especially when the target concept is uncertain or stochastically defined and/or the background knowledge of the problem is limited. In this case, it is useful to apply a hybrid learning approach which two or more mutually compensative algorithms (e.g. a low false positive algorithm and a low false negative algorithm) are effectively combined. In this paper, we report a series of experiments with a hybrid learning approach in promoter prediction problem. Three available public software systems are tested and two of them (McPromoter and PROMOTER SCAN) are hierarchically combined and tested. The result shows that the hybrid learning model in this problem is quite plausible (better than any of the two base systems in accuracy and low false alarms) and many other learning algorithms could be more useful in this approach than being applied independently.
