HAKKE: A Multi-Strategy Prediction System for Sequences

Naohiro Furukawa [1] (furukawa@i.kyushu-u.ac.jp)
Satoshi Matsumoto [1] (matumoto@i.kyushu-u.ac.jp)
Ayumi Shinohara [1] (ayumi@i.kyushu-u.ac.jp)
Takayoshi Shoudai [1] (shoudai@i.kyushu-u.ac.jp)
Satoru Miyano [2] (miyano@ims.u-tokyo.ac.jp)

[1] Department of Informatics,
Graduate School of Information Science and Electrical Engineering,
Kyushu University
6-10-1 Hakozaki, Higashi-ku, Fukuoka 812-81 Japan
[2] Human Genome Center, Institute of Medical Science,
University of Tokyo
4-6-1 Shirokanedai, Minato-ku, Tokyo 108 Japan

We developed a machine learning system HAKKE which is suitable for predicting functional regions from sequences, such as protein-coding region prediction, and transmembrane domain prediction. HAKKE is a hybrid system cooperated by a number of algorithms of a pool to make an accurate prediction. The system uses an extension of the weighted majority algorithm in order to fit the strength of each algorithm into given training examples. In this paper, we describe the core of the system and show some experimental results on transmembrane domain and alpha-helix predictions.