Extending hidden Markov models to allow conditioning on previous observations.

Area of Science:

• Computational molecular biology
• Bioinformatics
• Machine learning in biology

Background:

• Standard Hidden Markov Models (HMMs) rely on the Markovian assumption, where current observations depend only on the current state.
• This assumption may limit accuracy in biological sequence analysis, necessitating extensions to incorporate historical data.

Purpose of the Study:

• To develop and evaluate a simple extension of HMMs that accounts for previous observations.
• To improve the predictive performance of HMMs in computational biology tasks by relaxing the first-order Markovian constraint.

Main Methods:

• Implemented an extended HMM by transforming the observation sequence using an extended alphabet.
• This transformation allows the use of existing HMM training and decoding algorithms.
• Investigated various encoding schemes and applied the method to predict transmembrane proteins and signal peptides.

Main Results:

• The extended HMM approach demonstrated performance improvements of 1.8%-8.2% on biological prediction tasks when sufficient data were available.
• Significant performance gains were observed for specific prediction methods like PRED-TMBB2, PRED-TAT, and HMM-TM.
• The enhanced methods are accessible via web servers for academic users.

Conclusions:

• The proposed simple extension of HMMs effectively incorporates past observation information, enhancing predictive accuracy in computational biology.
• This approach offers a practical way to improve existing HMM-based prediction tools for biological sequences.
• The availability of these improved methods as web servers facilitates their adoption in biological research.