Two methods for improving performance of an HMM and their application for gene finding

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Hidden Markov models (HMMs) are widely used for gene finding, incorporating submodels for various genomic regions.
• Traditional methods often estimate HMM components independently, which may not be optimal for the overall model performance.
• The standard maximum likelihood estimation (MLE) criterion focuses on maximizing DNA sequence probability, not prediction accuracy.

Purpose of the Study:

• To develop an improved method for estimating HMMs for gene finding.
• To introduce a new training criterion that maximizes the probability of correct predictions.
• To enhance the performance of gene-finding algorithms like HMM-gene.

Main Methods:

• Estimating the entire HMM as a whole from labeled sequences, rather than independent sub-sequences.
• Utilizing a conditional maximum likelihood (CML) criterion to maximize prediction accuracy.
• Developing an approximate algorithm to find the most probable prediction summed over all paths yielding the same prediction.

Main Results:

• The proposed CML criterion is more effective for training gene-finding HMMs than standard MLE.
• The new approximate algorithm efficiently computes the most probable prediction.
• These methods significantly contribute to the high performance of the HMM-gene finder.

Conclusions:

• Training HMMs using conditional maximum likelihood enhances gene-finding accuracy.
• The developed algorithm and training strategy lead to substantial performance improvements in gene prediction.
• This approach offers a more effective way to build and train HMMs for genomic analysis.