Generalizations of Markov model to characterize biological sequences

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Standard kth order Markov models predict nucleotide probability based on k preceding units.
• These models do not account for joint dependencies among multiple nucleotides or long-range dependencies (gap > 0).

Purpose of the Study:

• To develop and evaluate generalized Markov models that capture joint and long-range dependencies in biological sequences.
• To assess if improved sequence classification accuracy can be achieved using alternative model parameters.

Main Methods:

• Developed a configurable Java tool to explore generalized Markov models.
• Evaluated model performance on CpG-poor promoters, all promoters, exons, and nucleosome positioning sequences.
• Compared classification accuracy using single nucleotide, dinucleotide, and trinucleotide units with varying gap lengths.

Main Results:

• Using dinucleotide and trinucleotide units significantly improved sequence classification accuracy compared to single nucleotide models.
• Optimal accuracy for nucleosome positioning sequences was observed at a gap length of 4.
• A periodicity of 10-11 base pairs was detected in nucleosome positioning sequences, suggesting potential structural preferences.

Conclusions:

• Extended standard Markov models to include joint and long-range dependencies.
• Generalized models are expected to enhance the accuracy of sequence analysis tools.
• The developed tool is available for download.