Probabilistic models for semisupervised discriminative motif discovery in DNA sequences

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Discriminative motif discovery methods identify transcription factor binding sites (TFBSs) by differentiating sequence sets.
• Generative models excel at using unlabeled sequences for motif detection, especially with limited labeled data.
• Hybrid approaches are needed to leverage the strengths of both generative and discriminative models.

Purpose of the Study:

• To develop a hybrid generative/discriminative model for semisupervised discriminative motif discovery.
• To enhance TFBS identification by incorporating unlabeled sequences into discriminative frameworks.
• To improve motif discovery performance, particularly when labeled training data is scarce.

Main Methods:

• Developed a hybrid model integrating generative and discriminative approaches.
• Implemented a semisupervised learning framework for motif discovery.
• Conducted numerical experiments using yeast ChIP-chip data for DNA motif discovery.

Main Results:

• The hybrid semisupervised model achieved superior performance compared to purely generative or discriminative methods.
• Semisupervised learning significantly improved motif discovery performance when labeled sequences were limited.
• Demonstrated enhanced sensitivity and specificity in TFBS identification.

Conclusions:

• Hybrid generative/discriminative models offer a powerful approach for semisupervised motif discovery.
• Utilizing unlabeled sequences via semisupervised learning boosts TFBS identification accuracy.
• The developed method provides a robust solution for motif discovery in genomics research.