Context dependent prediction in DNA sequence using neural networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Understanding DNA sequence structure is crucial for biological insights.
• Predicting DNA sequences aids in deciphering genomic information.

Purpose of the Study:

• To develop and evaluate a model for predicting DNA sequences.
• To compare the performance of neural networks against traditional models like Markov models.

Main Methods:

• Training a neural network model to predict missing DNA bases using flanking sequence contexts.
• Utilizing likelihood-ratio tests to assess model performance.
• Analyzing prediction accuracy across different genomic regions and species.
• Employing Fourier transforms to identify periodic signals in predictions.

Main Results:

• The neural network achieved 54% accuracy on the human genome, outperforming Markov models by 2%.
• Performance was uniform across chromosomes, with higher accuracy in repetitive sequences (~70%) than coding regions (~40%).
• Periodic signals were detected, potentially linked to nucleosome positioning and GC/AT content, with similar signals observed in other large genomes.

Conclusions:

• Neural networks show significant potential for DNA sequence modeling, outperforming Markov models.
• Periodic signals in DNA sequences offer new avenues for research into genomic structure.
• Further exploration of neural network architectures could enhance DNA sequence prediction capabilities.