Genome annotation across species using deep convolutional neural networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Deep neural networks, particularly convolutional neural networks (CNNs), are increasingly used in genomics.
• CNNs identify functional genomic sequences but struggle with genome-wide application due to differing positive/negative example ratios in training vs. whole-genome data.

Purpose of the Study:

• To assess the genome-wide performance of CNNs trained with varying positive/negative example ratios.
• To investigate the cross-species applicability of CNNs for gene start site prediction.

Main Methods:

• Trained CNNs using gene start sequences (RefGene) as positive examples and random sequences as negative examples.
• Evaluated model performance across different ratios of positive to negative training data.
• Tested the ability of trained models to predict gene start sites in a related species.

Main Results:

• CNN performance on whole genomes is sensitive to the ratio of positive to negative examples in the training set.
• Models trained with specific data ratios demonstrate effective genome-wide prediction.
• Successfully predicted gene start sites in a related species, indicating cross-species applicability.

Conclusions:

• Optimizing training data ratios is crucial for robust genome-wide CNN performance in genomics.
• CNNs trained on one species can annotate genomes of related species, facilitating large-scale genomic analysis.
• This approach aids in identifying conserved sequence motifs recognized by chromatin-associated proteins across species.