DeepTSS: multi-branch convolutional neural network for transcription start site identification from CAGE data

Area of Science:

• Genomics
• Molecular Biology
• Bioinformatics

Background:

• Cap Analysis of Gene Expression (CAGE) is crucial for understanding transcription but is affected by transcriptional and technical noise.
• Noisy CAGE data leads to inaccurate transcription start site (TSS) annotation and quantification of regulatory regions.
• Computational methods are needed to improve the signal-to-noise ratio in CAGE data.

Purpose of the Study:

• To develop DeepTSS, a novel computational method for processing CAGE samples.
• To accurately predict single-nucleotide TSS using a combination of genomic signal processing, DNA features, evolutionary conservation, and deep learning.
• To enhance the signal-to-noise ratio in CAGE data for improved downstream analyses.

Main Methods:

• DeepTSS integrates genomic signal processing (GSP), structural DNA features, evolutionary conservation, and raw DNA sequence.
• Deep learning (DL) models, specifically convolutional layers, are employed for pattern identification and classification.
• The method processes CAGE samples to provide single-nucleotide TSS predictions.

Main Results:

• DeepTSS demonstrated superior performance compared to existing algorithms in CAGE data processing.
• Achieved 98% precision and 96% sensitivity (95.4% accuracy) for protein-coding gene TSS prediction.
• Positive predictions showed high overlap with active chromatin (96.66%), transcription factors (98.27%), and H3K4me3 peaks (92.04%).

Conclusions:

• DeepTSS effectively removes biological and technical noise from CAGE data, overcoming limitations of the experimental protocol.
• The DL-based approach eliminates the need for manual feature selection, as convolutional layers automatically identify relevant patterns.
• DeepTSS enhances the utility of CAGE, advancing research in coding and non-coding gene expression regulation.