DeepTSS

DeepTSS predicts transcription start sites in Cap Analysis of Gene Expression (CAGE) data by integrating genomic signal processing, structural DNA features, evolutionary conservation, and a deep learning framework that combines raw DNA sequences with additional genomic features to reduce transcriptional and technical noise.

Key Features:

• Genomic Signal Processing (GSP): Applies GSP techniques to enhance the signal-to-noise ratio in CAGE datasets for more accurate TSS annotation and regulatory region quantification.
• Structural DNA Features: Incorporates DNA structural properties as features to provide biological context for TSS predictions.
• Evolutionary Conservation Evidence: Integrates conservation data to prioritize conserved genomic regions likely to be functionally relevant.
• Deep Learning Framework: Employs a deep learning architecture that combines raw DNA sequences and additional genomic features, using convolutional layers to automatically identify predictive patterns without manual feature selection.

Scientific Applications:

• Accurate TSS prediction: Reports 98% precision, 96% sensitivity, and 95.4% overall accuracy in protein-coding gene annotations.
• Regulatory region annotation: Positive predictions overlap 96.66% with active chromatin regions, 98.27% with transcription factor binding sites, and 92.04% with H3K4me3 peaks.
• Gene expression regulation studies: Enables investigation of coding and non-coding gene expression regulation by improving identification of true transcription initiation events.

Methodology:

Integrates genomic signal processing, structural DNA features, and evolutionary conservation within a deep learning model that combines raw DNA sequences with other genomic features using convolutional layers; evaluated using experimental data, protein-coding gene annotations, and computationally-derived genome segmentations by chromatin states.