DeepBound

DeepBound identifies transcript boundaries from RNA sequencing (RNA-seq) read alignments using deep convolutional neural fields to improve transcript assembly and the discovery of novel genes and transcripts.

Key Features:

• Deep convolutional neural fields: Uses deep convolutional neural fields to learn hidden distributions and patterns of transcript boundaries.
• AUC-based optimization: Integrates the AUC (area under the curve) score into the optimization objective to enhance boundary prediction accuracy.
• Label-imbalance handling: Addresses label-imbalance issues inherent in boundary training data through the AUC-based objective.
• Simulated training data: Utilizes simulated RNA-seq datasets for training to provide large labeled sample sets for the deep probabilistic graphical model.
• Noisy/weak signal targeting: Specifically tackles noisy and weak signals at transcript boundaries that complicate accurate determination from RNA-seq reads.
• Complementary to spliced-read junction detection: Focuses on boundary identification distinct from splicing junction detection, which is addressed by spliced reads.
• Empirical performance: Demonstrated superior performance compared to existing methods on simulation datasets from two species and on biological datasets.

Scientific Applications:

• Transcript assembly: Improves full-length transcript assembly by providing more accurate boundary locations from RNA-seq data.
• Novel gene and transcript discovery: Enables discovery of novel genes and transcripts through enhanced boundary detection.
• Gene expression and functional studies: Facilitates more complete transcript models for analysis of gene expression patterns and gene function.
• Method benchmarking: Serves as a benchmark for evaluating transcript boundary identification methods on simulated and biological datasets.

Methodology:

Implements deep convolutional neural fields trained on simulated RNA-seq datasets with an AUC (area under the curve)-based optimization objective to learn hidden distributions of transcript boundaries and address label-imbalance.