DeepSplice

DeepSplice classifies candidate splice junction sequences from RNA-seq data using convolutional neural networks to reduce false positives and improve splice junction identification for transcriptome analysis.

Key Features:

• Deep Learning Framework: Utilizes convolutional neural networks (CNNs) for classification of splice junction sequences.
• Model Training Data: Model inferred from sequences of annotated exon junctions.
• Benchmark Performance: Demonstrates superior performance on the HS3D benchmark dataset.
• Validation Against Annotations: Achieves high accuracy when validated against GENCODE annotations.
• Large-scale Application and Candidate Reduction: Applied to Rail-RNA alignments of 21,504 human RNA-seq datasets, reducing approximately 43 million putative splice junction candidates to around 3 million high-confidence novel junctions.

Scientific Applications:

• Splice Junction Refinement: Refines identification of splice junctions in large-scale RNA-seq datasets by classifying candidate junctions.
• False Positive Reduction: Reduces false positives from ab initio aligners to improve downstream splice variant discovery and abundance estimation.
• Novel Junction Discovery: Enables discovery of high-confidence novel splice junctions in human transcriptomes.

Methodology:

Model inferred from sequences of annotated exon junctions and implemented as a convolutional neural network that classifies candidate splice junctions derived from primary RNA-seq data; evaluated on the HS3D benchmark and validated against GENCODE annotations; applied to Rail-RNA alignments of 21,504 human RNA-seq datasets, reducing approximately 43 million candidates to around 3 million high-confidence junctions.