SpliceVisuL

SpliceVisuL visualizes and interprets splice junction predictions from bidirectional long short-term memory (Bi-LSTM) networks to reveal sequence motifs underlying canonical and non-canonical splicing.

Key Features:

• Visualization Techniques: A range of visualization techniques infer sequence information learned by recurrent neural networks (RNNs) during splice junction identification on genomic sequences.
• Perturbation-Based Visualization: The method modifies input nucleotide sequences to observe changes in model predictions, enabling identification of canonical donor and acceptor junction motifs.
• Back-Propagation-Based Visualization: This technique leverages the neural network embedding space to detect non-canonical splicing motifs learned by the model.
• Single Nucleotide and Span Analysis: Provides inspection at both single-nucleotide resolution and spans of consecutive nucleotides to localize sequence features relevant to splice junctions.

Scientific Applications:

• Identify Canonical and Non-Canonical Splicing Motifs: Facilitates discovery of both standard and atypical splicing patterns that affect gene expression regulation.
• Enhance Interpretability of Neural Models: Provides visual insights into features learned by Bi-LSTM networks to support interpretation and validation of model predictions.

Methodology:

Bidirectional LSTM (Bi-LSTM) recurrent neural networks process nucleotide sequences to predict splice junctions; perturbation-based visualization modifies input sequences to measure prediction changes; back-propagation-based visualization inspects the network embedding space; techniques are adapted for genomic sequence characteristics and support single-nucleotide and span-level analyses.