layerUMAP

layerUMAP visualizes hidden-layer representations from deep learning models to aid interpretation of biological sequence classification.

Key Features:

• Visualization of Hidden Layer Outputs: Extracts and projects outputs from hidden layers at multiple depths to reveal intermediate representations learned by models.
• UMAP-based Dimensionality Reduction: Applies UMAP (Uniform Manifold Approximation and Projection) to reduce high-dimensional learned representations while preserving data structure.
• Integration with autoBioSeqpy: Operates on representations produced by autoBioSeqpy deep learning workflows for biological sequence analysis.

Scientific Applications:

• Model interpretation for biological sequence classification: Enables inspection of layer-wise representations to interpret how deep learning models perform classification on biological sequences.
• Representation analysis and model diagnosis: Supports analysis of alternative representations and identification of layer-specific model behavior and potential areas for improvement.

Methodology:

Extracts outputs from hidden layers across model depths and integrates these learned representations with UMAP for dimensionality reduction; processes representations produced by autoBioSeqpy.