Visualizing complex feature interactions and feature sharing in genomic deep neural networks.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Existing deep learning visualization tools often overlook feature combinations in intermediate layers.
• Current methods may not adequately reveal complex model decision-making processes by focusing on limited input examples.

Purpose of the Study:

• To introduce DeepResolve, a novel analysis framework for visualizing feature contributions and interactions in deep convolutional models of genome function.
• To enable a more comprehensive understanding of how low-level features are combined to influence model decisions.

Main Methods:

• DeepResolve employs gradient ascent to explore intermediate feature maps in deep convolutional neural networks.
• The framework visualizes individual and combinatorial contributions of input features to network decisions.
• It analyzes feature sharing across different genomic tasks to infer shared biological mechanisms.

Main Results:

• DeepResolve effectively visualizes feature contributions and interactions, identifying negative features and non-additive interactions.
• The framework recovers similarities between poorly correlated classes, offering insights missed by traditional methods.
• It demonstrates shared decision-making structures across various genome annotations (histone marks, DNase hypersensitivity, transcription factor binding) and identifies biologically relevant TF groups.

Conclusions:

• DeepResolve provides a powerful method for visualizing complex feature patterns and interactions in genomic deep convolutional networks.
• The framework uncovers feature sharing and class similarities that suggest novel biological mechanisms.
• DeepResolve complements existing visualization tools, offering unique and valuable insights into deep learning models for genomics.