Interpreting cis-regulatory interactions from large-scale deep neural networks.

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Deep neural networks (DNNs) are increasingly used for predicting gene expression from DNA sequences.
• Evaluating and interpreting these large-scale genomic DNNs remains challenging.
• Current interpretation methods, like motif analysis, struggle with complex sequence interactions.

Purpose of the Study:

• To introduce cis-regulatory element model explanations (CREME), a novel toolkit for interpreting genomic DNNs.
• To understand the gene regulatory rules learned by DNNs.
• To provide high-resolution insights into genome regulatory architecture.

Main Methods:

• Developed CREME, an in silico perturbation toolkit for analyzing genomic DNNs.
• Applied CREME to Enformer, a state-of-the-art DNN for gene expression prediction.
• Analyzed cis-regulatory elements and their functional sequence elements.

Main Results:

• CREME successfully identified cis-regulatory elements that enhance or silence gene expression.
• Characterized complex interactions between cis-regulatory elements.
• Provided multi-scale interpretations of genomic organization, from elements to fine-mapped sequences.

Conclusions:

• CREME offers a powerful approach to translate DNN predictions into mechanistic insights of gene regulation.
• The toolkit enables a deeper understanding of the regulatory architecture of the genome.
• Facilitates the interpretation of complex genomic DNNs across various scales.