Interpretable prediction of mRNA abundance from promoter sequence using contextual regression models

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Machine learning models predict gene expression from DNA promoter sequences.
• Interpreting these models and understanding DNA motif grammar (cooperation, distance constraints) remains challenging.
• Existing interpretation methods are often slow or fail to capture complex combinatorial rules.

Purpose of the Study:

• To develop interpretable neural network models for predicting mRNA expression levels from DNA sequences.
• To uncover DNA motif grammar and regulatory insights from promoter sequences.
• To address limitations of previous interpretation approaches in machine learning for genomics.

Main Methods:

• Designed interpretable neural network models for gene expression prediction.
• Applied a novel Contextual Regression framework to extract weighted features.
• Clustered samples based on gene expression levels for motif analysis.
• Analyzed motif co-occurrence and locations to identify regulatory grammars.

Main Results:

• Extracted weighted features to cluster samples with distinct gene expression levels.
• Identified motifs exhibiting active or repressive regulation of gene expression within clusters.
• Uncovered grammars of motif combination, including cooperative motif communities.
• Discovered specific distance constraints between motif pairs.

Conclusions:

• The developed interpretable neural network models offer new insights into promoter sequence regulatory architecture.
• The Contextual Regression framework facilitates the discovery of motif cooperation and distance constraints.
• This approach enhances understanding of how DNA sequence motifs collectively influence gene expression.