Predictive modeling of gene expression and localization of DNA binding site using deep convolutional neural networks

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Genomic sequencing has advanced, yet the arrangement of transcription factor binding sites on regulatory DNA remains largely unknown.
• Massively Parallel Reporter Assays (MPRAs) offer a way to measure gene expression driven by thousands of DNA variants, but current analysis often assumes independent base pair contributions.
• Understanding these interactions is crucial for deciphering transcriptional control.

Purpose of the Study:

• To develop a novel computational framework for analyzing MPRA data that accounts for correlations between distant bases in regulatory DNA.
• To create a deep learning model capable of predicting gene expression levels from raw DNA sequences and identifying transcription factor binding sites at single-base pair resolution.

Main Methods:

• Developed the Deep learning Adaptable Regulatory Sequence Identifier (DARSI), a convolutional neural network.
• Trained DARSI using MPRA data to predict gene expression levels directly from regulatory DNA sequences.
• Validated DARSI's predictions by benchmarking against curated databases of known transcription factor binding sites.

Main Results:

• DARSI accurately predicts known transcription factor binding sites.
• DARSI identifies novel, unmapped binding sites within regulatory regions.
• The model demonstrates high accuracy in predicting gene expression from DNA sequences.

Conclusions:

• DARSI provides a new framework for MPRA data analysis, enabling the identification of transcription factor binding sites with single-base pair resolution.
• The tool generates experimentally actionable predictions, facilitating the theory-experiment cycle for a predictive understanding of transcriptional control.
• DARSI's ability to predict novel binding sites opens avenues for future experimental validation and discovery of transcription factors.