Modulation of DNA-protein Interactions by Proximal Genetic Elements as Uncovered by Interpretable Deep Learning

Area of Science:

• Genomics
• Molecular Biology
• Bioinformatics

Background:

• Transcription factors (TF) regulate cellular machinery by binding to specific DNA motifs.
• TF-DNA interactions exhibit high selectivity, despite numerous potential binding sites and accessible genomic regions.

Purpose of the Study:

• To develop a deep-learning framework for identifying and characterizing genetic elements that confer selectivity to TF-DNA binding.
• To analyze the role of upstream and downstream sequence elements in modulating TF-DNA interactions.

Main Methods:

• An interpretable recurrent neural network architecture was developed to analyze sequence context features.
• The framework was applied to model 26 transcription factors, scoring TF-DNA binding at base-pair resolution.

Main Results:

• Significant differences in DNA context feature activation were observed between bound and unbound sequences.
• The framework provided interpretability, enabling the identification and annotation of DNA elements modulating TF-DNA binding.
• Data processing significantly impacts model performance.

Conclusions:

• The deep-learning framework offers novel insights into non-coding genetic elements influencing TF-DNA interactions.
• This approach advances the understanding of gene regulation mechanisms by dissecting sequence-context-dependent TF binding selectivity.