DNA promoter task-oriented dictionary mining and prediction model based on natural language technology.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Promoters are crucial DNA sequences regulating gene expression and are vital for understanding gene regulatory networks.
• Accurate promoter identification is essential for deciphering gene expression patterns.
• Current deep learning methods for promoter prediction, including BERT, face limitations due to arbitrary DNA sequence segmentation during pre-training.

Purpose of the Study:

• To develop a novel DNA sequence segmentation method to improve promoter prediction accuracy.
• To introduce a refined dictionary for DNA sequence segmentation and BERT pre-training.
• To enhance the interpretability of deep learning models in DNA sequence analysis.

Main Methods:

• Developed a novel DNA sequence segmentation approach with a refined dictionary.
• Utilized the refined dictionary for BERT pre-training.
• Employed a BERT-Inception architecture integrating BERT with an Inception neural network for multi-granularity information capture.

Main Results:

• The BERT-Inception model demonstrated improved performance on several downstream tasks related to promoter identification.
• The novel segmentation method and architecture enhance the ability to capture information across multiple granularities in DNA sequences.
• The study introduces deep learning interpretability, offering new insights into DNA sequence information.

Conclusions:

• The proposed BERT-Inception architecture with novel DNA sequence segmentation significantly advances promoter prediction.
• This approach provides a more refined way to pre-train BERT models for genomic sequence analysis.
• The enhanced interpretability offers valuable perspectives for understanding DNA sequence information and gene regulation.