LangMoDHS: A deep learning language model for predicting DNase I hypersensitive sites in mouse genome

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• DNAse I hypersensitive sites (DHSs) are key genomic regions for identifying cis-regulatory elements.
• Existing methods for DHS detection face practical limitations, creating a need for improved approaches.

Purpose of the Study:

• To develop and validate a novel deep learning-based language model, LangMoDHS, for accurate prediction of DHSs.
• To assess the performance of LangMoDHS against state-of-the-art methods and analyze the contributions of its components.

Main Methods:

• LangMoDHS integrates convolutional neural networks (CNNs), bi-directional long short-term memory (Bi-LSTM), and attention mechanisms to process DNA sequences.
• The model was evaluated using 5-fold cross-validation and independent tests across 14 tissues and 4 developmental stages.
• Information entropy indices were employed to investigate sequence motifs within DHSs.

Main Results:

• LangMoDHS demonstrated competitive or superior performance compared to the latest DHS prediction method, iDHS-Deep.
• Empirical experiments highlighted the significant contributions of CNN, Bi-LSTM, and attention components to prediction accuracy.
• The analysis provided novel insights into the sequence characteristics of DHSs.

Conclusions:

• LangMoDHS represents an effective deep learning approach for predicting DHSs, addressing a gap in current methodologies.
• The model's architecture and analysis of sequence motifs offer valuable tools for genomic research.
• A user-friendly web server for LangMoDHS is publicly available for broader scientific application.