Dbert2_LR: A deep learning-based model for predicting cis-regulatory elements in crops.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cis-regulatory elements (CREs) control gene expression and are vital for agronomic traits.
• Identifying CREs in large, repetitive crop genomes like cotton is difficult.

Purpose of the Study:

• To develop a high-accuracy deep learning framework for identifying CREs in crop genomes.
• To improve the functional annotation of complex plant genomes.

Main Methods:

• Developed Dbert2_LR, a hybrid deep learning model integrating DNABERT-2 with RNN and LSTM networks.
• Applied the model to Arabidopsis thaliana and upland cotton for CRE classification.
• Conducted in-silico saturation mutagenesis (ISM) for model interpretability.

Main Results:

• Dbert2_LR achieved high accuracy in classifying promoters, enhancers, and non-regulatory sequences.
• Outperformed benchmark models with macro-averaged F1 scores of 0.890 (Arabidopsis) and 0.637 (cotton).
• ISM analysis confirmed biological interpretability, linking predictions to known transcription factor binding motifs.

Conclusions:

• Dbert2_LR is a powerful tool for functional annotation of crop genomes.
• The study facilitates CRE-based molecular breeding design for improved crop traits.