An interpretable block-attention network for identifying regulatory feature interactions

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Regulatory features are crucial in understanding health and disease.
• Self-attention networks (SANs) show promise for complex biological predictions.
• Existing SANs face challenges in biological applications due to high memory usage and lack of interpretability.

Purpose of the Study:

• To develop a deep learning model, ISANREG, that overcomes SAN limitations for biological regulatory interaction prediction.
• To enhance the interpretability of self-attention mechanisms in biological sequence analysis.
• To provide a framework for predicting transcription factor-bound motif instances and DNA-mediated TF-TF interactions.

Main Methods:

• Implementation of an Interpretable Self-Attention Network for REGulatory interactions (ISANREG).
• Integration of block self-attention and attention-attribution mechanisms.
• Utilizing self-attention attribution scores for prediction and interpretation at single-nucleotide resolution.

Main Results:

• ISANREG successfully predicts transcription factor-bound motif instances.
• The model accurately identifies DNA-mediated TF-TF interactions.
• ISANREG provides interpretable self-attention scores, offering insights into regulatory element contributions.

Conclusions:

• ISANREG offers an interpretable and efficient deep learning framework for biological regulatory analysis.
• The model addresses key limitations of traditional SANs in computational biology.
• ISANREG serves as a foundation for future interpretable deep learning models in genomics and related fields.