MAResNet

MAResNet predicts transcription factor binding sites (TFBS) in DNA sequences using multi-scale bottom-up and top-down attention integrated within a residual network to improve prediction accuracy and interpretability.

Key Features:

• Multi-Scale Attention Mechanism: Implements a multi-scale attention mechanism at the initial stage to extract representative sequence features across different scales.
• Integration with Residual Network (ResNet): Stacks attention modules within a ResNet architecture to generate attention-aware features that capture complex TFBS patterns.
• Bottom-Up and Top-Down Attention: Combines bottom-up and top-down attention strategies to refine feature extraction and capture intricate sequence characteristics associated with transcription factor binding.
• Interpretability of Deep-Learning Models: Exposes attention-aware features learned across modules to aid interpretation of how predictions are formed as network depth increases.
• Visualization of Learned Features: Uses TMAP (TreeMap) to visualize learned features that contribute to TFBS prediction.

Scientific Applications:

• Gene Expression Analysis: Supports investigation of gene regulation by predicting TFBS that influence transcriptional control.
• Biological Development Studies: Facilitates study of developmental biology by identifying transcription factor binding patterns relevant to cellular differentiation and organismal development.
• Drug Design and Personalized Medicine: Informs drug-target identification and personalized medicine approaches by identifying TFBS relevant to therapeutic intervention.

Methodology:

MAResNet stacks attention modules within a residual network and applies multi-scale bottom-up and top-down attention to produce attention-aware features, visualizes learned features with TMAP, and was evaluated on 690 ChIP-seq datasets achieving an average AUC of 0.927.