DSAC

DSAC predicts transcription factor binding sites (TFBSs) from DNA sequences using a dual-branch network that combines convolutional operations and self-attention to capture local and long-range sequence features.

Key Features:

• Dual-Branch Architecture: A dual-branch network integrates convolutional operations and self-attention mechanisms to combine local feature extraction with global dependency modeling.
• Comprehensive Feature Extraction: Interactive fusion of convolutional and self-attention representations enhances learning of detailed local features and long-distance dependencies within DNA sequences.
• Efficient Network Design: A lightweight, computationally efficient architecture leverages both convolutional and self-attention components without compromising predictive performance.
• Predictive Performance: Evaluation on 165 ChIP-seq datasets shows DSAC outperformed five other deep learning methods in predicting TFBSs from sequence alone.

Scientific Applications:

• TFBS Prediction: Identification of transcription factor binding sites from genomic sequence data to map regulatory elements.
• Transcriptional Regulation Studies: Analysis of transcriptional regulation mechanisms by locating binding sites that influence gene expression.
• ChIP-seq Benchmarking: Benchmarking predictive accuracy using ChIP-seq datasets to evaluate model performance in TFBS prediction.

Methodology:

A dual-branch neural network where a convolutional branch extracts local sequence motifs and a self-attention branch captures long-distance dependencies, with interactive fusion of their representations for TFBS prediction.