MSDenseNet

MSDenseNet predicts DNA–protein binding from raw DNA sequences by combining multi-scale dense convolutional networks with a fault-tolerant coding scheme to improve accuracy in studies of gene expression regulation.

Key Features:

• Fault-Tolerant Coding Scheme: Transforms raw DNA sequences into fusion coding using fault-tolerant coding (FTC) that accounts for partial motif variations and sequencing errors from sequencing technologies.
• Multi-Scale Convolutional Architecture: Implements multi-scale convolutions within dense layers and preceding dense blocks to capture features at multiple scales and accelerate convergence during training.
• Dense Connectional Deep Neural Networks: Employs dense connectional deep neural network structures to mine complex relationships between internal attributes of fusion sequence features for DNA–protein binding prediction.

Scientific Applications:

• DNA–Protein Binding Prediction: Predicts DNA–protein binding (DPB) to support analysis of DNA-binding proteins and regulatory mechanisms underlying gene expression.
• ChIP-seq Benchmarking: Validated on 690 ChIP-seq datasets, achieving an average Area Under the Curve (AUC) of 0.933 and outperforming existing state-of-the-art methods.

Methodology:

Transforms raw DNA sequences into fusion coding via fault-tolerant coding (FTC), applies multi-scale convolutions within dense layers and preceding dense blocks, and uses dense-connection deep neural networks with accelerated training convergence.