DeeProPre

DeeProPre identifies promoter regions in eukaryotic genomes using deep learning to classify promoters (TATA-box, without TATA-box, mixed) and support studies of promoter structure and gene expression regulation.

Key Features:

• Supervised embedding: Maps DNA sequences into a high-dimensional space via a supervised embedding layer to capture complex sequence relationships.
• Convolutional neural networks (CNNs): Two 1D convolutional layers extract local sequence patterns relevant to promoter signals.
• Bidirectional LSTM (BiLSTM): BiLSTM models capture both upstream and downstream sequence context around transcription start sites (TSS).
• Attention mechanism: Attention weights focus the model on relevant subsequences to improve interpretability and prediction accuracy.
• Sigmoid-activated classifier: A fully connected layer with Sigmoid activation outputs probabilities for promoter classes.
• Promoter categories: Produces classifications for promoters with TATA-box, without TATA-box, and mixed data types.
• Target regions: Detects promoter regions located upstream or at the 5' end of transcription start sites in eukaryotic genomes.

Scientific Applications:

• Promoter classification: Distinguishes TATA-box, non-TATA-box, and mixed promoters to improve promoter annotation in eukaryotic genomes.
• Genome annotation: Supports identification of promoter locations upstream or at the 5' end of TSS for eukaryotic gene models.
• Promoter structure-function studies: Enables investigation of promoter architecture and its influence on gene expression regulation and transcription mechanisms.

Methodology:

DeeProPre applies a supervised embedding layer to map sequences into a high-dimensional space, uses two 1D convolutional layers for local feature extraction, employs a bidirectional LSTM for contextual modeling, incorporates an attention mechanism to weight relevant positions, and uses a fully connected Sigmoid-activated output layer to produce promoter class probabilities.