GCBLANE
Graph-enhanced Convolutional BiLSTM Attention Network
A research project developing a novel deep learning model for Transcription Factor Binding Site (TFBS) prediction. Existing methods struggle with the complexity of DNA sequences; GCBLANE leverages Graph Neural Networks and Attention mechanisms to capture both local and global dependencies.
Technical Architecture
We developed GCBLANE (Graph-enhanced Convolutional BiLSTM Attention Network), a hybrid deep learning architecture that processes DNA sequences using one-hot encoding and de Bruijn graphs to capture sequential and structural dependencies. The architecture consists of:
- GNN Module: Graph Neural Network on de Bruijn graph representations of k-mers to model multi-hop structural relationships.
- CNN Module: Convolutional layers to extract local motifs from encoded sequences.
- Multi-Head Attention Module: Self-Attention applied to focus on important TF binding motifs.
- BiLSTM Module: Bidirectional LSTM layers to capture forward and backward contextual dependencies in sequences.
[REF_IMG]: GCBLANE ArchitectureResults
The below tables demonstrate GCBLANE's superior performance compared to various other state-of-the-art methods across different dataset sizes and feature types.
| Model | All | Small | Medium | Large |
|---|---|---|---|---|
| GCBLANE | 0.943 | 0.904 | 0.930 | 0.973 |
| MSDenseNet | 0.933 | 0.897 | 0.921 | 0.973 |
| MAResNet | 0.927 | 0.883 | 0.914 | 0.972 |
| SAResNet | 0.920 | 0.876 | 0.907 | 0.966 |
| Model | Accuracy | ROC AUC | PR AUC | Input Type |
|---|---|---|---|---|
| GCBLANE | 0.887 | 0.9495 | 0.949 | Sequence Only |
| BERT-TFBS | 0.851 | 0.919 | 0.920 | Sequence Only |
| TBCA | 0.823 | 0.894 | 0.899 | Sequence & Shape |
| DSAC | 0.816 | 0.887 | 0.883 | Sequence & Shape |
| DeepSTF | 0.814 | 0.883 | 0.890 | Sequence & Shape |