Identifying Cancer Driver Genes Using a Neural Network Framework With Cross-Attention Mechanism

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying cancer driver genes is crucial for developing targeted therapies.
• Current deep learning methods struggle with weak feature representations due to ignored feature interconnections.
• Improved accuracy in driver gene identification can accelerate drug discovery and cancer treatment development.

Purpose of the Study:

• To propose a novel graph neural network framework, GTCM, for enhanced identification of cancer driver genes.
• To improve feature representations by effectively learning connections among different gene feature sets.
• To enhance the accuracy of cancer driver gene identification compared to existing methods.

Main Methods:

• Utilized a graph neural network framework (GTCM) integrating graph convolutional networks (GCN), Transformer with cross-attention, and a multi-layer perceptron (MLP) classifier.
• Employed GCN to learn gene feature representations from three distinct gene association networks.
• Applied Transformer with cross-attention to dynamically capture inter-feature relationships, enhancing representations for cancer driver genes.
• Used MLP for the final prediction of cancer driver genes.

Main Results:

• Ablation experiments confirmed that Transformer with cross-attention significantly improves GCN-learned feature representations.
• The proposed GTCM framework demonstrated improved identification rates for cancer driver genes.
• GTCM outperformed existing representative methods, achieving superior performance in Area Under the Receiver Operating Characteristic Curves (AUROCC) and Area Under Precision-Recall Curves (AUPRC).

Conclusions:

• The GTCM framework effectively enhances feature representations for cancer driver gene identification.
• GTCM offers a significant advancement in accurately identifying cancer driver genes, potentially accelerating therapeutic development.
• The proposed method shows promising results and outperforms current state-of-the-art approaches in driver gene identification.