SMG: self-supervised masked graph learning for cancer gene identification.

Area of Science:

• Cancer genomics and bioinformatics.
• Application of deep learning in molecular biology.

Background:

• Identifying cancer genes (CGs) is vital for understanding cancer progression.
• High-throughput data and deep learning offer new ways to model genomic data.
• Limited labeled data poses a challenge for pinpointing CGs.

Purpose of the Study:

• To propose a novel deep learning framework, self-supervised masked graph learning (SMG), for identifying cancer genes.
• To leverage multi-omic features within protein-protein interaction (PPI) networks.
• To address the challenge of limited labeled data in cancer genomics.

Main Methods:

• Developed a self-supervised masked graph learning (SMG) framework.
• Utilized a graph neural network (GNN)-based autoencoder for network reconstruction (pretext task).
• Employed task-specific fine-tuning for downstream predictions on node and graph levels.

Main Results:

• Demonstrated the superiority of the SMG method on node-level tasks (CG identification, essential genes, healthy driver genes) and a graph-level task (disease subnetwork identification).
• Benchmarking experiments across eight PPI networks confirmed SMG's effectiveness.
• Showcased SMG's advantage in multi-omic feature engineering compared to state-of-the-art methods.

Conclusions:

• The proposed SMG framework effectively identifies cancer genes and related biological entities.
• SMG offers a robust approach for analyzing complex genomic data with limited labels.
• This method advances the application of deep learning in cancer genomics research.