TL;DR
This paper introduces SMG, a novel self-supervised learning framework that leverages multi-omic data and protein-protein interaction networks to identify cancer genes (CGs). By masking and reconstructing graph structures, SMG effectively learns gene relationships, outperforming existing methods on various cancer gene identification tasks, even with limited labeled data.
What Problem Does This Paper Address?
Identifying cancer genes (CGs) is crucial for understanding cancer progression and developing targeted therapies. However, the sheer volume of genomic data and the complex, interconnected nature of biological systems make pinpointing these genes a significant challenge. Traditional methods often struggle with limited labeled data and fail to fully capture the intricate relationships within biological networks. While previous approaches have used protein-protein interaction (PPI) networks and multi-omic data, they often fall short in effectively utilizing unlabeled data or fully exploiting the rich topological information within these networks. This paper aims to address this gap by proposing a self-supervised learning method that can learn robust gene representations from unlabeled data, thereby improving CG identification accuracy.
Methods
The core of the proposed method, SMG (self-supervised masked graph learning), consists of two main stages:
- SMG Reconstruction (Pretext Task): This stage employs a graph neural network (GNN)-based autoencoder. To train it, nodes in the multi-omic featured PPI networks are randomly masked. The GNN autoencoder then learns to reconstruct these masked nodes by leveraging the information from their neighbors. This process allows the model to capture intricate interaction relationships and topological information within the network without relying on labeled data. For node classification tasks, a GCN encoder/decoder is used, while for graph classification, a GIN encoder/decoder is employed to better capture global information.
- Task-Specific Fine-Tuning (Downstream Task): After the self-supervised pre-training, the GNN encoder is used to generate feature embeddings for the nodes. These embeddings are then fed into a task-specific layer for final prediction. This stage can be applied to various tasks, including node-level identification of CGs, essential genes (EGs), and healthy driver (HD) genes, as well as graph-level identification of disease subnetworks. For node-level tasks, a learnable classification layer is used with a weighted masked binary cross-entropy loss. For graph-level tasks, a linear SVM is employed on the graph representations.
Datasets
The study utilized the following datasets:
- Multi-omic Data: Gene mutation, copy number, DNA methylation, and gene expression data from 29,446 samples across 16 cancer types, primarily sourced from The Cancer Genome Atlas (TCGA) [5].
- Protein-Protein Interaction (PPI) Networks: Data was collected from multiple public databases including CPDB [30], STRING-db [31], IRefIndex [32], Multinet [33], and PCNet [34]. The authors also used the latest versions of CPDB (2021) and IRefIndex (2022) for evaluation.
- Labeled Gene Datasets:
- Cancer Genes (CGs): Positive labels were obtained from the Network of Cancer Genes (NCG) [35] and DigSee [36]. Negative samples were randomly sampled from unlabelled genes.
- Essential Genes (EGs): Data was summarized from 16 human EG datasets from the DEG database [37]. Genes present in at least five datasets were considered positive.
- Healthy Driver (HD) Genes: Labeled HD genes were obtained from NCG [35].
- Disease Subnetworks: Disease subtype labels were directly from TCGA [5].
The paper does not introduce new benchmark datasets but rather uses existing ones for comprehensive evaluation.
Key Insights and Ideas
- Self-supervised learning is powerful for under-resourced biological data: The success of SMG demonstrates that self-supervised learning can effectively learn meaningful representations from unlabeled biological data, a critical advantage in fields like cancer genomics where labeled data is scarce.
- Graph structure and multi-omic features are complementary: SMG's ability to integrate both the topological information of PPI networks and the features from multi-omic data is key to its superior performance.
- Masked reconstruction is an effective pretext task: The strategy of masking nodes and reconstructing them forces the model to learn deep contextual relationships within the graph, leading to more robust gene embeddings.
- Generalization across tasks: SMG shows strong performance not only on CG identification but also on other related tasks like EG and HD gene identification, and even graph-level disease subnetwork identification, highlighting its versatility.
- Interpretability aids validation: The use of GNNExplainer and Integrated Gradient provides insights into which genes and interactions are most important for predictions, lending credibility to the model's findings and offering avenues for further biological investigation.
Key Results
- Superior Performance on CG Identification: SMG consistently outperformed state-of-the-art methods (including EMOGI, MTGCN, GAT, DGI, and GraphMAE2) across eight different PPI networks for CG identification, achieving an average AUPRC improvement of 7.4% over the second-best model.
- Strong Performance on Other Node-Level Tasks: SMG also demonstrated significant improvements on Essential Gene (EG) identification (average 8.7% AUPRC improvement) and Healthy Driver (HD) gene identification, even under extreme data scarcity (up to 96.39% AUPRC).
- Effective Graph-Level Performance: For disease subnetwork identification, SMG achieved the highest AUPRC of 0.87, outperforming other methods.
- Qualitative Validation: t-SNE visualizations showed distinct distribution patterns for positive and negative genes in the feature space, indicating SMG's ability to capture label-related semantic information.
- Identification of Potential CGs: The study identified several potential CGs, such as Dystonin (DST), KMT2B, and SETD2, and provided biological context and supporting literature for their roles in cancer.
Limitations and Future Work
- Limitations:
- The paper acknowledges that while SMG performs well, the performance of deep learning models can be affected by the "negative transfer phenomenon" in graph self-supervised training, where complex pre-training might negatively impact downstream tasks.
- The study primarily focuses on PPI networks; integrating other biological networks could potentially further enhance performance.
- Future Work:
- The authors envision SMG as a useful tool for predicting novel potential CGs and elucidating molecular mechanisms from PPI network data.
- The interpretability tools used (GNNExplainer, Integrated Gradient) suggest future work in exploring underlying molecular mechanisms associated with identified cancer genes.
- The framework's generalization capability suggests its potential application to other biological prediction tasks beyond cancer gene identification.
Summary updated