TL;DR
This paper introduces PGLCN, a novel biologically informed graph neural network that integrates multi-omics data to predict tumor mutation burden (TMB) and identify key biological pathways involved in cancer immunotherapy response. PGLCN not only achieves superior prediction accuracy compared to traditional methods but also offers enhanced interpretability, revealing potential therapeutic targets like Toll-like receptors and DNA repair pathways for gastric cancer.
What Problem Does This Paper Address?
Predicting response to cancer immunotherapy, particularly immune checkpoint inhibitors (ICIs), remains a significant challenge. While Tumor Mutation Burden (TMB) is a promising biomarker, its correlation with ICI efficacy is not always straightforward due to the complex biological landscape of tumors. This complexity necessitates a deeper understanding and contextualization of TMB. Current methods for extracting meaningful biological insights from vast multi-omics datasets are often formidable. The paper aims to bridge this gap by developing a model that can accurately predict TMB status and, crucially, provide interpretable insights into the underlying biological mechanisms driving this prediction, thereby aiding in the development of more effective personalized therapies.
Methods
The core innovation is the PGLCN (Pathway-informed Graph Convolutional Network), a biologically informed graph neural network. Its architecture comprises three main modules:
- Biological Graph Formation: This module constructs a patient-specific graph. Nodes represent biological pathways, and node features are derived from multi-omics data (mRNA expression, Copy Number Variation (CNV), and DNA methylation) for each pathway. Pathway interactions are defined using the Reactome database, forming an adjacency matrix. Principal Component Analysis (PCA) is used for dimensionality reduction of the multi-omics data.
- Graph Learning Module: This module learns an optimal graph structure that best serves the subsequent graph convolutional layers. It aims to learn pathway interaction strengths by optimizing a non-negative function that establishes pairwise relationships between pathways.
- Graph Convolutional Module: This module processes the learned graph structure and node features through graph convolutional layers to predict TMB status.
- Graph Explanation Module: This module, inspired by GNNExplainer, learns graph and feature masks to identify crucial subgraph structures and node features responsible for the model's predictions, thereby enhancing interpretability.
The model is trained using a cross-entropy loss function, with a combined loss function incorporating a regularization term to control graph sparsity.
Datasets
The study utilized several datasets:
- TCGA (The Cancer Genome Atlas) datasets: Collected from UCSC and cBioPortal, these provided multi-omics data (mRNA expression, CNV, DNA methylation) and clinical information for cancer patients across 33 human cancer types.
- Reactome database: Used to establish pathway hierarchy relationships and gene sets for constructing the biological graphs.
- Immunotherapy datasets: Specifically, the tide database and the Kim cohort (immune blockade therapy cohort of Gastric cancer) were used for integrative analysis.
Synthetic datasets were also constructed to evaluate the interpretability of the PGLCN model, featuring different graph motifs (BA-House, BA-Community, BA-Grid, Tree-Cycles, Tree-Grids).
Key Insights and Ideas
- Multi-omics integration is key: The study highlights that combining different omics data (mRNA, CNV, methylation) significantly improves TMB prediction accuracy compared to using individual omics alone.
- Biological priors enhance interpretability: By incorporating pathway interaction information from Reactome into the graph structure, PGLCN not only predicts TMB but also provides biologically meaningful explanations for its predictions.
- TMB status is cancer-type specific: While distinct biological variations were observed across TMB statuses in STAD, COAD, and UCEC, these variations were not uniform across all 33 cancer types, suggesting context-specific biological mechanisms.
- Toll-like Receptors (TLRs) and DNA Repair pathways are critical: The model's interpretability revealed that the Toll-like receptor family and DNA repair pathways are strongly associated with TMB status in gastric cancer, suggesting their potential as biomarkers or therapeutic targets.
- Interpretable AI for drug discovery: The work demonstrates how interpretable graph neural networks can accelerate the discovery of novel therapeutic strategies by pinpointing critical biological pathways.
Key Results
- Superior TMB Prediction: PGLCN achieved high AUC values for TMB status prediction in STAD (0.976 ± 0.007), COAD (0.994 ± 0.007), and UCEC (0.947 ± 0.023), outperforming traditional machine learning models (Tables 1-3).
- Enhanced Interpretability: The model demonstrated strong interpretability across various synthetic datasets, with accuracy scores close to or exceeding GNNExplainer (e.g., BA-House: 1.0 for PGLCN vs. 0.99926 for GNNExplainer).
- Informative Omics: mRNA expression was found to be the most informative omics type for TMB prediction, though combinations of omics data yielded better results.
- Identification of Biomarkers: The analysis identified the Toll-like receptor family and DNA repair pathways as potential combined biomarkers for gastric cancer immunotherapy. Specifically, mutations in TLRs were prevalent, and their expression levels correlated with TMB status and immune cell infiltration.
- Therapeutic Synergy: The findings suggest a potential synergistic targeting strategy involving TLRs and DNA repair pathways in conjunction with TMB status for gastric cancer immunotherapy.
Limitations and Future Work
- Model Training Dependency: The authors acknowledge that PGLCN, like all deep learning models, requires training, and hyperparameters significantly influence performance.
- Prior Knowledge Reliance: The model's application to different tasks is dependent on the biological priors used to construct the graph structure.
- Future Directions: The paper suggests that further exploration of the interplay between TLRs, the tumor microenvironment, and TMB could lead to more refined therapeutic strategies. The integration of PGLCN with immunotherapy datasets also opens avenues for discovering novel therapeutic targets and understanding resistance mechanisms. The authors also propose that future work could focus on developing more robust methods for integrating diverse biological data and enhancing the interpretability of complex biological networks.
Summary updated