Integrating somatic mutation profiles with structural deep clustering network for metabolic stratification in pancreatic cancer: a comprehensive analysis of prognostic and genomic landscapes

Min Zou, Honghao Li, Dongqing Su, Yuqiang Xiong, Haodong Wei, Shiyuan Wang, Hongmei Sun, Tao Wang, Qilemuge Xi, Yongchun Zuo, Lei YangBriefings in Bioinformatics 2024 · DOI

TL;DR

Pancreatic cancer's complexity makes it hard to treat. This study introduces a novel method to classify pancreatic cancer subtypes by analyzing somatic mutations and their impact on metabolic pathways. This approach successfully identifies distinct subtypes with different prognoses and immune profiles, paving the way for more personalized treatment strategies.

What Problem Does This Paper Address?

Pancreatic cancer is a deadly disease with a poor prognosis, largely due to its aggressive nature and significant heterogeneity. While somatic mutations are known drivers of cancer development, their high dimensionality and sparsity make them challenging to analyze directly for patient stratification. Existing methods often struggle to capture the underlying biological complexity from this type of data. This paper aims to bridge this gap by developing a method to effectively leverage somatic mutation data to uncover biologically and clinically relevant subtypes of pancreatic cancer, focusing on metabolic pathways as a key stratifying feature.

Methods

The core of this research lies in a multi-step computational pipeline:

  1. Network Propagation for Smoothed Mutation Profiles: The authors first address the sparsity of somatic mutation data by using a network propagation algorithm (specifically, random walk with restart) on protein-protein interaction (PPI) networks (STRING database). This process smooths the discrete mutation data, incorporating network information to create a more continuous and biologically informed "smoothed somatic mutation profile."
  2. Metabolic Pathway Activity Calculation: Using the smoothed mutation profiles, the study calculates the activity levels of 114 metabolic pathways per patient via single-sample gene set enrichment analysis (ssGSEA).
  3. Structural Deep Clustering Network (SDCN): To identify distinct patient subtypes based on metabolic pathway activity, the researchers employ SDCN. This method integrates autoencoders with graph convolutional networks, allowing it to effectively capture structural relationships within the data. They first identify prognostic-related metabolic pathways using univariate Cox regression and then cluster patients based on these pathways.
  4. Differential Gene and Functional Enrichment Analysis: To understand the biological differences between identified subtypes, they perform differential gene expression analysis and functional enrichment analyses (GO, KEGG, GSEA) on the network propagation-based gene profiles.
  5. Machine Learning-Based Metabolic Risk Score (MRS): A prognostic model is built using an ensemble of machine learning algorithms to create a metabolic risk score, aiming to predict patient survival.
  6. XGBoost Classifier: An XGBoost classifier is trained to predict patient cluster subtypes based on metabolic pathway activity, demonstrating high accuracy.

Datasets

The primary dataset used in this study is a collection of somatic mutational profiles, genomic data, and clinical data from 1942 pancreatic cancer patients obtained from the Memorial Sloan Kettering Cancer Center (MSKCC), as reported by Nguyen et al. [33]. This dataset was downloaded from the cBioPortal for Cancer Genomics. The authors also mention using data from other cancer cohorts within the MSKCC database for validation and the TCGA pancreatic adenocarcinoma (PAAD) cohort.

Key Insights and Ideas

  • Somatic Mutations as a Foundation for Metabolic Stratification: The study demonstrates that even sparse and high-dimensional somatic mutation data can be effectively transformed and utilized to infer metabolic pathway activity, a crucial aspect of cancer biology.
  • Network Propagation Bridges Data Sparsity: The use of network propagation on PPI networks is a key innovation, transforming discrete mutation data into a more informative representation that captures biological context.
  • SDCN Uncovers Clinically Relevant Subtypes: The application of structural deep clustering successfully identifies distinct pancreatic cancer subtypes that are not only biologically meaningful but also have significant prognostic implications.
  • Metabolic Subtypes Reflect Immune Landscape: The identified metabolic subtypes are strongly associated with different tumor immune microenvironments, suggesting a link between metabolism and immune response in pancreatic cancer.
  • Predictive Power for Immunotherapy: The study suggests that these subtypes may predict response to immunotherapy, offering a potential avenue for personalized treatment selection.

Key Results

  • Two Distinct Metabolic Subtypes: The SDCN algorithm identified two distinct metabolic subtypes of pancreatic cancer patients in the training cohort, which showed a statistically significant difference in overall survival (P-value = 2.26E-8). This was validated in a testing cohort.
  • Prognostic Metabolic Pathways: 76 metabolic pathways were found to be significantly associated with overall survival in pancreatic cancer patients.
  • High Predictive Accuracy: An XGBoost classifier achieved an impressive 97.87% accuracy in predicting cluster subtypes based on metabolic pathway activity in the training cohort.
  • Differential Immune Signatures: The two subtypes exhibited significantly different infiltration levels of various immune cells and distinct immune scores, stromal scores, and tumor purity.
  • Immunotherapy Response Prediction: Cluster 2 subtype patients showed a significantly higher proportion of response to Immune Checkpoint Blockade (ICB) treatment compared to cluster 1 subtype patients.

Limitations and Future Work

The authors acknowledge several limitations:

  • Data Source Dependency: The study relied solely on the MSKCC database. Future work should involve validation with other somatic mutation databases.
  • Multi-Omics Integration: While the study effectively uses somatic mutations, corroborating findings with mRNA expression or other omics data would strengthen the biological insights.
  • Immunotherapy Data: Predictive data for immunotherapy response was derived from somatic mutation data; further validation with actual clinical immunotherapy data is needed.
  • PPI Network Scope: The study used the STRING database for PPIs. Exploring other types of biological networks (e.g., signaling, metabolism) could potentially enhance the network propagation model's performance.

Future work could also involve applying this methodology to other cancer types and exploring the clinical utility of the developed metabolic risk score and nomogram model for personalized treatment decisions.

Summary updated

BibTeX
@article{zou2024integrating,
  author = {Min Zou and Honghao Li and Dongqing Su and Yuqiang Xiong and Haodong Wei and Shiyuan Wang and Hongmei Sun and Tao Wang and Qilemuge Xi and Yongchun Zuo and Lei Yang},
  title = {{Integrating somatic mutation profiles with structural deep clustering network for metabolic stratification in pancreatic cancer: a comprehensive analysis of prognostic and genomic landscapes}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbad430},
}
Figure