Identifying Prognostic Genes for Cancer Through Graph Analysis
What is Cancer Prognosis Prediction? Cancer prognosis prediction refers to estimating outcomes such as a patient’s survival period or the likelihood of recurrence after surgery.
In recent years, the development of prognosis prediction models using artificial intelligence techniques like machine learning has been actively pursued. In particular, many studies are utilizing gene expression data to identify factors that influence cancer prognosis and to build predictive models.
However, due to the heterogeneous nature of cancer, there remain significant challenges in improving the accuracy of prognosis prediction.
Could inaccurate predictions potentially prevent patients from receiving the most effective treatments?
To address these research challenges, we introduce a study that applies graph machine learning to gene expression data and biological network data[1].
By Identifying gene biomarker through network centrality analysis,
it is possible to significantly enhance the accuracy of cancer prognosis prediction.
Background
Limitations of Traditional Cancer Prognosis Prediction Models
Previous studies on prognosis prediction have often focused solely on analyzing gene expression levels or trained a single model using the entire dataset.
However, due to the genetic instability and heterogeneity of cancer, such approaches have shown limitations in predictive accuracy.
Moreover, a major drawback has been the insufficient consideration of the complex relationships within biological networks.
Core Idea
Cancer Prognosis Prediction Using Clustering-Based PageRank
In this study, we propose a prognosis prediction model that first segments heterogeneous cancer patient data using K-means clustering. Within each cluster, a modified PageRank algorithm based on biological networks is applied to identify prognostic genes. These genes are then used with a K-NN classifier to predict patient outcomes.
PageRank is a graph-based algorithm originally developed by Google to assess the importance of web pages. It calculates centrality by considering not only the number of links but also the influence of the connected neighbors. In this sense, it is similar to eigenvector centrality, but it is particularly effective in networks with directionality and weighted edges.
In the field of biology, PageRank is widely used to identify key genes or biomarkers within gene interaction networks or protein-protein interaction networks.
One of its main advantages is its robustness to noise, allowing it to reliably identify core nodes even within complex relational structures.
※ Biomarker: a measurable indicator of some biological state or condition(source: Wikipedia)
Solution
Step 1. Construction of Gene Networks Across Prognostic Groups
In this study, six gene expression datasets along with clinical data and one biological network dataset were collected.
Gene Expression data & Clinical data(Nodes)
Gene Expression data: Gene expression data from six cancer types - Breast cancer(BRCA), Acute myeloid leukemia(LAML), Liver cancer(LIHC), Lung cancer(LUAD), Pancreatic cancer(PAAD), Stomach cancer(STAD)
Clinical data: Patient's death status(OS event), Survival time(OS duration), Favorable prognosis(good), Unfavorable prognosis(poor)
FI network data(Links): Composed of biological pathways, PPI (protein-protein interaction) data, gene co-expression data, GO annotations, and curated information from various research articles.
First, the z-score normalized gene expression data is reduced to two principal components using Principal Component Analysis (PCA).
Then, K-means clustering is applied to divide the samples with similar expression patterns into two clusters.
Next, gene importance (ranking) is determined using a PageRank algorithm that calculates the ratio between the PageRank score derived from a weighted network and the score from an unweighted network for each cluster.
Finally, the top N genes from each cluster are selected as biomarkers to be used for prognosis prediction.
In this study, the PageRank algorithm was adapted by quantifying gene expression differences between prognosis groups within each cluster using t-statistics, which were then used as edge weights in the biological network. As a result, genes that are connected to others with significant prognostic differences and biological importance receive higher PageRank scores.
This approach allows the analysis to focus on pathways that are more likely to contribute to actual prognosis prediction by assigning greater importance to the connections between genes with significant expression differences across prognosis groups.
Step 2. Prediction Model Construction
The expression data of prognostic biomarker genes identified through PageRank was used to train a K-NN classifier.
The final input data structure used in the K-NN classifier is as follows.
<Example of Training Data for the Prediction Model>
※ Biomarker: Yellow
Step 3. Comparison with Traditional Biomarker Identification Algorithms
In this study, to evaluate the performance of the proposed model (CPR), the authors compared its predictive accuracy with several other models: a non-clustering version of CPR (NCPR), NetRank which also uses PageRank, WGCNA which identifies prognosis-related gene modules based on co-expression patterns, and the model proposed by Wu & Stein.
Using PageRank Models: CPR, NCPR, NetRank
Without PageRank Models: WGCNA, Wu & Stein
Result
Demonstrating the Effectiveness of Graph-Based Analysis in Identifying Prognostic Biomarkers
To evaluate the performance of the CPR model, the AUC values were compared between models that used the PageRank algorithm and those that did not.
As a result, models incorporating PageRank generally demonstrated superior performance.
The following presents the results of performance comparisons between the proposed classification model and several established models.
The conventional NetRank model showed slightly higher AUC values than the CPR model in LAML and PAAD, with differences ranging from 0.002 to 0.06. In LUAD, only the Wu & Stein model outperformed CPR by 0.06 in AUC.
However, in the remaining cancer types, the CPR model demonstrated higher AUC values—ranging from 0.036 to 0.116—compared to other models, confirming its superior ability to accurately identify prognostic genes.
This approach, which combines clustering with network centrality, outperforms existing methods across all cancer types. Notably, by accounting for cancer heterogeneity and enabling group-specific analysis, it enhances both biological interpretability and predictive precision compared to models based solely on gene expression levels.
This study proposes a novel approach that applies network analysis algorithms to biological networks in order to identify key genes and predict cancer prognosis. The findings demonstrate that network analysis can serve as a powerful tool for cancer prognosis prediction.
Furthermore, it suggests the potential of extending network-based methods beyond prognosis prediction to a wide range of precision medicine applications.
CYRAM Inc.
1660-4230 | netminer@cyram.com
References
[1] Jongwan Choi, Jaegyoon Ahn. (2018).Identification of Heterogeneous Prognostic Genes and Prediction of Cancer Outcome using PageRank. Journal of KIISE Vol.45 No.1, 61-68.