Pan-Cancer Metastasis Prediction Based on Graph Deep Learning Method.

Area of Science:

• Computational biology
• Cancer research
• Machine learning in oncology

Background:

• Tumor metastasis is a primary cause of cancer mortality, necessitating accurate prognosis for therapeutic planning.
• Gene expression and transcriptome analysis are crucial for understanding metastasis mechanisms.
• Existing machine learning models struggle with limited real-world data and complex feature extraction for predicting tumor outcomes.

Purpose of the Study:

• To develop an advanced feature extraction strategy for improved prediction of cancer metastasis risk.
• To enhance the accuracy of patient prognosis by analyzing gene expression data.
• To overcome limitations of current machine learning pipelines in cancer outcome prediction.

Main Methods:

• Constructed a gene regulation network and extracted gene expression features using a relational graph convolutional network (RGCN).
• Utilized a convolutional neural network (CNN) to predict metastasis risk, treating high-dimensional features as image pixels.
• Performed ten cross-validations on 1,779 cancer patient cases from The Cancer Genome Atlas (TCGA).

Main Results:

• The directed relation-graph convolutional network achieved an Area Under the Curve (AUC) of 0.837 and an Area Under the Precision-Recall Curve (AUPRC) of 0.717.
• The proposed model significantly outperformed an existing network-based method, which reported an AUC of 0.707 and AUPRC of 0.555.
• Demonstrated superior performance in predicting patient metastasis risk compared to current approaches.

Conclusions:

• The developed directed relation-graph convolutional network offers an advanced feature extraction strategy for cancer metastasis prediction.
• The model shows significant potential for improving patient risk stratification and therapeutic decision-making in oncology.
• This approach addresses key challenges in analyzing complex gene expression data for more accurate cancer outcome prediction.