Diagnostic classification of cancers using extreme gradient boosting algorithm and multi-omics data.

Area of Science:

• Computational biology
• Oncology
• Bioinformatics

Background:

• Accurate cancer staging is crucial for effective patient surveillance and treatment strategies.
• The increasing volume of biological data necessitates advanced computational methods for cancer prognosis.
• Machine learning (ML) is integral to modern cancer prediction, moving beyond traditional biostatistics.

Purpose of the Study:

• To develop and evaluate an extreme gradient boosting (XGBoost) classification model for distinguishing early-stage from late-stage cancers.
• To assess the performance of XGBoost using multi-omics data from The Cancer Genome Atlas (TCGA).
• To compare XGBoost with other ML methods and identify the most effective molecular data types for cancer staging.

Main Methods:

• Applied the XGBoost classification model to multi-omics data (DNA methylation, mRNA expression, miRNA expression) from four cancer types obtained from TCGA.
• Compared XGBoost performance against other popular machine learning algorithms.
• Utilized autoencoder for multi-omics data integration and performed bioinformatics analyses on gene importance rankings.

Main Results:

• The XGBoost model achieved statistically significant or comparable predictive performance against other ML methods.
• DNA methylation data demonstrated superior accuracy and stability in discriminating between early and late cancer stages compared to mRNA and miRNA expression.
• Integration of multi-omics data via autoencoder enhanced cancer stage classification accuracy.

Conclusions:

• XGBoost is a powerful and effective method for predicting cancer patient stages using multi-omics data.
• DNA methylation is a key molecular biomarker for cancer staging.
• The study identified novel candidate genes associated with cancer stages, potentially aiding in understanding disease pathogenesis and developing new therapeutics.