MASE-GC: a multi-omics autoencoder and stacking ensemble framework for gastric cancer classification.

Area of Science:

• Computational oncology
• Bioinformatics
• Genomics

Background:

• Gastric cancer (GC) is a leading cause of cancer mortality worldwide.
• Accurate GC classification is crucial for diagnosis, prognosis, and personalized treatment.
• Existing single-omics or single-model approaches fail to capture tumor complexity, limiting their effectiveness.

Purpose of the Study:

• To develop a novel computational framework, MASE-GC, for precise gastric cancer classification.
• To integrate diverse multi-omics data (exon, mRNA, miRNA, DNA methylation) for enhanced accuracy.
• To improve sensitivity, specificity, and generalizability in GC stratification.

Main Methods:

• MASE-GC utilizes modality-specific autoencoders for feature extraction and weighted fusion.
• A stacking ensemble of five base learners (SVM, RF, DT, AdaBoost, CNN) with an XGBoost meta-classifier is employed.
• A robust preprocessing pipeline addresses noise, high dimensionality, and class imbalance.

Main Results:

• MASE-GC achieved superior classification performance on the TCGA-STAD cohort (accuracy 0.981, F1-score 0.9883).
• Ablation studies highlighted the contributions of autoencoders and ensemble components (CNN, Random Forest).
• Independent validation on external cohorts confirmed MASE-GC's robustness and generalizability (accuracy > 0.958).

Conclusions:

• MASE-GC provides an effective and generalizable framework for gastric cancer classification.
• The integration of multi-omics fusion, ensemble learning, and robust preprocessing enhances diagnostic capabilities.
• MASE-GC shows strong potential for clinical translation in precision diagnostics and treatment planning.