Multi-omics driven computational framework for cancer molecular subtype classification

Area of Science:

• Computational biology
• Bioinformatics
• Precision oncology

Background:

• AI is crucial for cancer molecular subtype classification, guiding prognosis and targeted therapies.
• Current AI applications face challenges due to non-standardized datasets, diverse omics data, and inconsistent evaluation metrics.
• These limitations hinder AI classifier comparability, reproducibility, and generalizability.

Purpose of the Study:

• To conduct a comprehensive comparative analysis of 35 AI classifiers across 153 datasets.
• To investigate the impact of different omics modalities and dataset configurations on AI performance.
• To identify optimal AI models and data types for robust cancer molecular subtype classification.

Main Methods:

• Comparative analysis of 35 AI classifiers on 153 datasets covering 8 omics modalities and 20 cancer types.
• Evaluation of classifier performance based on macro-accuracy (MACC) and other metrics.
• Investigation of 6 research questions regarding data configurations, omics modalities, and model types (ML vs. DL).

Main Results:

• RPPA, Gistic2-all-data-by-genes (CNV), HM27 (Meth), and HiSeqV2-exon (Exon) configurations showed better performance.
• RNASeq, miRNA, CNV, and Exon modalities generally achieved higher MACC than Meth., Array, SNP, and RPPA.
• Traditional machine learning (ML) models excelled on small datasets, while deep learning (DL) models performed better on large, high-dimensional datasets.
• SVM achieved the highest mean MACC, with NN, ResNet18, DEEPGENE, and MLP also showing strong results.
• DL classifiers outperformed ML classifiers in 12 out of 20 cancers.

Conclusions:

• Specific data configurations and omics modalities are superior for AI-based cancer classification.
• The choice between ML and DL models depends on dataset size and dimensionality.
• Findings provide insights for developing standardized, reproducible, and efficient AI pipelines for precision oncology.