Integration of pan-cancer multi-omics data for novel mixed subgroup identification using machine learning methods

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer is a complex disease with significant heterogeneity, even among patients with tumors from different organs.
• Identifying patient subgroups with similar molecular characteristics, irrespective of tumor origin, is crucial for developing effective treatment strategies.
• Current approaches often overlook shared molecular alterations across diverse cancer types.

Purpose of the Study:

• To develop a machine learning (ML) pipeline for identifying novel, multi-omics-based patient subgroups in pan-cancer analysis.
• To investigate the clinical and molecular characteristics of these identified subgroups.
• To create and validate classification models for predicting subgroup membership in unseen samples.

Main Methods:

• Concatenation and non-linear dimensionality reduction of multi-omics data (mRNA, miRNA, DNA methylation, protein expression) from pan-cancer samples using ML algorithms.
• Clustering of the projected data to identify novel multi-omics-based subgroups.
• Clinical characterization of subgroups, including overall survival (OS) and disease-free survival (DFS) analysis.
• Development and validation of decision-level and feature-level fused classification models for subgroup identification.

Main Results:

• Identification of novel ML-derived patient subgroups within pan-cancer data.
• Significant differences in overall survival (OS) and disease-free survival (DFS) were observed across the identified subgroups (p-value < 0.0001).
• Subgroups comprised patients from different tumor types but shared similar molecular alterations, including immune microenvironment, mutation profiles, and enriched pathways.
• Validated classification models accurately assigned class labels to validation samples, confirming subgroup molecular characteristics.

Conclusions:

• Novel multi-omics-based patient subgroups can be identified in pan-cancer analysis using machine learning.
• Patients with different tumor types can exhibit similar molecular characteristics, challenging traditional organ-specific classifications.
• The developed classification models are effective for identifying these novel subgroups and can inform the design of tailored treatment regimens based on subgroup-specific molecular profiles.