A novel machine learning-based workflow to capture intra-patient heterogeneity through transcriptional multi-label...

Area of Science:

• Computational biology
• Genomics
• Precision medicine

Background:

• Accurate patient classification into molecular subtypes is crucial for understanding and treating complex diseases.
• Current cancer subtyping methods often oversimplify patient molecular profiles, failing to capture intra-sample heterogeneity.
• Recognizing multiple co-occurring subtype traits is essential for precise patient characterization and personalized treatment strategies.

Purpose of the Study:

• To develop a novel computational workflow, MULTI-STAR, for reliable multi-label patient subtyping.
• To address the limitations of existing methods that neglect the co-occurrence of multiple molecular subtypes within a single patient.
• To enable more precise patient characterization and improve personalized treatment decisions.

Main Methods:

• Developed MULTI-STAR, a computational workflow leveraging gene expression profiles for multi-label patient subtyping.
• Adapted similarity-based techniques for multi-label characterization and trained single-sample predictors.
• Employed machine learning to identify and rank the best-performing multi-label classifiers.

Main Results:

• MULTI-STAR classifiers accurately recognize all contributing subtypes, distinguishing primary from secondary assignments.
• Demonstrated superior performance in multi-label subtyping for breast and colorectal cancer compared to existing methods.
• Showcased improved prognostic value, particularly for overall survival predictions, and single-sample applicability.

Conclusions:

• Multi-label subtyping is essential for capturing comprehensive molecular traits and patient heterogeneity.
• MULTI-STAR offers a reproducible and generalizable approach for advanced patient characterization.
• This method provides clinically relevant insights, advancing precision medicine and personalized treatments for heterogeneous diseases.