Classification of non-TCGA cancer samples to TCGA molecular subtypes using compact feature sets

Area of Science:

• Computational biology
• Genomics
• Machine learning in oncology

Background:

• Molecular subtypes, defined by The Cancer Genome Atlas (TCGA), are crucial for understanding cancer biology, prognosis, and treatment.
• Existing subtype discovery methods are often not applicable for classifying new patient samples from diverse studies.
• A need exists for robust models to assign established molecular subtypes to new clinical specimens.

Purpose of the Study:

• To develop machine learning models capable of classifying new cancer specimens into predefined TCGA molecular subtypes.
• To address the limitation of subtype discovery methods in classifying samples from external datasets.
• To facilitate the clinical application of molecular subtyping by enabling subtype assignment for new tumors.

Main Methods:

• Applied five distinct machine learning approaches to multi-omic data from 8,791 TCGA tumor samples across 26 cancer cohorts and 106 subtypes.
• Developed models utilizing a minimal set of features for efficient classification of new samples.
• Validated selected predictive models using independent external datasets to assess generalizability.

Main Results:

• Successfully built and validated machine learning models for classifying new samples into established TCGA molecular subtypes.
• Identified key features driving subtype classification, providing insights into the biological underpinnings of each subtype.
• Evaluated the performance of different machine learning algorithms and their suitability for multi-omic cancer data.

Conclusions:

• The developed machine learning models represent a significant step towards the clinical utility of molecular subtyping.
• Containerized versions of top-performing models are provided as a public resource for each cancer and data type.
• This work bridges the gap between molecular subtype discovery and practical application in clinical settings and research.