Cancer detection via one-shot learning: integrating gene expression and genomic mutation analysis.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Cancer is a complex disease with diverse tumor microenvironments (TMEs) influenced by multiple genetic factors.
• Molecular classification of cancer is crucial for precise therapeutic strategies, but traditional machine learning (ML) methods have limitations.
• Existing ML approaches often rely solely on gene expression data, neglecting crucial genomic alterations and requiring large sample sizes.

Purpose of the Study:

• To develop a novel ML-based method for cancer detection that integrates gene expression and genomic mutation data.
• To create a type-agnostic representation for characterizing TMEs and enabling generalization to unseen cancer types.
• To enhance the interpretability of cancer detection models using SHapley Additive exPlanations (SHAP) values.

Main Methods:

• A one-shot learning framework using Siamese Neural Networks was implemented for cancer detection.
• Cancer detection was redefined as a similarity-based classification task to handle data scarcity.
• A SHAP-based explainability technique was developed to analyze the contributions of gene expression and mutation data.

Main Results:

• Integrating mutational profiles with gene expression data improved cancer type detection accuracy.
• The method revealed significant mutation patterns associated with different cancer types.
• The model demonstrated the ability to generalize to unseen cancer types.

Conclusions:

• The proposed method enhances cancer type detection by providing a comprehensive understanding of TMEs.
• The SHAP-based explainability technique aids in identifying key biomarkers for immunotherapy success.
• This approach addresses limitations of existing methods by leveraging integrated genomic data and improving model interpretability.