MFmap: A semi-supervised generative model matching cell lines to tumours and cancer subtypes

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Translating in vitro cancer cell line findings to clinical practice is challenging due to the need for accurate model selection.
• Existing methods may not fully capture the complex relationships between cell line data and individual patient tumors.

Purpose of the Study:

• To introduce MFmap (model fidelity map), a novel machine learning model for predicting cancer cell line subtypes and their similarity to individual tumors.
• To enable more effective translation of preclinical cancer research to personalized clinical applications.

Main Methods:

• MFmap utilizes a semi-supervised generative approach to compress high-dimensional omics data (gene expression, copy number variation, mutation data).
• The model generates low-dimensional latent representations that are informed by cancer subtypes.
• Model performance was validated across ten diverse cancer datasets.

Main Results:

• MFmap achieved high accuracy in cancer subtype prediction, with a test set F1 score exceeding 90%.
• The model successfully demonstrated the translation of subtype-specific drug sensitivities to individual breast cancer and glioblastoma tumor samples.
• Extracted latent representations revealed known and novel subtype-specific features and facilitated the analysis of cell-state transitions.

Conclusions:

• MFmap is a robust semi-supervised method offering both generative and predictive capabilities in cancer research.
• The model enhances the selection of cell line models for preclinical studies and aids in personalized medicine by linking cell line data to patient tumors.
• MFmap's approach holds potential for broader applications within computational biology.