A network-based deep learning methodology for stratification of tumor mutations.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Tumor stratification is crucial for diagnosis, prognosis, and personalized cancer treatment.
• Cancer's heterogeneity, driven by combinations of mutated genes, makes accurate tumor subdivision challenging.

Purpose of the Study:

• To develop a network-embedding based stratification (NES) methodology for identifying clinically relevant patient subtypes.
• To leverage somatic mutation profiles and human interactome data for robust tumor classification.

Main Methods:

• Encoded genes on the human protein-protein interactome using network embedding.
• Constructed patient vectors by integrating somatic mutation profiles from 7344 tumor exomes across 15 cancer types.
• Utilized lightGBM for classification and unsupervised clustering for subtype identification.

Main Results:

• Achieved an AUC of approximately 0.89 for cancer type prediction and 0.78 for tumor stage prediction.
• Demonstrated that network embedding-based patient features are reliable for patient stratification.
• Identified patient subtypes significantly correlated with survival across 12 of 15 cancer types.

Conclusions:

• The NES methodology provides a powerful network-based deep learning approach for personalized cancer medicine.
• Identified patient subtypes offer new insights into cancer heterogeneity and patient outcomes.
• The study highlights the potential of integrating network information with mutation data for advancing cancer research.