Predicting cancer type from tumour DNA signatures

Area of Science:

• Genomics
• Computational Biology
• Oncology

Background:

• Accurate cancer type and origin identification is vital for effective patient treatment.
• Cancer of unknown primary (CUP) presents diagnostic challenges and is associated with poor patient survival.
• Gene alteration data from tumor DNA offers potential for identifying cancer types.

Purpose of the Study:

• To evaluate the utility of gene alteration data from tumor DNA for cancer type identification.
• To assess the predictive performance of machine-learning models using genomic features.
• To determine the optimal combination of gene alteration types for accurate cancer diagnosis.

Main Methods:

• Utilized sequenced tumor DNA data from 6640 samples across 28 cancer types via cBioPortal.
• Employed machine-learning techniques including linear support vector machines, L1-regularized logistic regression, and random forest.
• Selected informative gene alterations for cancer-type prediction.

Main Results:

• Linear support vector machines demonstrated the highest predictive accuracy.
• Prediction using 100 somatic point-mutated genes achieved 49.4% accuracy.
• Incorporating copy number alterations with somatic point mutations using 50 genes yielded 77.7% accuracy.

Conclusions:

• Neither somatic point mutations nor copy number alterations alone are sufficient for broad cancer type differentiation.
• Combining both somatic point mutations and copy number alterations significantly enhances diagnostic performance.
• Genomic profiling holds promise for developing advanced cancer diagnostic tools.