ML Model Uses Methylation Patterns to Identify Tissue of Origin 

Objective:

To develop a machine learning model that predicts tissue of origin in cancers of unknown primary using a focused set of DNA methylation patterns.

Key Findings:

• The model achieved an area under the curve of 0.998 and classification accuracy of 0.954 in internal validation.
• In the TCGA held-out test cohort, the model maintained a high area under the curve of 0.998 and accuracy of 0.947.
• Errors in classification reflected biological similarities rather than model failures.

Interpretation:

The model effectively captures tumor biology through DNA methylation patterns, offering a promising tool for identifying tissue of origin in cancers of unknown primary.

Limitations:

• The study primarily focused on a limited set of cancer types.
• Further validation is needed in cohorts of patients with true cancers of unknown primary.

Conclusion:

The developed classifier has the potential to improve site-directed therapy decision-making for patients with cancers of unknown primary, complementing existing genomic profiling methods.