Explainable AI Model Reveals Informative Mutational Signatures for Cancer-Type Classification.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence in Oncology

Background:

• Cancer type prediction traditionally relies on driver genes and mutations.
• Omics technologies provide new genetic data for improved cancer diagnosis.
• Mutational signatures offer insights into DNA repair malfunctions, aiding cancer diagnosis.

Purpose of the Study:

• To compare machine learning approaches for cancer type prediction.
• To leverage artificial intelligence and omics data for enhanced cancer diagnosis accuracy.
• To investigate the utility of mutational signatures versus driver gene mutations for cancer classification.

Main Methods:

• Compared unsupervised and supervised machine learning models, including deep and artificial neural networks.
• Utilized layerwise relevance propagation for feature extraction in cancer-type prediction.
• Optimized neural network architecture using ten-fold cross-validation and grid search with driver gene mutations, mutational signatures, and topological mutation information from the PCAWG dataset.

Main Results:

• Using whole genome or intergenic/intronic regions increased mutation information relevance by over 10% compared to exome data.
• Mutational signatures were more relevant for cancer-type discrimination than topological mutation information for most cancer types.
• The approach successfully discriminated between 17 primary sites and 24 cancer types.

Conclusions:

• Informative mutational signatures provide superior cancer type prediction compared to driver gene mutations, adding diagnostic value.
• Mutational signatures can differentiate cancer types from the same primary site based on distinct relevant mutations.
• Analysis of mutational signatures enables the assignment of specific impaired DNA repair mechanisms and aids in cancer type classification.