CancerNet: a unified deep learning network for pan-cancer diagnostics

Area of Science:

• Computational biology
• Genomics
• Artificial intelligence in medicine

Background:

• Cancer remains a leading global cause of death, underscoring the need for early detection and origin localization.
• Technological advancements in artificial intelligence (AI) offer new avenues for cancer diagnostics.
• The Cancer Genome Atlas (TCGA) provides a rich dataset for understanding cancer epigenomes.

Purpose of the Study:

• To develop a unified AI framework for detecting and classifying cancers and their tissues of origin.
• To leverage epigenomic features for early cancer detection.
• To create a model encompassing 33 distinct cancer types.

Main Methods:

• Development of a novel AI framework for cancer diagnostics.
• Utilizing epigenomic dysregulation patterns as biomarkers.
• Training a unified model on data from 33 TCGA cancer types.

Main Results:

• The AI model achieved an overall F-measure exceeding 99% for cancer detection and classification across 33 tissues of origin.
• The model successfully differentiated between cancerous, pre-cancerous, and metastatic lesions.
• It accurately distinguished true cancer-related hypomethylation from age-related epigenetic changes.

Conclusions:

• The computational model robustly identifies the tissue of origin for primary, metastatic, and unknown primary cancers.
• It demonstrates capability in characterizing pre-cancerous samples, advancing early detection efforts.
• Broad deployment of this AI model promises accurate diagnosis for a wider patient population.