DeepClassPathway: Molecular pathway aware classification using explainable deep learning.

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Human Papillomavirus (HPV)-associated head and neck cancers have a favorable prognosis, but the underlying biological mechanisms remain unclear.
• Understanding these mechanisms is crucial for improving patient outcomes and developing targeted therapies.

Purpose of the Study:

• To develop an explainable artificial intelligence (AI) model, DeepClassPathway, for predicting HPV status in head and neck squamous cell carcinoma.
• To identify patient-specific molecular pathways that drive the AI's classification decisions, offering insights into the tumor's biology.

Main Methods:

• A convolutional neural network (CNN) classifier was trained on transcriptome data from head and neck squamous carcinoma patients, transformed into 2D-treemaps.
• Explainability was achieved using Grad-CAM saliency maps to quantify pathway contributions to individual predictions.
• Model stability was verified through pathway shuffling within the treemaps.

Main Results:

• The CNN ensemble achieved high classification performance with ROC-AUC of 0.96 and PR-AUC of 0.90.
• Key molecular pathways identified for HPV-positive classification included KRAS, spermatogenesis, bile acid metabolism, and inflammation.
• Pathways associated with HPV-negative classification included MYC targets, epithelial-mesenchymal transition, and protein secretion.

Conclusions:

• An explainable CNN approach was successfully developed and applied to transcriptome data for HPV status classification.
• This method allows for classification while simultaneously highlighting the importance of specific molecular pathways in individual patient cases.
• The findings provide a foundation for understanding the molecular drivers of HPV-associated head and neck cancers and can inform personalized treatment strategies.