Explainable deep learning for identifying cancer driver genes based on the Cancer Dependency Map.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Identifying cancer driver genes and mutations is crucial for understanding tumor progression and developing targeted therapies.
• Existing methods face challenges in accurately detecting these critical genetic alterations.

Purpose of the Study:

• To leverage the Cancer Dependency Map (DepMap) for identifying potential cancer driver genes and inferring multi-driver mutations.
• To develop robust and interpretable deep learning models for cancer driver gene discovery and mutation pattern analysis.

Main Methods:

• Developed xNNDriver, a supervised deep learning model linking gene mutation status to genome-wide dependency scores.
• Utilized xAEDriver, an unsupervised explainable autoencoder for inferring multiple driver variant representations (DVRs) simultaneously.
• Applied models to the Cancer Dependency Map (DepMap) dataset.

Main Results:

• xNNDriver successfully identified established drivers (e.g., NRAS, KRAS, HRAS, BRAF) and novel candidates (e.g., VDR).
• The model's interpretability revealed downstream pathways regulated by driver genes.
• xEADriver-based stratification of cell lines showed distinct drug response profiles, indicating translational potential.

Conclusions:

• The developed framework provides a powerful and interpretable approach for cancer driver gene discovery.
• The models offer new insights into cancer biology and mutation patterns.
• This work has significant implications for advancing targeted cancer therapies.