Overcoming Interpretability in Deep Learning Cancer Classification.

Area of Science:

• Genomics
• Machine Learning
• Computational Biology

Background:

• Deep learning has significantly advanced machine learning and data-driven science, particularly in genomics.
• Applications in genomics include predicting regulatory elements and cancer classification.
• A key limitation of deep learning is its lack of interpretability.

Purpose of the Study:

• To address the interpretability challenge in deep learning for cancer classification.
• To adapt and validate a feature importance scoring methodology for deep learning models in genomics.

Main Methods:

• Utilized a quasi-recurrent neural network for cancer classification based on FASTA sequencing data.
• Employed Gradient-based Class Activation Mapping (Grad-CAM) for feature importance scoring.
• Developed a novel nucleotide-to-genomic-region Grad-CAM scoring methodology.

Main Results:

• Successfully formulated and validated a Grad-CAM scoring methodology for deep learning cancer classification.
• Demonstrated the utility of Grad-CAM for assessing feature importance in genomic data.
• Identified potential novel candidate genes, genomic elements, and mechanisms relevant to cancer.

Conclusions:

• The developed Grad-CAM methodology enhances the interpretability of deep learning models in cancer genomics.
• This approach facilitates the identification of key genomic features driving cancer classification.
• The findings provide a foundation for future cancer research by highlighting potential novel targets and mechanisms.