DeepCC: a novel deep learning-based framework for cancer molecular subtype classification.

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Molecular subtyping is crucial for personalized cancer therapy and understanding tumor heterogeneity.
• Current gene expression-based methods face challenges like platform variability, batch effects, and classifying individual samples.

Purpose of the Study:

• Introduce DeepCC, a novel deep learning framework for supervised cancer classification.
• Evaluate DeepCC's performance against established methods for colorectal and breast cancer.
• Assess DeepCC's robustness to missing data and its ability to improve single-sample prediction.

Main Methods:

• Developed DeepCC, a supervised cancer classification framework utilizing deep learning on functional spectra of biological pathway activities.
• Applied DeepCC to colorectal and breast cancer datasets for classification and single-sample prediction.
• Compared DeepCC with Random Forests, SVM, GBM, and multinomial logistic regression.
• Conducted simulation analyses to test robustness against missing data.

Main Results:

• DeepCC classifiers and single-sample predictors demonstrated superior sensitivity, specificity, and accuracy compared to RF, SVM, GBM, and logistic regression.
• DeepCC exhibited robustness to missing data, as shown by simulation analyses.
• Learned deep features by DeepCC enhanced subtype separation and reduced unclassifiable samples.

Conclusions:

• DeepCC offers a platform-independent and robust framework for cancer molecular subtyping.
• The method facilitates single-sample prediction, paving the way for clinical implementation of personalized cancer therapies.
• DeepCC effectively addresses limitations of existing methods, improving classification accuracy and reducing ambiguity.