BCDForest: a boosting cascade deep forest model towards the classification of cancer subtypes based on gene...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Cancer subtype classification is crucial for diagnosis and therapy.
• Deep learning models show promise but struggle with small, high-dimensional biological data.
• Standard deep forest models face overfitting and diversity challenges with limited sample sizes.

Purpose of the Study:

• To propose BCDForest, a novel deep learning model for cancer subtype classification.
• To address limitations of standard deep forest models in handling small-scale, high-dimensional biological data.
• To enhance ensemble diversity and feature discriminability for improved classification performance.

Main Methods:

• Introduced a multi-class-grained scanning method to train diverse binary classifiers.
• Incorporated classifier fitting quality into representation learning.
• Implemented a boosting strategy to emphasize important features within cascade forests.

Main Results:

• BCDForest demonstrated superior performance in cancer subtype classification on microarray and RNA-Seq datasets.
• The proposed methods effectively mitigated overfitting and enhanced model robustness.
• Systematic experiments confirmed consistent outperformance against state-of-the-art methods.

Conclusions:

• BCDForest offers an effective solution for cancer subtype classification using deep learning on challenging biological datasets.
• The multi-class-grained scanning and boosting strategies improve deep forest model robustness for small-scale data.
• This approach provides a valuable tool for analyzing high-dimensional, small-scale biological data in cancer research.