RN-Autoencoder: Reduced Noise Autoencoder for classifying imbalanced cancer genomic data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression datasets are crucial for cancer classification in the genomic era.
• These datasets often suffer from high dimensionality and class imbalance, hindering classifier performance.
• These inherent data characteristics negatively impact the accuracy of cancer classification models.

Purpose of the Study:

• To introduce the Reduced Noise-Autoencoder (RN-Autoencoder) as a novel preprocessing method for imbalanced genomic datasets.
• To address the dual challenges of high dimensionality and class imbalance in gene expression data for improved cancer classification.
• To enhance the precision and performance of cancer classification using machine learning algorithms.

Main Methods:

• Utilizes an autoencoder for feature reduction, effectively solving the curse of dimensionality.
• Applies Reduced Noise-Synthesis Minority Over Sampling Technique (RN-SMOTE) to address class imbalance in the reduced dataset.
• Evaluates the RN-Autoencoder model using various classifiers and imbalanced genomic datasets with differing imbalance ratios.

Main Results:

• The RN-Autoencoder significantly improved classifier performance compared to original and reduced data.
• Achieved notable increases in test accuracy across multiple datasets, including colon, leukemia, DLBCL, and WDBC.
• Demonstrated superior performance against current state-of-the-art methods, with accuracy improvements up to 18-19% on specific datasets.

Conclusions:

• The RN-Autoencoder is an effective model for cancer classification using imbalanced gene expression data.
• The model successfully reduces dimensionality and handles class imbalance, leading to improved classifier performance.
• RN-Autoencoder achieved 100% classification performance in some cases and outperformed existing methods on benchmark datasets.