Using transfer learning approaches to predict RNA-Seq gene expression data for cancer classification.

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Accurate cancer type classification is vital for effective early detection and treatment strategies.
• RNA-Sequencing (RNA-Seq) data provides critical insights into gene expression but presents analytical challenges due to its high dimensionality.
• Previous analyses often focused on single cancer types, limiting comprehensive classification and gene discovery.

Purpose of the Study:

• To classify diverse cancer types using RNA-Seq data.
• To identify promising genes for cancer classification.
• To evaluate the efficacy of Transfer Learning (TL) algorithms for multi-class cancer type classification.

Main Methods:

• Utilized RNA-Seq data from the Mendeley repository.
• Transformed RNA-Seq values into 2D images for analysis.
• Employed five Transfer Learning (TL) algorithms: VGG16, VGG19, Resnet50, Resnet101, and Resnet152.
• Applied four data splitting strategies and compared performance with and without data augmentation.

Main Results:

• The optimal data split for classification accuracy was determined to be 70-30.
• VGG16 achieved the highest overall accuracy at 95%.
• Comparative analysis highlighted the performance variations among different TL models and splitting strategies.

Conclusions:

• VGG16 demonstrated superior performance as a TL algorithm for classifying cancer types from RNA-Seq data.
• The image-based transformation of RNA-Seq data is an effective approach for cancer classification.
• The study underscores the potential of TL for complex genomic data analysis in oncology.