Gene Swin transformer: new deep learning method for colorectal cancer prognosis using transcriptomic data

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Transcriptome sequencing is crucial for tumor research but generates complex data.
• Identifying clinically relevant information from vast gene expression data is challenging.
• Effective prognostic models are needed for colorectal cancer treatment.

Purpose of the Study:

• To develop a novel method for analyzing transcriptome data for cancer prognosis.
• To create a reliable prognostic model for colorectal cancer using deep learning.
• To identify key prognostic markers in colorectal cancer.

Main Methods:

• Developed the Gene Swin Transformer method to convert transcriptomic data into Synthetic Image Elements (SIEs).
• Utilized 12 datasets including TCGA for generating SIEs.
• Trained and evaluated deep learning models (BeiT, ResNet, Swin Transformer, ViT Transformer) for survival time classification.

Main Results:

• The enhanced Swin-T model demonstrated superior performance in predicting colorectal cancer prognosis.
• Achieved weighted precision (0.708), recall (0.692), and F1 scores (0.705).
• Identified PEX10 gene as a significant prognostic marker via attention analysis and bioinformatics.

Conclusions:

• The Gene Swin model effectively transforms RNA sequencing data into SIEs for prognosis prediction.
• This approach enables data-driven, automated classification and prediction using transcriptomic data.
• Presents a novel clinical strategy for cancer treatment and prognosis forecasting.