VTrans: A VAE-Based Pre-Trained Transformer Method for Microbiome Data Analysis.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Predicting cancer patient survival and risk is crucial for understanding microbial composition.
• Deep learning shows promise for analyzing patient survival risks from microbial data.
• Limited sample sizes and high dimensionality in cancer datasets cause overfitting in deep learning models.

Purpose of the Study:

• To propose a deep learning model, VTrans, for predicting cancer patient survival risk using microbial data.
• To address overfitting and improve data representation in deep learning models for cancer survival prediction.
• To explore the potential of VTrans for integrating microbial multi-omics data.

Main Methods:

• Developed VTrans, a deep learning model combining Transformer encoder and variational autoencoder (VAE).
• Employed pretraining and fine-tuning strategies to enhance model performance with limited data.
• Assessed VTrans on three cancer datasets from The Cancer Genome Atlas Program.

Main Results:

• VTrans outperformed conventional machine learning and other deep learning models in predicting survival risk.
• Pretraining significantly boosted VTrans performance.
• VAE encoding proved more effective than positional encoding for data representation.
• Saliency maps identified key microbes contributing to classification.

Conclusions:

• VTrans effectively predicts cancer patient survival risk using microbial data.
• Pretraining and VAE encoding are crucial for VTrans's superior performance.
• The model offers insights into microbe-specific contributions to cancer survival prediction.