Predicting chemotherapy response using a variational autoencoder approach.

Area of Science:

• Computational biology
• Genomics
• Machine learning in oncology

Background:

• Transcriptome-wide RNA-seq profiles are valuable for predicting cancer chemotherapy response.
• Limited clinically annotated tumor profiles hinder fully-supervised machine learning.
• Variational autoencoders (VAEs) can generate meaningful latent features from high-dimensional data.

Purpose of the Study:

• To develop and evaluate a semi-supervised approach for predicting chemotherapy response using VAE-encoded tumor transcriptome features.
• To compare the performance of VAE features against original gene expression data and other dimensionality reduction techniques.

Main Methods:

• Applied variational autoencoder (VAE) for unsupervised encoding of tumor transcriptome data.
• Utilized regularized gradient boosted decision trees (XGBoost) for classification.
• Tested the approach on five cancer types: colon, pancreatic, bladder, breast, and sarcoma.

Main Results:

• VAE encoding preserved cancer type identity, indicating retention of biological information.
• VAE-derived features significantly improved prediction performance (AUC-ROC, AUC-PR) over original profiles, PCA, and ICA in four of five cancer types.
• The semi-supervised VAE-XGBoost model demonstrated enhanced accuracy in predicting chemotherapy response.

Conclusions:

• VAE is effective for nonlinear, low-dimensional embedding of high-dimensional 'omics' data.
• VAE features retain biologically relevant patterns, improving cancer subtyping and chemotherapy response prediction.
• This approach offers a more accurate method for transcriptome-based chemotherapy response prediction compared to traditional methods.