Style transfer with variational autoencoders is a promising approach to RNA-Seq data harmonization and analysis

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transcriptomic data analysis is crucial for identifying disease biomarkers and predicting treatment outcomes.
• Current methods often require data harmonization and accurate outcome prediction.
• Style transfer offers a promising approach to address these challenges by controlling sample-specific attributes.

Purpose of the Study:

• To develop and evaluate a novel style transfer approach for transcriptomic data.
• To enable effective data harmonization and treatment outcome prediction.
• To control for technical and biological factors in transcriptomic datasets.

Main Methods:

• The study proposes a style transfer solution utilizing Conditional Variational Autoencoders (Y-Autoencoders) and adversarial feature decomposition.
• Neural network classifiers trained on real expression data were employed for quantitative evaluation of style transfer quality.
• The approach allows for the control of various sample attributes, such as gender, biological state, and treatment.

Main Results:

• The proposed model achieved the highest style prediction accuracy across all tested datasets.
• The model demonstrated comparable or superior semantics prediction accuracy compared to existing methods.
• This indicates successful control over style components while preserving essential biological information.

Conclusions:

• The developed style transfer method effectively harmonizes transcriptomic data and predicts treatment outcomes.
• The approach offers a robust solution for controlling sample-specific attributes in high-dimensional transcriptomic data.
• This work advances the application of style transfer in precision medicine and biomarker discovery.