Conditional out-of-distribution generation for unpaired data using transfer VAE.

Area of Science:

• Machine Learning
• Computational Biology
• Genomics

Background:

• Generative models excel at sampling high-dimensional data based on low-dimensional descriptors.
• Conditional variational autoencoders (CVAEs) struggle with out-of-distribution generation due to difficulties in learning compact joint distributions across conditions.
• CVAEs lack explicit incentives to learn compact representations relating different conditions during training.

Purpose of the Study:

• To develop a generative model that overcomes the limitations of CVAEs in handling out-of-distribution samples.
• To improve the generalization capabilities of generative models by explicitly relating different conditions.
• To enhance the accuracy and robustness of generative models for high-dimensional data analysis, particularly in single-cell RNA sequencing.

Main Methods:

• Introduced transfer VAE (trVAE), a novel generative model building upon the CVAE framework.
• Implemented a distribution matching technique using maximum mean discrepancy in the decoder layer.
• Applied trVAE to high-dimensional image and single-cell RNA sequencing datasets for benchmarking.

Main Results:

• trVAE demonstrated significantly improved generalization by matching distributions across conditions.
• Benchmarking showed higher robustness and accuracy compared to existing methods on image and single-cell RNA-seq data.
• Achieved substantial improvements in predicting cell-type-specific responses to perturbations in gene expression data, enhancing predictions by 65% for key genes.

Conclusions:

• trVAE effectively addresses the out-of-distribution generation problem in conditional generative models.
• The model offers enhanced accuracy and robustness for high-dimensional data, with significant implications for single-cell genomics.
• The developed trVAE model and its reproducibility resources are publicly available.