trVAE

trVAE performs conditional generation of high-dimensional samples from low-dimensional descriptors to predict condition-specific responses in data such as single-cell RNA-seq.

Key Features:

• Conditional Out-of-Distribution Generation: Generates out-of-distribution samples by learning compact joint distributions across multiple conditions, addressing limitations of conventional CVAEs.
• Maximum Mean Discrepancy (MMD): Incorporates MMD in the decoder layer following the bottleneck as a regularizer to improve reconstruction and enable transformations across conditions.
• Improved Generalization: Demonstrates increased robustness and accuracy relative to CVAEs, handling minority classes and multiple conditions in high-dimensional data such as single-cell RNA sequencing.
• Application in Cellular Biology: Benchmarked on high-dimensional images and single-cell RNA-seq, it predicts cellular responses to perturbations and diseases and improves prediction of cell-type-specific genes by 65%.
• Quantitative Improvements: Improves Pearson correlations for high-dimensional estimated means and variances with ground truth from 0.89 to 0.97 and from 0.75 to 0.87, respectively.

Scientific Applications:

• Single-cell RNA-seq Analysis: Predicts cellular responses to treatments and diseases and cell-type-specific expression changes from single-cell gene expression data.
• Style Transfer Problems: Frames condition-to-condition transformation as a style-transfer problem to manage and predict changes across experimental conditions.

Methodology:

Extends Conditional Variational Autoencoders by explicitly relating conditions during training; incorporates MMD regularization in the decoder layer following the bottleneck; implemented in Keras (TensorFlow < 2.0) and trained on datasets such as Kang and Haber.