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

Abstract

Motivation The transcriptomic data is being frequently used in the research of biomarker genes of different diseases and biological states. The most common tasks there are data harmonization and treatment outcome prediction. Both of them can be addressed via the style transfer approach. Either technical factors or any biological details about the samples which we would like to control (gender, biological state, treatment etc.) can be used as style components. Results The proposed style transfer solution is based on Conditional Variational Autoencoders, Y-Autoencoders and adversarial feature decomposition. In order to quantitatively measure the quality of the style transfer, neural network classifiers which predict the style and semantics after training on real expression were used. Comparison with several existing style-transfer based approaches shows that proposed model has the highest style prediction accuracy on all considered datasets while having comparable or the best semantics prediction accuracy. Availability https://github.com/NRshka/stvae-source Supplementary information FigShare.com (https://dx.doi.org/10.6084/m9.figshare.9925115)

Learning interpretable cellular responses to complex perturbations in high-throughput screens

“…Aside from CPA, existing methods [17,26] such as scGen [16] have also been shown capable of predicting single-cell perturbation responses when the dataset contains no combinatatorial treatment or dose-dependent perturbations. Therefore, it may be beneficial to benchmark CPA against such methods on less complicated scenarios with a few perturbations while it might not be practical considering current trend on the generation of massive perturbation studies [4,5,12].…”

“…While differential expression compares each condition separately with a control, modeling a joint latent space with a conditional variational autoencoder [17,26,27] is highly uninterpretable and not amenable to the prediction of the effects of combinations of conditions. Our goal here is to factorize the latent space of neural networks to turn them into interpretable, compositional models.…”

“…Prior work models the effects of perturbations on gene expression as separate processes: for example, differential expression compares each condition separately with control, and a conditional variational autoencoder [17,26,29]-while modeling a joint latent space-is highly unintepretable and not amenable to predict the effect of combinations of conditions. In contrast, our goal here is to factorize the latent space of neural networks to turn them into interpretable compositional models.…”

“…In contrast, deep learning (DL) models do not face these limitations. Recently, DL methods have been used to model gene expression latent spaces from scRNA-seq data [22][23][24][25], and describe and predict single-cell responses [16,17,20,26]. However, current DL-based approaches also have limitations: they model only a handful of perturbations; can be difficult to interpret; cannot handle combinatorial treatments; and cannot incorporate continuous covariates such as dose and time, or discrete covariates such as cell types, species, and patients while preserving interpretability.…”

“…In contrast, deep learning (DL) models do not face these limitations. Recently, DL has been suggested not only to model gene expression latent spaces from scRNA-seq data [22][23][24][25], but also to describe and predict single-cell responses [16,17,20,26]. Yet, current DL-based approaches model only a handful of perturbations; are not interpretable; cannot handle combinatorial treatments; and are unable to incorporate continuous (e.g.…”