CoupleVAE – coupled variational autoencoders for predicting perturbational single-cell RNA sequencing data

As single-cell RNA sequencing becomes more advanced, scientists are now able to analyze the genetic activity of individual cells in extraordinary detail. This is especially useful for studying how cells respond to things like drugs, infections, or other environmental changes—called perturbations. But actually performing these experiments for every possible condition can be expensive and time-consuming.

That’s where artificial intelligence steps in. A research team led by researchers at Xi’an Jiaotong University has developed a new deep learning model called CoupleVAE, which helps predict how cells will behave after a perturbation using RNA sequencing data.

$Workflow of CoupleVAE. CoupleVAE takes the gene expression of controlled cells ($\mathbf{x}_{c}$) and perturbed ($\mathbf{x}_{p}$) cells as input. The whole procedure includes three processes. Inference process: the latent representations $\mathbf{z}_{c}$ and $\mathbf{z}_{p}$ of controlled cells ($\mathbf{x}_{c}$) and perturbed cells ($\mathbf{x}_{p}$) are obtained by two encoders. Coupling process: $\mathbf{z}_{c}$ is transformed to $\hat{\mathbf{z}}_{p}$ through a nonlinear mapping, and $\mathbf{z}_{p}$ is transformed to $\hat{\mathbf{z}}_{c}$ through inverse mapping by the coupler. Generative process: gene expression of the controlled cell ($\hat{\mathbf{x}}_{c}$) and the perturbed cell ($\hat{\mathbf{x}}_{p}$) are reconstructed from $\mathbf{z}_{c}$, $\hat{\mathbf{z}}_{c}$ and $\mathbf{z}_{p}$, $\hat{\mathbf{z}}_{p}$ through the corresponding decoder. Components of this figure are created using Servier Medical Art templates, which are licensed under a Creative Commons Attribution 3.0 Unported License from https://smart.servier.com.$

CoupleVAE takes the gene expression of controlled cells (⁠xc⁠) and perturbed (⁠xp⁠) cells as input. The whole procedure includes three processes. Inference process: the latent representations zc and zp of controlled cells (⁠xc⁠) and perturbed cells (⁠xp⁠) are obtained by two encoders. Coupling process: zc is transformed to z^p through a nonlinear mapping, and zp is transformed to z^c through inverse mapping by the coupler. Generative process: gene expression of the controlled cell (⁠x^c⁠) and the perturbed cell (⁠x^p⁠) are reconstructed from zc⁠, z^c and zp⁠, z^p through the corresponding decoder.

CoupleVAE uses a type of neural network called variational autoencoders (VAEs). It works by learning patterns from both unperturbed (normal) and perturbed (affected) single-cell data, then connecting them in a way that allows it to predict what will happen to a cell’s gene activity under new conditions. In simple terms, it’s like giving the model a “before” picture of a cell and asking it to predict the “after.”

The researchers tested CoupleVAE on three datasets—focused on infection, stimulation, and even data from different species—and found it was more accurate than other methods. This means CoupleVAE could become an important tool for designing experiments, understanding disease mechanisms, and accelerating drug development, all without needing to run expensive lab tests every time.

Availability – The software is available at https://github.com/LiminLi-xjtu/CoupleVAE.

Wu Y, Liu J, Xiao Y, Zhang S, Li L. (2025) CoupleVAE: coupled variational autoencoders for predicting perturbational single-cell RNA sequencing data. Briefings in Bioinformatics 26(2): bbaf126. [article]