Abstract:Single-cell data is driving new insights into the spatiotemporal dynamics of cells and individual disease susceptibility. However, accurately identifying cell states across diverse cohorts remains challenging, as both biological variation and technical biases cause distributional shifts in the data. Separating these effects is crucial for capturing cellular heterogeneity and ensuring interpretability. To address this, we developed inVAE, a conditionally invariant deep generative model based on variational autoencoders. inVAE models the latent space as a combination of invariant variables, encoding true biological signals, and spurious variables, capturing technical biases. By conditioning the prior distribution of cells on biological covariates, such as disease variants, inVAE identifies high-resolution cell states in the invariant representation. Enforcing independence between the two representations disentangles biological signals from noise, enabling a more interpretable and generalizable model with a causal semantic. inVAE outperformed existing methods across four human cellular atlases of the human heart and lung, while uncovering novel cell states. It precisely stratified cell atlas donors based on the genetic impact of pathogenic variants, and excelled in predicting cell types and disease in unseen data, proving its generalizability as a reference model for label transfer. Furthermore, inVAE accurately identified temporal cell states and trajectories from developmental datasets, and captured spatial cell states in a spatially resolved atlas. In summary, inVAE provides a powerful method for integrating multivariate single-cell transcriptomics data. By leveraging prior knowledge such as metadata, it effectively accounts for biological variation and improves latent space interpretability by disentangling biological and technical sources of variation. These capabilities enable deeper insights into cellular heterogeneity and its role in disease progression.

Out-of-distribution Prediction with Disentangled Representations for Single-Cell RNA Sequencing Data

Conditional Out-of-distribution Generation for Unpaired Data Using Transfer VAE.

Unsupervised Deep Disentangled Representation of Single-Cell Omics

Learning interpretable latent autoencoder representations with annotations of feature sets

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

Prediction of context-specific regulatory programs and pathways using interpretable deep learning

Conditional out-of-sample generation for unpaired data using trVAE

Conditionally Invariant Representation Learning for Disentangling Cellular Heterogeneity

scVAE: variational auto-encoders for single-cell gene expression data

scNODE : Generative Model for Temporal Single Cell Transcriptomic Data Prediction

A deep adversarial variational autoencoder model for dimensionality reduction in single-cell RNA sequencing analysis

Evaluating unsupervised disentangled representation learning for genomic discovery and disease risk prediction

Disentangling Factors of Variation in Deep Representations Using Adversarial Training.

eDOC: Explainable Decoding Out-of-domain Cell Types with Evidential Learning

Scgen Predicts Single-Cell Perturbation Responses

Network Embedding-Based Representation Learning for Single Cell RNA-seq Data

A Supervised Contrastive Framework for Learning Disentangled Representations of Cell Perturbation Data

Biologically Interpretable VAE with Supervision for Transcriptomics Data Under Ordinal Perturbations

inVAE: Conditionally invariant representation learning for generating multivariate single-cell reference maps

Interpretable factor models of single-cell RNA-seq via variational autoencoders

VASC: Dimension Reduction and Visualization of Single-cell RNA-seq Data by Deep Variational Autoencoder.