scREPA: Predicting single-cell perturbation responses with cycle-consistent representation alignment.

Area of Science:

• Computational Biology
• Genomics
• Biotechnology

Background:

• Modeling cellular responses to perturbations is crucial for disease research and drug development.
• Existing computational methods struggle with sparse, noisy, and high-dimensional single-cell RNA sequencing (scRNA-seq) data.
• Single-cell foundation models (scFMs) offer improved representations for complex biological data.

Purpose of the Study:

• To develop a novel framework, scREPA, for robust single-cell perturbation prediction.
• To leverage pretrained scFMs to enhance variational autoencoder (VAE) based models for scRNA-seq data.
• To improve the generalization and accuracy of predicting cellular responses.

Main Methods:

• Proposed scREPA framework aligning VAE latent embeddings with scFM representations.
• Introduced Cycle-Consistent Representation Alignment for dual consistency in VAE-generated data.
• Utilized optimal transport for aligning unpaired control and perturbed data distributions during inference.

Main Results:

• scREPA significantly outperforms existing methods in predicting differentially expressed genes and whole-transcriptome responses.
• Demonstrated strong generalization capabilities across diverse datasets, unseen conditions, and cross-study settings.
• Maintained robust performance even with noisy or limited scRNA-seq data.

Conclusions:

• scREPA provides a powerful and generalizable approach for single-cell perturbation prediction.
• The representation alignment strategy effectively addresses challenges in scRNA-seq data analysis.
• This framework advances the understanding of cellular heterogeneity and perturbation dynamics.