scVAEAT: An Integrative Attention-Augmented Variational Autoencoder for Predicting Single-Cell Perturbation...

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Predicting single-cell responses to perturbations is vital for understanding disease and developing precision medicine.
• Challenges include cellular heterogeneity and the lack of paired pre- and post-perturbation samples in single-cell RNA sequencing (scRNA-seq) data.

Purpose of the Study:

• To present scVAEAT, a novel deep learning framework for accurate prediction of gene expression variations after cellular perturbations.
• To address the limitations of existing methods in handling unpaired scRNA-seq data.

Main Methods:

• Integration of variational autoencoders with attention-enhanced optimal transport.
• Development of an attention-augmented encoding module for multiscale feature representation.
• Implementation of an attention-based optimal transport (OTA) strategy to align unperturbed and perturbed cell states without paired data.

Main Results:

• scVAEAT achieved high accuracy in predicting perturbation-induced expression changes on human PBMC and intestinal epithelial cell datasets, outperforming existing methods (mean R² up to 0.97).
• Model performance was validated through ablation studies and robustness analyses.
• The framework accurately recapitulated known interferon-stimulated gene responses and infection-associated markers, demonstrating high biological fidelity.

Conclusions:

• scVAEAT provides a powerful computational tool for dissecting regulatory mechanisms of human genetic and environmental responses using unpaired scRNA-seq data.
• The framework has significant implications for functional genomics, drug development, and advancing precision medicine.