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

Understanding how individual cells respond to genetic or environmental perturbations is crucial for deciphering disease mechanisms and advancing precision medicine. However, predicting single-cell perturbation responses from standard single-cell RNA sequencing (scRNA-seq) data remains challenging because of cellular heterogeneity and the lack of paired pre- and post-perturbation samples in the dataset. Here, we present scVAEAT, a novel deep learning framework that integrates variational autoencoders with attention-enhanced optimal transport to accurately predict gene expression variations following perturbations, such as interferon-β stimulation or parasitic infection. scVAEAT introduces two key innovations: (1) an attention-augmented encoding module that captures multiscale feature representations through combined global and local attention mechanisms, and (2) an attention-based optimal transport (OTA) strategy that aligns unperturbed and perturbed cell states without requiring paired data. Evaluated on benchmark human datasets, including peripheral blood mononuclear cells (PBMC) and intestinal epithelial cells, scVAEAT consistently outperformed existing methods in predicting perturbation-induced expression changes, achieving mean R² scores of up to 0.97 across diverse cell types. Ablation studies confirmed the contribution of each module to model performance, whereas robustness analyses demonstrated stability across hyperparameter settings. Notably, scVAEAT accurately recapitulated known interferon-stimulated gene responses and infection-associated markers, underscoring its high biological fidelity. By in silico perturbation modeling and validation experiments using unpaired scRNA-seq data, scVAEAT offers a powerful tool for dissecting the regulatory mechanisms underlying human genetic and environmental responses, with broad implications for functional genomics and drug development.