scPreGAN

scPreGAN predicts single-cell expression responses to perturbations by using a deep generative model to infer perturbed single-cell RNA-sequencing (scRNA-seq) profiles from unperturbed data.

Key Features:

• Autoencoder and GAN integration: scPreGAN combines an autoencoder with a generative adversarial network (GAN) to model and generate perturbed single-cell expression profiles.
• Feature extraction across conditions: The autoencoder extracts common features from both unperturbed and perturbed datasets to capture underlying data structure.
• Generative prediction: The GAN component uses the extracted features to predict how cells would express genes under perturbation conditions.
• Improved prediction accuracy: The model demonstrates superior prediction accuracy compared to three state-of-the-art methods by modeling complex distributions of cell expression data.
• Fidelity to real data: Predicted expression abundances closely mirror observed expression in real perturbed datasets.
• Experimental validation: Performance has been validated on three real-world datasets.

Scientific Applications:

• Drug response prediction: Predicts how different cell types may respond to specific drugs, supporting in silico evaluation of treatment effects prior to administration.
• Biological mechanism exploration: Aids investigation of cellular responses and mechanisms at single-cell resolution in fields such as oncology, immunology, and developmental biology.

Methodology:

scPreGAN uses an autoencoder to extract common features from unperturbed and perturbed scRNA-seq data and a GAN that leverages those features to generate predicted perturbed single-cell expression profiles.