SynCellNet: generative AI framework for single-cell RNA sequencing data generation and validation

Abstract

Generative AI plays a pivotal role in addressing the challenges of data scarcity and class imbalance in single-cell transcriptomics. In this work, we present SynCellNet, a generative AI technique for synthesis of single-cell gene expression profiles. SynCellNet can be used to enrich real data by producing statistically similar synthetic AI-generated data. This strategy can also alleviate the need to share real datasets between labs, thus addressing privacy and anonymity concerns. Instead of operating directly on raw gene expressions, our approach leverages Genomaps, a structured 2D representation that encodes gene–gene interactions, enabling the application of deep convolutional architectures while preserving the genomic structure. SynCellNet synthesizes Genomaps that are conditioned on cell phenotype, and includes a reverse transformation module to recover gene expression values from synthetic Genomaps, followed by a Gaussian Copula post-processing step that restores the marginal count distributions and gene–gene correlation structure of the real data. To assess the validity of our synthesized data, we feed them to a convolutional neural network (CNN) based classifier that is trained on two independent datasets, each with two classes (patient-derived colorectal organoid (PDO) and PBMC). The test results indicate an average classification accuracy of 90.16%. Structural similarity and reconstruction quality of synthetic data are further validated through SSIM and PSNR metrics, with the strongest agreement on the Stem class of the PDO dataset, where real-to-synthetic comparisons reach SSIM values exceeding 0.90 and PSNR up to 30.36 dB, the highest fidelity of any class. PSNR remains consistently high, around 28–30 dB, across all four classes on both datasets, confirming high visual and structural alignment. Additionally, the CDFs of real and synthetic samples demonstrate significant statistical similarity in both the Genomap and gene expression domains across all classes. A comprehensive benchmarking study against scGAN and scVI, two generative AI baselines, covering gene-level statistics, cell-level structure, and differential expression conservation is conducted. The study shows that SynCellNet alone faithfully captures the cell-level structure of the real data, while the Gaussian Copula step further elevates SynCellNet to match or outperform both baselines across the majority of metrics on both datasets. These results confirm that SynCellNet, together with the Gaussian Copula step, generates class-specific, biologically meaningful synthetic data, offering a robust solution for data augmentation and downstream single-cell analysis.