AI
New AI Method Generates Realistic Synthetic Cell Data While Resisting Noise
Researchers introduce LapDDPM, a diffusion model that creates high-quality synthetic single-cell RNA data while remaining robust against structural variability and technical artifacts.
Creating synthetic single-cell RNA sequencing data that mimics real biology remains a major hurdle in computational biology research. Most existing approaches struggle to accurately map the intricate geometry of how cells differentiate or fall apart when real-world noise and variations enter the picture. Researchers have now proposed LapDDPM, a graph-based diffusion model designed to generate high-fidelity and biologically plausible cell data while staying stable against such disruptions.
How It Works#
LapDDPM combines graph diffusion probabilistic modeling with a novel technique called spectral adversarial perturbation. Rather than treating cellular data as simple point clouds, the method represents cells and their relationships as a graph, embedding biological constraints directly into the model's structure. During training, the system deliberately adds controlled noise to the edges of this graph along its main spectral modes—essentially stress-testing itself to learn which relationships matter most. This acts as a form of distributionally robust optimization, making the model invariant to unwanted structural noise while preserving genuine biological signals.
Broader Applications#
Beyond standard single-cell RNA analysis, the team extended LapDDPM to handle spatial transcriptomics data, which tracks where genes are expressed within tissue samples, as well as multi-modal datasets that combine multiple measurement types. In each case, data generation is framed as a robust inverse problem on cellular graphs—a more principled way to handle the inherent uncertainty in the data.
Validation and Results#
The researchers evaluated LapDDPM on several benchmark datasets including PBMC3K (immune cells), Dentate Gyrus (brain tissue), HLCA (lung cells), Visium (spatial data), and 10x Multiome (multi-modal data). Compared to existing state-of-the-art methods, LapDDPM showed substantial improvements in distribution matching, preserving the underlying cellular manifold structure, and generating cell states that remain useful for downstream biological tasks.
