scDiffusion-X

scDiffusion-X is a latent diffusion model for generating and translating single-cell multi-omics data, developed by Erpai Luo and colleagues at Tsinghua University. Single-cell multi-omics assays that jointly profile complementary molecular layers — such as gene expression and chromatin accessibility in the same cell — are powerful but constrained by cost, throughput, and incomplete modality coverage. scDiffusion-X addresses this bottleneck by learning to synthesize high-fidelity in-silico multi-omics profiles and to predict one modality directly from another.

The model extends the earlier scRNA-seq generator scDiffusion from the same group to the multi-modal setting. Its central innovation is a Dual-Cross-Attention (DCA) module that adaptively learns hidden relationships between molecular modalities during the diffusion process, providing a more flexible and interpretable coupling than fixed fusion schemes. Beyond generation, the authors repurpose the DCA module through a gradient-based interpretation framework, turning the generative model into a discovery tool for cell-type-specific gene regulatory networks (GRNs).

Originally posted on bioRxiv in February 2025 and subsequently published in Nature Communications, scDiffusion-X ships with a pretrained "miniatlas" checkpoint, making it directly usable for data augmentation and cross-modality prediction without training from scratch.

#Key Features

• Multi-omics generation: Synthesizes joint single-cell profiles across molecular modalities while preserving cellular heterogeneity and global data structure, with strong scalability to large datasets.
• Dual-Cross-Attention module: Adaptively models intricate relationships between modalities within the latent diffusion backbone, offering a more flexible and interpretable alternative to conventional fusion approaches.
• Cross-modality translation: Predicts one molecular modality from another with high fidelity, enabling imputation of unmeasured layers and providing calibrated uncertainty quantification.
• Gene regulatory network inference: A gradient-based interpretation framework transforms the DCA module into a discovery tool that infers comprehensive, cell-type-specific heterogeneous GRNs.
• Pretrained miniatlas checkpoint: A ready-to-use model and the filtered training dataset are distributed alongside the MIT-licensed code, lowering the barrier to adoption.

#Technical Details

scDiffusion-X operates as a latent diffusion model: raw multi-omics measurements are encoded into compact latent representations, a diffusion process is trained to denoise these latents, and the Dual-Cross-Attention module mediates information exchange between modality-specific streams at each denoising step. This design lets the model capture cross-modal dependencies rather than treating modalities independently, which is essential for realistic modality translation and for quantifying predictive uncertainty. The authors benchmark scDiffusion-X against existing multi-omics simulators, reporting improved fidelity in reproducing cellular heterogeneity and global structure across paired datasets. The implementation is released under the MIT license, with the pretrained miniatlas model and the openproblem_filtered training data available for download and documentation hosted on Read the Docs.

#Applications

scDiffusion-X benefits computational biologists who need realistic synthetic multi-omics data to augment limited experimental cohorts, balance rare cell populations, or stress-test downstream integration methods. Its modality-translation capability lets researchers impute an unmeasured layer — for example inferring chromatin accessibility from transcriptomes — when only single-modality data are available, while the accompanying uncertainty estimates flag low-confidence predictions. The gradient-based GRN inference framework further supports mechanistic studies, allowing investigators to dissect cell-type-specific regulatory relationships and to generate hypotheses about perturbation responses.

#Impact

scDiffusion-X demonstrates that latent diffusion, coupled with an adaptive cross-attention mechanism, can serve as a unified engine for multi-omics generation, translation, and interpretation rather than for a single narrow task. By combining state-of-the-art generative modeling with an interpretability layer that surfaces regulatory structure, it bridges data simulation and biological discovery in single-cell genomics. Publication in Nature Communications and release of a pretrained checkpoint under a permissive license position the model for adoption in single-cell method development. As with any generative simulator, downstream biological conclusions depend on the coverage and quality of the training data, and synthetic profiles complement rather than replace experimental measurement.