Improving scDiffusion with Sparsity-Biased Classifier-Free Guidance

2026-07-31

AI Summary

This work addresses a key limitation in current diffusion models for single-cell RNA sequencing (scRNA-seq) data generation: the unconditional branch retains excessive gene-specific structure, which diminishes the efficacy of classifier-free guidance (CFG). To overcome this, the authors propose Sparse-Biased CFG (SB-CFG), which replaces the standard unconditional prediction during sampling with an under-informed reference that preserves only sparse statistical information. This enhances the contrast between conditional and unconditional outputs, thereby strengthening guidance without requiring model retraining. SB-CFG represents the first approach to incorporate intentionally degraded sparse references into the guidance mechanism of scRNA-seq diffusion models. Experiments across five public datasets demonstrate that SB-CFG consistently outperforms standard CFG in preserving marker gene expression fidelity, maintaining cell-type consistency, and retaining data sparsity.

Abstract

Single-cell RNA sequencing (scRNA-seq) has become an essential tool in modern cellular biology, and generating accurate synthetic scRNA-seq data is becoming increasingly important. Although diffusion models have achieved promising results in conditional scRNA-seq generation, existing guidance strategies, including classifier guidance and classifier-free guidance (CFG), rely on an unconditional branch trained to approximate the true marginal distribution, which may retain substantial gene-specific structure and limit guidance effectiveness. Inspired by recent work showing that diffusion models can be effectively guided using intentionally degraded references, we propose a sparsity-biased classifier-free guidance (SB-CFG) strategy for scRNA-seq generation. Rather than approximating the assumed "neutral" marginal distribution, SB-CFG introduces a deliberately under-informative sparse reference for the unconditional branch, removing gene identity while preserving only coarse sparsity statistics. This "bad" reference amplifies the contrast between conditional and unconditional predictions, leading to stronger and more effective guidance during sampling. We evaluated SB-CFG as a training-free sampling modification on five publicly available scRNA-seq datasets. Experimental results demonstrate consistent improvements over standard CFG-based sampling in terms of marker gene expression fidelity, cell-type consistency, and sparsity preservation, indicating that SB-CFG better captures biologically meaningful gene expression patterns.

Problem

• scRNA-seq
• diffusion models
• classifier-free guidance
• sparsity
• conditional generation

Innovation

• sparsity-biased guidance
• classifier-free guidance
• diffusion models
• single-cell RNA sequencing
• synthetic data generation