scDD: scRNA-seq dataset distillation in latent codes with single-step conditional diffusion generator

Abstract

The single-cell RNA sequencing (scRNA-seq) technology has profiled hundreds of millions of human cells across organs, diseases, developmental stages and perturbations. However, the original scRNA-seq datasets are redundant and have an ever-increasing data scale, which pose significant challenges for cross-platform data sharing and scalable foundation model construction. To address this, we propose novel dataset distillation technology in scRNA-seq analysis tasks to distill/condense the original scRNA-seq dataset into a synthetic, smaller and discriminative dataset. Unfortunately, the synthetic datasets distilled by existing dataset distillation methods have inferior cross-architecture generalization and inter-class discriminability. In light of this, (1) We propose scDD, a scRNA-seq dataset distillation framework in latent codes, which distills the original dataset information into a compact latent space, and generates a synthetic dataset with cross-architecture generalization by avoiding direct disruption to gene expression values. Then, (2) We propose a single-step conditional diffusion generator named SCDG within the scDD framework, through high-fidelity generation and category-condition guidance of the generator, SCDG ensures that the generated synthetic dataset retains scRNA-seq data characteristics and inter-class discriminability. Finally, we propose a comprehensive and robust benchmark to evaluate the performance of scRNA-seq dataset distillation in different data analysis tasks. It is validated that our proposed method can achieve $7.61\%$ absolute and $15.70\%$ relative improvement over previous state-of-the-art methods on average across task. In particular, our method also achieves an average $26.51\%$ absolute improvement in cross-architecture generalization.