LapDDPM: A Conditional Graph Diffusion Model for scRNA-seq Generation with Spectral Adversarial Perturbations

Lorenzo Bini, Stéphane Marchand-Maillet · 2025

Read LapDDPM: A Conditional Graph Diffusion Model for scRNA-seq Generation with Spectral Adversarial Perturbations by Lorenzo Bini, Stéphane Marchand-Maillet (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-13344.txt.

Introduction

Lorenzo Bini and Stéphane Marchand-Maillet introduce LapDDPM in a 2025 paper on making fake single-cell gene data. Real data of this kind are wide, sparse, and full of noise, so models that learn from them often miss the true shape of cell growth.

Their model joins graph-based learning with a score-based method that builds data step by step. During training it shakes the links of the cell graph along its main modes, which forces the model to stay sound when the data are messy. The authors also extend the model to maps of genes in tissue and to data with more than one type. Tests on five named data sets beat older tools on match, shape, and later use.

Catalog record

CREATORLorenzo Bini, Stéphane Marchand-Maillet
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-13344

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

LapDDPM: A Conditional Graph Diffusion Model for scRNA-seq Generation with Spectral Adversarial Perturbations Lorenzo Bini, Stéphane Marchand-Maillet Abstract Generating high-fidelity and biologically plausible synthetic single-cell RNA sequencing (scRNA-seq) data, especially with conditional control, is challenging due to its high dimensionality, sparsity, and complex biological variations. Existing generative models often struggle to capture these unique characteristics and ensure robustness to structural noise in cellular networks. We introduce LapDDPM, a novel conditional Graph Diffusion Probabilistic Model for robust and high-fidelity scRNA-seq generation. LapDDPM uniquely integrates graph-based representations with a score-based diffusion model, enhanced by a novel spectral adversarial perturbation mechanism on graph edge weights. Our contributions are threefold: we leverage Laplacian Positional Encodings (LPEs) to enrich the latent space with crucial cellular relationship information; we develop a conditional score-based diffusion model for effective learning and generation from complex scRNA-seq distributions; and we employ a unique spectral adversarial training scheme on graph edge weights, boosting robustness against structural variations. Extensive experiments on diverse scRNA-seq datasets demonstrate LapDDPM’s superior performance, achieving high fidelity and generating biologically-plausible, cell-type-specific samples. LapDDPM sets a new benchmark for conditional scRNA-seq data generation, offering a robust tool for various downstream biological applications. 1 Introduction Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity, enabling the characterization of gene expression at an unprecedented resolution ( Tang et al. 2009 ; Saliba et al. 2014 ) . However, the generation of scRNA-seq data is often resource-intensive, limited by sample availability, and inherently noisy due to biological and technical variations ( Kiselev et al. 2019 ) . This has spurred significant interest in developing computational methods for generating synthetic scRNA-seq data that closely mimics real biological samples ( Eraslan et al. 2019 ) . Such synthetic data can serve various purposes, including data augmentation for rare cell types, benchmarking of computational tools, and in-silico perturbation studies. Recent advancements in generative modeling, particularly Diffusion Probabilistic Models (DDPMs) ( Ho et al. 2020 ) , have shown remarkable success in synthesizing high-fidelity data across various domains. Concurrently, flow-based generative models ( Dinh et al. 2014 ; Dinh et al. 2017 ; Palma et al. 2025 ) have also gained prominence for their ability to learn complex data distributions through invertible transformations, offering exact likelihood estimation. The application of Graph Neural Networks (GNNs) has also gained traction in single-cell biology, leveraging the inherent graph-like structures of cellular relationships (e.g., cell-cell similarity networks or lineage trees) ( Zhao et al. 2021 ; Zhou et al. 2022 ) . Integrating these powerful paradigms presents a promising avenue for generating scRNA-seq data, where GNNs can capture complex cellular interactions and DPMs or flow models can model the intricate gene expression distributions. Despite these advances, existing generative models for scRNA-seq data often face several limitations. Many struggle with the unique characteristics of scRNA-seq data, such as its high dimensionality, sparsity (due to dropout events), zero-inflation, and complex, non-Gaussian distributions ( Kiselev et al. 2019 ) . Furthermore, while some methods incorporate graph information, they may lack mechanisms to ensure robustness against structural noise or subtle variations in cell-cell interaction networks, which are common in real biological contexts. The ability to generate data conditioned on specific biological metadata (e.g., cell type, tissue origin) is also crucial for practical applications, yet remains a challenge for many unconditional models. To address these limitations, we introduce LapDDPM (Laplacian Diffusion Denoising Probabilistic Model), a novel conditional generative framework for scRNA-seq data. LapDDPM integrates graph-based representations with a score-based diffusion model, further enhanced by a unique spectral adversarial perturbation mechanism applied directly to the graph’s edge weights. Our method aims to synthesize realistic scRNA-seq count matrices that faithfully reproduce the statistical properties and cellular heterogeneity of real data, conditionally on cell types. Our approach makes three key contributions: • We propose a robust graph-based data representation module for scRNA-seq, which constructs k-Nearest Neighbors (k-NN) graphs from gene expression data and incorporates Laplacian Positional Encodings (LPEs) to enrich the structural information provided to the generative model. • We develop a conditional score-based diffusion model that operates in a learned latent space. This model effectively captures the complex, high-dimensional distribution of scRNA-seq data and enables the generation of samples conditioned on specific cell types or tissue labels. • We introduce a novel spectral adversarial perturbation mechanism applied to the graph’s edge weights during training. This strategy enhances the robustness of our encoder to structural variations and noise in the cellular interaction graph, leading to more stable and reliable generative performance. Extensive experiments across diverse scRNA-seq datasets, including PBMC3K, Dentate Gyrus, Tabula Muris, and the Human Lung Cell Atlas (HLCA), comprehensively validate LapDDPM’s effectiveness. Our results demonstrate that LapDDPM consistently achieves superior quantitative metrics (RBF-kernel MMD and 2-Wasserstein distance), indicating high fidelity in capturing the underlying data distributions and cellular heterogeneity. The ability to generate conditional samples further underscores LapDDPM’s utility in synthesizing biologically plausible, cell-type-specific scRNA-seq profiles. The rest of the paper is organized as follows: Section 2 discusses related work on generative models for scRNA-seq, graph neural networks, and diffusion models. Section 3 presents the detailed methodology of LapDDPM, including its architecture, graph-based data representation, spectral adversarial perturbations, and training objective. Section 5 provides computational complexity analysis and comprehensive ablation studies. Section 4 presents the comprehensive experimental results across various datasets and tasks. Finally, Section 6 concludes with a summary of our findings and a discussion of promising future research directions. 2 Related Work Our work on LapDDPM bridges several active research areas: generative models for single-cell RNA sequencing (scRNA-seq) data, the GNNs in computational biology, the rapidly evolving field of DDPMs, and the crucial aspect of adversarial robustness in graph-structured data. 2.1 Generative Models for scRNA-seq Data The increasing availability of scRNA-seq data has driven the development of computational methods for generating synthetic cellular profiles. Early approaches often adapted models from general machine learning, such as Variational Autoencoders (VAEs) ( Lopez et al. 2018 ; Eraslan et al. 2019 ) and Generative Adversarial Networks (GANs) ( Ghosh et al. 2020 ; Li et al. 2021 ) . VAE-based models learn a low-dimensional latent representation and reconstruct gene expression, often accounting for sparsity and zero-inflation. GANs aim to learn a mapping from a simple prior distribution to the complex data distribution through an adversarial training process between a generator and a discriminator. More recently, flow-based models ( Chen et al. 2021 ) have been explored for their exact likelihood estimation and invertible mappings. While these models have demonstrated success, they often face challenges in capturing the intricate multi-modal distributions, preserving biological heterogeneity, and ensuring robustness to noise inherent in real scRNA-seq data. Our work extends these efforts by leveraging graph structures and advanced diffusion models for enhanced generative fidelity and robustness. 2.2 Graph Neural Networks in Single-Cell Biology The inherent relational nature of single-cell data, where cells can be viewed as nodes in a graph connected by biological similarity (e.g., gene expression similarity, spatial proximity, or lineage relationships), makes GNNs a natural fit for analysis. GNNs have been applied to various tasks in single-cell biology, including cell type annotation ( Wang et al. 2021 ) , trajectory inference ( Qi et al. 2021 ) , and spatial transcriptomics analysis ( Zhou et al. 2022

Read the complete text