scGND: Graph neural diffusion model enhances single-cell RNA-seq analysis

Yuchen Liu, Anqi Zou, Simon Liang Lu, Jou-Hsuan Lee, Juexin Wang, Chengqi ZhangbioRxiv 2024 · DOI

Abstract

Single-cell sequencing technologies have played a pivotal role in advancing biomedical research over the last decade. With the evolution of deep learning, a variety of models based on deep neural networks have been developed to improve the precision of single-cell RNA sequencing (scRNA-seq) analysis from multiple angles. However, deep learning models currently used in scRNA-seq analysis frequently suffer from a lack of interpretability. In this study, we present a novel physics-informed graph generative model, termed Single Cell Graph Neural Diffusion (scGND). This model is founded on solid mathematical concepts and provides enhanced interpretability. Unlike methods that focus solely on gene expression in individual cells, scGND concentrates on the cell-cell interaction graph, incorporating two key physical concepts: local and global equilibrium. We show that achieving a balance between local and global equilibrium significantly improves the geometric properties of the graph, aiding in the extraction of inherent biological insights from the cell-cell interaction graph at multiple scales. The effectiveness of scGND has been proven through benchmark tests involving five independent scRNA-seq datasets from various tissues and species. scGND consistently achieves better or comparable results comparing with several established competitors in both clustering and trajectory analysis. scGND represents a comprehensive generative model based on cell graph diffusion, demonstrating considerable promise for both theoretical and practical applications in scRNA-seq data analysis.

Introduction

Analyzing single-cell RNA sequencing (scRNA-seq) data typically involves two distinct perspectives: the discrete viewpoint, which categorizes cells into distinct types, and the continuous viewpoint, which examines transitions and developmental trajectories. While traditional tools like Seurat or Scanpy offer interpretability, they often struggle with high noise levels and lack the flexibility of deep learning. Conversely, many existing deep learning models function as "black boxes," providing high performance without explaining the underlying biological rationale or the geometric relationships between cells.

The Single Cell Graph Neural Diffusion (scGND) model addresses these challenges by employing a physics-informed framework grounded in diffusion theory. By treating scRNA-seq data as a dynamic diffusion process, scGND makes the latent space more meaningful and provides a unified approach to both discrete and continuous analysis.

Methods

The core of scGND is a graph neural diffusion algorithm that processes cell-cell interaction graphs. The workflow consists of several key stages:

  1. Graph Construction: Raw gene expression counts are converted into a cell-cell interaction graph (specifically a K-Nearest Neighbors graph), where nodes represent cells and edges represent similarities.
  2. Physics-Informed Diffusion: Instead of standard neural network layers, scGND utilizes a diffusion equation derived from heat physics to simulate the spread of information across the cell graph over time.
  3. Learnable Diffusivity: An attention mechanism acts as "diffusivity," allowing the network to learn which cell-cell connections are biologically significant (high diffusivity) and which are noise (low diffusivity).
  4. Equilibrium Dynamics: The model leverages two physical concepts: Local Equilibrium, which enhances similarities within clusters for better cell type identification, and Global Equilibrium, which highlights inter-cluster connections to reveal developmental trajectories.
The architecture of the scGND network, showing the encoder-decoder structure and the diffusion layers.

Results

The performance of scGND was validated across several independent benchmark datasets, focusing on both clustering and trajectory inference.

Clustering Performance

On datasets such as Klein, Zeisel, and Tabula Muris, scGND consistently outperformed established competitors, including Seurat, Scanpy, and the transformer-based scGPT. For instance, on the Zeisel dataset, scGND achieved an Adjusted Rand Index (ARI) of 0.86, significantly higher than Seurat's 0.72.

Trajectory Accuracy

In trajectory inference tests, scGND demonstrated exceptional accuracy. It perfectly reproduced the reference lineage for the Oligodendrocyte dataset (Accuracy = 1.0) and achieved the highest accuracy on the complex Planaria dataset (0.69), surpassing Monocle3 (0.61).

Robustness

The model maintained high performance even when the dimensionality of the latent space was significantly reduced (from 50D to 15D), indicating that it captures stable and essential biological features.

Key Insights and Discussion

A primary advantage of scGND is the use of "diffusion time" as a control parameter. Shorter diffusion times emphasize local equilibrium, which is ideal for clustering and identifying discrete cell types. Longer times emphasize global equilibrium, facilitating the analysis of continuous developmental trajectories.

Visualizing the transition from non-equilibrium data to local and global equilibrium states over diffusion time.

Furthermore, by focusing on the geometry of the cell graph rather than just gene counts, scGND can capture "fractal" structures—sub-clusters within clusters that represent subtle biological states. This demonstrates that cell types and trajectories are not competing viewpoints but rather different stages of the same physical process.

Limitations and Future Work

Despite its strengths, scGND is computationally intensive and may face memory limitations with datasets exceeding 100,000 cells. Future developments will focus on extending the model to handle temporal (time-series) data and spatial transcriptomics. Additionally, the researchers aim to utilize learned attention scores to simulate gene regulatory networks, potentially uncovering the mechanisms controlling cell fate.

Summary updated

Abstract

Single-cell sequencing technologies have played a pivotal role in advancing biomedical research over the last decade. With the evolution of deep learning, a variety of models based on deep neural networks have been developed to improve the precision of single-cell RNA sequencing (scRNA-seq) analysis from multiple angles. However, deep learning models currently used in scRNA-seq analysis frequently suffer from a lack of interpretability. In this study, we present a novel physics-informed graph generative model, termed Single Cell Graph Neural Diffusion (scGND). This model is founded on solid mathematical concepts and provides enhanced interpretability. Unlike methods that focus solely on gene expression in individual cells, scGND concentrates on the cell-cell interaction graph, incorporating two key physical concepts: local and global equilibrium. We show that achieving a balance between local and global equilibrium significantly improves the geometric properties of the graph, aiding in the extraction of inherent biological insights from the cell-cell interaction graph at multiple scales. The effectiveness of scGND has been proven through benchmark tests involving five independent scRNA-seq datasets from various tissues and species. scGND consistently achieves better or comparable results comparing with several established competitors in both clustering and trajectory analysis. scGND represents a comprehensive generative model based on cell graph diffusion, demonstrating considerable promise for both theoretical and practical applications in scRNA-seq data analysis.
BibTeX
@misc{liu2024scgnd,
  author = {Yuchen Liu and Anqi Zou and Simon Liang Lu and Jou-Hsuan Lee and Juexin Wang and Chengqi Zhang},
  title = {{scGND: Graph neural diffusion model enhances single-cell RNA-seq analysis}},
  year = {2024},
  doi = {10.1101/2024.01.28.577667},
  howpublished = {bioRxiv},
}
Figure