[Paper Review] A Survey on Graph Diffusion Models: Generative AI in Science for Molecule, Protein and Material

Mengchun Zhang, Maryam Qamar+5 more|arXiv (Cornell University)|Jan 1, 2023

Bioinformatics and Genomic Networks26 citations

TL;DR

This survey provides a comprehensive overview of graph diffusion models in generative AI for scientific applications, focusing on molecule, protein, and material generation. It reviews foundational graph generation methods, details the mechanisms of diffusion models on graphs, and evaluates their performance, challenges, and open research questions in scientific AI.

ABSTRACT

Diffusion models have become a new SOTA generative modeling method in various fields, for which there are multiple survey works that provide an overall survey. With the number of articles on diffusion models increasing exponentially in the past few years, there is an increasing need for surveys of diffusion models on specific fields. In this work, we are committed to conducting a survey on the graph diffusion models. Even though our focus is to cover the progress of diffusion models in graphs, we first briefly summarize how other generative modeling methods are used for graphs. After that, we introduce the mechanism of diffusion models in various forms, which facilitates the discussion on the graph diffusion models. The applications of graph diffusion models mainly fall into the category of AI-generated content (AIGC) in science, for which we mainly focus on how graph diffusion models are utilized for generating molecules and proteins but also cover other cases, including materials design. Moreover, we discuss the issue of evaluating diffusion models in the graph domain and the existing challenges.

Motivation & Objective

• To address the growing need for specialized surveys on diffusion models in graph-structured data, particularly in scientific domains.
• To provide a systematic review of graph generation techniques based on autoregressive, VAE, GAN, and especially diffusion models.
• To evaluate the performance of graph diffusion models in generating molecules, proteins, and materials using standardized metrics.
• To identify and discuss key challenges such as scalability, irregular graph structure, evaluation criteria, and interpretability in graph diffusion modeling.
• To guide researchers toward open problems and future directions in applying diffusion models to scientific graph generation tasks.

Proposed method

• Reviews prior graph generation methods including autoregressive models (e.g., GraphRNN), VAEs, normalizing flows, and GANs, highlighting their limitations in capturing structural invariance.
• Introduces the core mechanism of diffusion models, including forward and reverse denoising processes, score-based modeling, and denoising score matching.
• Applies diffusion modeling to graph data by defining node and edge noise schedules, and uses score-based generative modeling on graph structures.
• Employs a unified framework for graph generation that models the joint distribution of nodes and edges through iterative denoising in latent space.
• Utilizes graph neural networks (GNNs) to estimate the score function in the reverse diffusion process, enabling end-to-end training.
• Evaluates models using statistical metrics (e.g., KLD, MMD), classifier-based metrics (e.g., FID), and task-specific metrics like validity, uniqueness, and novelty.

Experimental results

Research questions

• RQ1 How do diffusion models compare to traditional graph generation methods (e.g., VAEs, GANs, autoregressive models) in generating realistic molecular and protein graphs?
• RQ2 What are the key architectural and training components that enable effective graph diffusion modeling on irregular, variable-sized graphs?
• RQ3 How can evaluation metrics such as validity, uniqueness, and novelty be meaningfully applied to assess the quality of generated scientific graphs?
• RQ4 What are the major scalability and structural challenges in applying diffusion models to large-scale or highly irregular graphs?
• RQ5 To what extent can diffusion models generate diverse, chemically valid, and functionally relevant molecules and materials?

Key findings

• Graph diffusion models outperform traditional methods in generating diverse and chemically valid molecules, with reported validity rates exceeding 95% in some models.
• The use of score-based generative modeling enables high-fidelity generation of molecular and protein structures by learning complex, high-order dependencies.
• Evaluation using FID and classifier-based metrics shows that generated graphs from diffusion models achieve high similarity to real graphs in latent space.
• Despite strong performance, current diffusion models face scalability issues due to quadratic complexity in reverse sampling, especially for large graphs.
• The lack of universal evaluation metrics remains a key challenge, as statistical and classifier-based metrics do not always correlate with downstream chemical or biological utility.
• Interpretability of the score function in graph diffusion remains limited, making it difficult to analyze or debug model behavior in scientific contexts.