Hierarchical Graph Latent Diffusion Model for Molecule Generation
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Hierarchical Graph Latent Diffusion, Molecule Generation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Recently, generative models based on the diffusion process have emerged as a promising direction for automating the design of molecules. However, directly adding continuous Gaussian noise to discrete graphs leads to the problem of the final noisy data not conforming to the standard Gaussian distribution. Current graph diffusion models either corrupt discrete data through a transition matrix or relax the discrete data to continuous space for the diffusion process. These approaches not only require significant computation resources due to the inclusion of the bond type matrix but also cannot easily perform scalable conditional generation, such as adding cross-attention layers, due to the lack of embedding representations. In this paper, we first introduce the Graph Latent Diffusion Model (GLDM), a novel variant of latent diffusion models that overcomes the mismatch problem of continuous diffusion space and discrete data space. Meanwhile, the latent diffusion framework avoids the issues of computational resource consumption and lack of embeddings for conditional generation faced by current graph diffusion models. However, it only utilizes graph-level embeddings for molecule generation, losing node-level and structural information. Therefore, we further ex- tend the GLDM to the Hierarchical Graph Latent Diffusion Model (HGLDM). By including node embeddings and subgraph embeddings that contain structural in- formation, our model significantly reduces computation time compared to the cur- rent graph diffusion models. We evaluate our model on three benchmarks through unconditional generation and conditional generation tasks, which demonstrate its superior performance.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6846
Loading