Zhangyang Gao*, Daize Dong*, Cheng Tan, Jun Xia, Bozhen Hu, Stan Z. Li

Published on The 41st International Conference on Machine Learning (ICML 2024).

Can we model Non-Euclidean graphs as pure language or even Euclidean vectors while retaining their inherent information? The Non-Euclidean property have posed a long term challenge in graph modeling. Despite recent graph neural networks and graph transformers efforts encoding graphs as Euclidean vectors, recovering the original graph from vectors remains a challenge. In this paper, we introduce GraphsGPT, featuring an Graph2Seq encoder that transforms Non-Euclidean graphs into learnable GraphWords in the Euclidean space, along with a GraphGPT decoder that reconstructs the original graph from GraphWords to ensure information equivalence. We pretrain GraphsGPT on $100$M molecules and yield some interesting findings:

• The pretrained Graph2Seq excels in graph representation learning, achieving state-of-the-art results on $8/9$ graph classification and regression tasks.
• The pretrained GraphGPT serves as a strong graph generator, demonstrated by its strong ability to perform both few-shot and conditional graph generation.
• Graph2Seq+GraphGPT enables effective graph mixup in the Euclidean space, overcoming previously known Non-Euclidean challenges.
• The edge-centric pretraining framework GraphsGPT demonstrates its efficacy in graph domain tasks, excelling in both representation and generation.

This is the official code implementation of ICML 2024 paper A Graph is Worth $K$ Words: Euclideanizing Graph using Pure Transformer.

The model checkpoints can be downloaded from 🤗 Transformers. We provide both the foundational pretrained models with different number of Graph Words $\mathcal{W}$ (GraphsGPT-nW), and the conditional version with one Graph Word (GraphsGPT-1W-C).

Model Name	Model Type	Model Checkpoint
GraphsGPT-1W	Foundation Model
GraphsGPT-2W	Foundation Model
GraphsGPT-4W	Foundation Model
GraphsGPT-8W	Foundation Model
GraphsGPT-1W-C	Finetuned Model

To get started with GraphsGPT, please run the following commands to install the environments.

git clone [email protected]:A4Bio/GraphsGPT.git
cd GraphsGPT
conda create --name graphsgpt python=3.12
conda activate graphsgpt
pip install -e .[dev]
pip install -r requirement.txt

We provide some Jupyter Notebooks in ./jupyter_notebooks, and their corresponding online Google Colaboratory Notebooks. You can run them for a quick start.

Example Name	Jupyter Notebook	Google Colaboratory
GraphsGPT Pipeline	example_pipeline.ipynb
Graph Clustering Analysis	clustering.ipynb
Graph Hybridization Analysis	hybridization.ipynb
Graph Interpolation Analysis	interpolation.ipynb

You should first download the configurations and data for finetuning, and put them in ./data_finetune. (We also include the finetuned checkpoints in the model_zoom.zip file for a quick test.)

To evaluate the representation performance of Graph2Seq Encoder, please run:

bash ./scripts/representation/finetune.sh

You can also toggle the --mixup_strategy for graph mixup using Graph2Seq.

For unconditional generation with GraphGPT Decoder, please refer to README-Generation-Uncond.md.

For conditional generation with GraphGPT-C Decoder, please refer to README-Generation-Cond.md.

To evaluate the few-shots generation performance of GraphGPT Decoder, please run:

bash ./scripts/generation/evaluation/moses.sh
bash ./scripts/generation/evaluation/zinc250k.sh

@article{gao2024graph,
  title={A Graph is Worth $K$ Words: Euclideanizing Graph using Pure Transformer},
  author={Gao, Zhangyang and Dong, Daize and Tan, Cheng and Xia, Jun and Hu, Bozhen and Li, Stan Z},
  journal={arXiv preprint arXiv:2402.02464},
  year={2024}
}

If you have any questions, please contact:

• Zhangyang Gao: [email protected]

• Daize Dong: [email protected]