CoAtGIN: Marrying Convolution and Attention for Graph-based Molecule Property Prediction

CoAtGIN is a lightweight graph neural network for molecular property prediction. It combines k-hop convolution for local neighborhoods with a virtual node and linear attention for global graph context, targeting a practical trade-off between accuracy and parameter count.

On PCQM4Mv2 (OGB-LSC), the default coat3211 configuration reaches 0.0901 validation MAE with about 6.4M parameters.

Publication

Xuan Zhang, Cheng Chen, Zhaoxu Meng, Zhenghe Yang, Haitao Jiang, and Xuefeng Cui. CoAtGIN: Marrying Convolution and Attention for Graph-based Molecule Property Prediction. In IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2022, pp. 374–379.

• IEEE Xplore: https://ieeexplore.ieee.org/document/9995324/
• DOI: https://doi.org/10.1109/BIBM55620.2022.9995324
• Preprint (bioRxiv): https://www.biorxiv.org/content/10.1101/2022.08.26.505499v1

@inproceedings{Zhang2022CoAtGIN,
  author    = {Zhang, Xuan and Chen, Cheng and Meng, Zhaoxu and Yang, Zhenghe
               and Jiang, Haitao and Cui, Xuefeng},
  title     = {{CoAtGIN}: Marrying Convolution and Attention for Graph-based
               Molecule Property Prediction},
  booktitle = {2022 IEEE International Conference on Bioinformatics and
               Biomedicine (BIBM)},
  year      = {2022},
  pages     = {374--379},
  doi       = {10.1109/BIBM55620.2022.9995324},
  url       = {https://ieeexplore.ieee.org/document/9995324/}
}

Method (short)

Level	Block	Role
Local	ConvMessage (VoVNet-style, iterated 1-hop)	Multi-hop neighborhood aggregation with degree scaling
Global	VirtMessage	Graph-level virtual node with recurrent residual state
Global	AttMessage (CosFormer-style linear attention)	Importance-weighted global aggregation, linear in nodes
Mix	GatedLinearBlock (GLU)	Per-layer residual mixer

Implementation lives under model/; the CoAtGIN-specific modules are in model/modify.py. Training entry point: model/train.py. See model/README.md for setup, training commands, and config naming (coat3211, etc.).

Leaderboard

Finalized submission CoAtGIN-tiny (test MAE 0.0908) on the OGB-LSC leaderboards (2022-09-20).

License

MIT — Copyright (c) 2022 Xuefeng Cui.