[论文解读] Hierarchical Inter-Message Passing for Learning on Molecular Graphs

Matthias Fey, Jan-Gin Yuen+1 more|arXiv (Cornell University)|Jun 22, 2020

Machine Learning in Materials Science参考文献 39被引用 34

一句话总结

简要：介绍一个分层GNN，联合处理分子图及其结点树表示，通过互信息传递提升表达能力和在分子性质预测任务上的性能。

ABSTRACT

We present a hierarchical neural message passing architecture for learning on molecular graphs. Our model takes in two complementary graph representations: the raw molecular graph representation and its associated junction tree, where nodes represent meaningful clusters in the original graph, e.g., rings or bridged compounds. We then proceed to learn a molecule's representation by passing messages inside each graph, and exchange messages between the two representations using a coarse-to-fine and fine-to-coarse information flow. Our method is able to overcome some of the restrictions known from classical GNNs, like detecting cycles, while still being very efficient to train. We validate its performance on the ZINC dataset and datasets stemming from the MoleculeNet benchmark collection.

研究动机与目标

• 激发并解决标准GNN在检测分子环及高阶结构方面的局限性。
• 提出一种双图架构，在分子图及其结点树表示上同时学习。
• 实现图间信息交换，以实现粗到细和细到粗的信息流。
• 在ZINC和MoleculeNet基准数据集上展示预测性能的提升，同时保持训练效率。

提出的方法

• 在分子图G及其结点树T上各自运行两个GNN，在每个图内进行消息传递。
• 实现互消息传递：X^(l) <- X^(l) + sigma(S Z^(l) W1^(l)) 将粗到细信息从T传递到G。
• 实现互消息传递：Z^(l) <- Z^(l) + sigma(S^T X^(l+1) W2^(l)) 将细到粗信息从G传递到T。
• 在结点树中用类别编码Z^(0)表示簇，覆盖单体、键、环、桥连化合物。
• 通过将图读出和树读出拼接来聚合两图的最终表示，形成分子嵌入。
• 基于成熟的GIN-E（图）和GIN（结点树）算子，使用Adam端到端训练。

实验结果

研究问题

• RQ1 分子图及其结点树表示之间的分层互消息传递是否能够提高表达能力和相对于标准GNN的预测性能？
• RQ2 粗到细/细到粗的信息交换是否有助于克服普通GNN固有的环检测限制？
• RQ3 在知名分子基准（ZINC、MoleculeNet、OGB）上，该架构在预测准确性与效率方面的影响如何？

主要发现

• On ZINC (10k) the proposed method achieves MAE 0.151 ± 0.006, outperforming all baselines including GIN-E (0.252 ± 0.014) and the full ZINC set (0.036 ± 0.002 vs 0.088 ± 0.002 for GIN-E).
• On ZINC Full, the method attains MAE 0.036 ± 0.002, significantly better than competitors.
• On a subset of MoleculeNet (HIV, MUV, Tox21), the method shows ROC-AUC improvements over NGF, RP-NGF, and GIN-E baselines (e.g., HIV: 84.81 ± 0.42 vs 83.83 ± 0.67 for GIN-E).
• On Open Graph Benchmark datasets ogbg-molhiv and ogbg-molpcba, the approach yields ROC-AUC 78.80 ± 0.82 and PRC-AUC 27.39 ± 0.17, outperforming GCN-E, GatedGCN-E, and GIN-E baselines.
• The architecture achieves strong performance with relatively shallow networks (2–3 layers) due to the enlarged receptive field from inter-graph communication.
• The method remains computationally efficient with marginal memory and runtime overhead compared to single-graph GNNs.