Supervised learning on molecules has incredible potential to be useful in chemistry, drug discovery, and materials science. Luckily, several promising and closely related neural network models invariant to molecular symmetries have already been described in the literature. These models learn a message passing algorithm and aggregation procedure to compute a function of their entire input graph. At this point, the next step is to find a particularly effective variant of this general approach and apply it to chemical prediction benchmarks until we either solve them or reach the limits of the approach. In this paper, we reformulate existing models into a single common framework we call Message Passing Neural Networks (MPNNs) and explore additional novel variations within this framework. Using MPNNs we demonstrate state of the art results on an important molecular property prediction benchmark; these results are strong enough that we believe future work should focus on datasets with larger molecules or more accurate ground truth labels.
核心贡献 · Key contributions
开发了一种 MPNN，在全部 13 个 QM9 目标上达到最先进水平，并在 11 个目标上达到化学精度。 Developed an MPNN achieving state-of-the-art on all 13 QM9 targets and chemical accuracy on 11.
引入多种 MPNN，仅使用分子拓扑就在 5 个目标上以化学精度预测 DFT。 Introduced multiple MPNNs predicting DFT to chemical accuracy on 5 targets using only molecular topology.
提出一种通用方法，在不增加计算或内存的情况下训练具有更大节点表示的 MPNN。 Proposed a general method to train MPNNs with larger node representations without increasing computation or memory.
将现有图神经网络模型统一为通用的消息传递神经网络（MPNN）框架。 Unified existing graph neural network models into a common Message Passing Neural Network (MPNN) framework.
通过主节点和 set2set 输出证明了长程相互作用对分子性质预测的重要性。 Demonstrated the importance of long-range interactions via master node and set2set output for molecular property prediction.
引入多塔架构，提高了可扩展性和泛化性能。 Introduced a multi-tower architecture that improves scalability and generalization performance.
局限 · Limitations
模型仅在 QM9 数据集中的小有机分子（最多 9 个重原子）上评估。 Models were evaluated only on small organic molecules (up to 9 heavy atoms) from QM9 dataset.
对更大分子或不同化学空间的泛化能力尚未验证。 Generalization to larger molecules or different chemical spaces remains unverified.
对空间信息的依赖可能限制其在缺乏 3D 构象时的适用性。 Spatial information reliance may limit applicability when 3D conformations are unavailable.
每个目标单独训练优于联合训练，限制了多任务效率。 Training one model per target outperforms joint training, limiting multi-task efficiency.
多塔变体无法与边缘网络消息函数有效结合。 The towers variant could not be effectively combined with the edge network message function.