摘要
我们介绍了一种直接在图上运行的卷积神经网络。这些网络允许端到端学习预测管道，其输入可以是任意大小和形状的图。我们提出的架构推广了基于圆形指纹的标准分子特征提取方法。实验结果表明，这些数据驱动的特征更具可解释性，并在多种任务中表现出更好的预测性能。
代码仓库
onakanob/Peptide_Graph_Autograd
GitHub 中提及
SystemicCypher/Neural-Molecule-Fingerprints
tf
GitHub 中提及
kimisyo/simple-GCN
pytorch
GitHub 中提及
nrel/m2p
GitHub 中提及
pgniewko/solubility
GitHub 中提及
HIPS/neural-fingerprint
官方
tf
GitHub 中提及
debbiemarkslab/neural-fingerprint-theano
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| drug-discovery-on-hiv-dataset | GraphConv | AUC: 0.822 |
| drug-discovery-on-muv | GraphConv | AUC: 0.836 |
| drug-discovery-on-pcba | GraphConv | AUC: 0.855 |
| drug-discovery-on-tox21 | GraphConv | AUC: 0.846 |
| drug-discovery-on-toxcast | GraphConv | AUC: 0.754 |
| graph-regression-on-lipophilicity | GC | RMSE: 0.655 |
| node-classification-on-citeseer-05 | GCN-FP | Accuracy: 43.9% |
| node-classification-on-citeseer-1 | GCN-FP | Accuracy: 54.3% |
| node-classification-on-citeseer-with-public | GCN-FP | Accuracy: 61.5% |
| node-classification-on-cora-05 | GCN-FP | Accuracy: 50.5% |
| node-classification-on-cora-1 | GCN-FP | Accuracy: 59.6% |
| node-classification-on-cora-3 | GCN-FP | Accuracy: 71.7% |
| node-classification-on-cora-with-public-split | GCN-FP | Accuracy: 74.6% |
| node-classification-on-pubmed-003 | GCN-FP | Accuracy: 56.2% |
| node-classification-on-pubmed-005 | GCN-FP | Accuracy: 63.2% |
| node-classification-on-pubmed-01 | GCN-FP | Accuracy: 70.3% |
| node-classification-on-pubmed-with-public | GCN-FP | Accuracy: 76.0% |