FM-PSEP

Our hierarchical graph convolutional neural network model FM-PSEP is an efficient general approach for multirelational link prediction in any hierarchical heterogeneous network.

Here, we are particularly concerned about the safety of polypharmacy, which is the concurrent use of multiple medications by a patient. Given a pair of drug, the model will predict how many polypharmacy side effects the drug pair will have, and what are the possibilities.

Code implementations base on the PyTorch-Geometric package, make sure it has been installed before running the code.

Data Drive

We construct a hierarchical heterogeneous graph of two node types: proteins and drugs. We think of the whole graph as three sub-graph.

• p-net: protein-protein association network
• d-net: drug-drug association network with multiple edge types (each edge type refers to a polypharmacy side effect)
• pd-net: drug-protein association network whose edges have the direction from protein node to drug node

Performance Comparision

We will firstly evaluate three auto-encoder approach (Hamilton et al., 2017), and compare them with our approach:

• GAE (Kipf and Welling, 2016): [paper]
• AROPE (Zhang et al., 2018): [paper], [code]
• LINE (Tang et al., 2015): [paper]

We firstly consider using DistMult (Yang et al., 2015[paper]) as decoder to predict the polypharmacy side effect of drug pairs.

Then, we will compare our model with the state-of-the-art polypharmacy side effect prediction models:

• DECAGON (Zitnik, 2018): [paper], [code]
• DeepDDI (Ryu et al., 2018): [paper], [code]
• mvGAE (Ma et al., 2018): [paper], [code]
• MR-GNN (Xu et al., 2019): [paper], [code]

Result by now

All the models are trained on GPU. torch-memlab helps to profile and inspect memory usage

Pre-train ppi embedding

2 layers GCN:

• 80 epochs(output):
• around 92% auprc.
• less than 2GB GPU memory cost.
• Time cost: 2 mins.

Interaction prediction (use drug ids as feature, dd-net):

21st July - RGCN + DistMult:

• parameter dim: 16, 16, 16, 16
• around 64% auprc (Limited to the number of parameters, ).
• around 10GB GPU memory cost.
• Time cost: 10 mins.
• to improve: change batch => mini-batch, add more parameters, biased sampling.

5th Aug - RGCN + NN-1:

• output file: [here]
• parameter dim: {16, 16, 8, 8}, {16, 963}
• 100 epoch
• auprc:0.9317, auroc:0.9451, ap@50:0.9317, time/epoch:60s
• maximum GPU memory cost: 7.5G

8th Aug - RGCN + DistMult:

• output file: [here]
• parameter dim: {16, 64, 32, 16}
• 100 epoch
• auprc:0.9479, auroc:0.9586, ap@50:0.9482, time/epoch:109s
• maximum GPU memory cost: 10.94G

- Evaluation (the top20 highest and lowest AUPRC)

8th Aug - RGCN + NN-1:

• output file: [here]
• parameter dim: {16, 64, 32, 16}, {16, 963}
• 100 epoch
• auprc:0.9437, auroc:0.9542, ap@50:0.9438, time/epoch:s
• maximum GPU memory cost: 11.G

10th Aug - FM-PSEP

• output file: [here]
• parameter dim: {16-8-8}, {4}, {16-16-8-8}
• 100 epoch
• auprc:0.9460, auroc:0.9566, ap@50:0.9463, time/epoch:68s
• maximum GPU memory cost: 10.7G

10th Aug - RGCN + DistMult

• output file: [here]
• parameter dim: {16-16-8-8}
• 100 epoch
• auprc:0.9462, auroc:0.9568, ap@50:0.9465, time/epoch:62s
• maximum GPU memory cost: 7.6G

Interaction prediction (use protein ids as feature):

7th Aug - HGCN + NN-1:

• output file: [here]
• parameter dim: {32, 16, 8}, {16, 963}
• 100 epoch
• auprc:0.7461, auroc:0.7432, ap@50:0.7331, time/epoch:28s
• maximum GPU memory cost: 7.94G