yeonghyeon / pipgcn-tf2

Unofficial TensorFlow implementation of Protein Interface Prediction using Graph Convolutional Networks.

[TensorFlow] Protein Interface Prediction using Graph Convolutional Networks

Unofficial TensorFlow implementation of Protein Interface Prediction using Graph Convolutional Networks [1].

Usage

$ unzip dataset.zip
$ python run.py

Dataset

Number of samples

Set	Complex	Positive	Negative
Training	140	12866 (9.1 %)	128660 (90.9 %)
Validation	35	3138 (0.2 %)	31380 (99.8 %)
Test	55	4871 (0.1 %)	4953446 (99.9 %)

Sample view for Ligand and Receptor respectively.

Feature description

Graph
│
├── Vertex: 70 features
│    ├── Windowed Position Specific Scoring Matrix: 20
│    ├── Relative Accessible Surface Area: 1
│    ├── Residue Depth: 2
│    ├── Protrusion Index: 6
│    ├── Hydrophobicity: 1
│    └── Half Sphere Amino Acid Composition: 40
└── Edge: 2 features
     ├── Average Atomic Distance: 1
     └── CCαO Angle: 1

Three equations for each neural network in this repository. Each equation is provided as net00_pipgcn_navg.py, net01_pipgcn_neavg.py, and net02_pipgcn_odepn.py

Result

• Refer that the results shown below are temporary results. (AUROC: 0.81421)
• The model tuning process is not finished completely.

Loss graph.

Results of the test process. The first figure (upper side) shows the negative interaction score between ligand and receptor. The lower side figures show a histogram of negative interaction score of negative labeled and positive labeled sample.

Requirements

• TensorFlow 2.3.0
• White Box Layer 0.1.7

Contribution

• YeongHyeon Park : PIPGCN implementation (including neural network, data management)
• Myung Jin Kim and Yeong Beom Kim : review (double-check)

Reference

[1] Fout, Alex M. Protein interface prediction using graph convolutional networks. Diss. Colorado State University, 2017.
[2] Supplementary Data for NIPS Publication: Protein Interface Prediction using Graph Convolutional Networks. https://zenodo.org/record/1127774#.WkLewGGnGcY