PDI
A paper accepted by the 2020 BIBM conference, which is to achieve high precision in protein-drug interaction by using GAT and Transformer.
Source code and dataset for "Structure Enhanced Protein-Drug Interaction Prediction using Transformer and Graph Embedding"
Reqirements:
- Pytorch=1.2.0
- torchvision=0.4.0
- cudatoolkit=10.0 (View version:
cat /usr/local/cuda/version.txt) - Python3
- transformers=2.7.0 (How to install: huggingface/transformers)
Prepare Data:
Pre-training phase:
- Large-scale protein sequences
Location: ./data/total_modified_enzy_seqs-v2.csvFormat: 1.9G, 4168590 lines |
Accession_Code | Recommended Name | EC Number | Organism | Source | No of amino acids | Sequence |
|---|---|---|---|---|---|---|---|
| 1 | P15807 | NaN | 1 | Saccharomyces cerevisiae (strain ATCC 204508 / S288c) | Swiss-Prot | 274 | MVKSLQLAHQLKDKKILLIGGGEVGLTRLYKLI... |
| 2 | Q5SRE7 | NaN | 1 | Homo sapiens | Swiss-Prot | 291 | MACLSPSQLQKFQQDGFLVLEGFLSAEECVAMQQRI... |
We just need to use the last column (protein sequences), getting sequences by ./pre-training/get_sequences.py and saving in ./data/corpus-enz416w.bigFormat: 3.0G, 4168590 lines |
Protein sequences |
|---|---|
| M V K S L Q L A H Q L K D K K I L L I G G G E V G L T R L Y K L I P T G C K L T L V S... | |
| M A C L S P S Q L Q K F Q Q D G F L V L E G F L S A E E C V A M Q Q R I G E I V A E M D V... |
The data is divided into train-set and dev-set by ./pre-training/train_dev_split.py, saving in ./data/corpus-enz416w-train.tsv and ./data/corpus-enz416w-dev.tsv.
Next, we need to generate protein sequences's vocabulary by ./pre-training/generate_vocab.py, saving in ./pre-training/protein_vocab.txt
Fine-tuning phase:
- The data of protein-drug interaction (PDBBind2016)
Location:./data/pdbbind2016.pkl
Format: 4.2G, 13435 lines
| PDB-ID | Affinity-Value | seq | rdkit_smiles | set | contact_map | |
|---|---|---|---|---|---|---|
| 0 | 3zzf | 0.4 | NGFSATRSTV... | CC(=O)NC@@HC(=O)O | train | [array([[ True, True, True, ..., False, False, False], ..., [False, False, False, ..., True, True, True]])]] |
| 1 | 11gs | 5.82 | PYTVVY... | CC[C@H](C(=O)...,CC[C@@H]... | train | [array([[ True, True, True, ..., False, False, False], ..., [False, False, False, ..., True, True, True]])] |
- train_df_nums: 11877, valid_df_nums: 1000, core2016_df_nums: 290, casf2013_df_nums: 195, casf2013_df_nums: 73
Model architecture
Cite:
If you use the code, please cite this paper:
Hu, Fan, et al. "Structure Enhanced Protein-Drug Interaction Prediction using Transformer and Graph Embedding." 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2020.