[paper]

This is the official code repository for the paper "Accurate Protein Function Prediction via Graph Attention Networks with Predicted Structure Information".

pytorch >=1.7.1
pytorch-geometric >= 1.7.0

To use GAT-GO with pre-processed data, we provided a data_loader which parses the pre-processed sequence features ussed in GAT-GO.
GAT-GO.py can be used to make prediction with pre-processed seuqnces. Examples can be found in this Notebook

python GAT-GO.py --ModelPath <PATH> --Device <CUDA device> --BatchSize <BatchSize> --SeqIDs <SeqIDs> --OutDir <OutDir>

**** Intput ****

--ModelPath : Path to trained model weights
--OutDir    : Output directory, where the result will be saved
--BatchSize : Batch size  
--Device    : CUDA device to be used for inferece

**** Output ****
GAT-GO_Results.pt will be saved at <OutDir> which is a serialized dictionary indexed by sequence identifiers provided in <SeqIDs>

To extract the GO-terms from the result, please see the Notebook example.

For each sequence in PDB/PDBmmseq dataset, a serialized dictionary stores the processed features used in GAT-GO. Details can be found below

1. data.seq: One-hot encoded primary sequence
2. data.pssm: Sequence profile constructed from MSA
3. data.x: Residue level sequence embedding generated from ESM-1b
4. data.edge_index: Contact map index
5. data.seq_embed: Sequence level embedding generated from ESMA-1b
6. data.label: GO term annotation
7. data.chain_id: Sequence identifier

Pre-processed data and pre-trained model can be downloaded at this link