search

Bhattacharya-Lab / EquiPPIS

E(3) equivariant graph neural network for PPI site prediction
GNU General Public License v3.0
8 stars 2 forks source link

how to train the EquiPPIS from the scratch? #2

Closed onlyonewater closed 1 year ago

onlyonewater commented 1 year ago

could you provide the script to train the EquiPPIS from the scratch?

onlyonewater commented 1 year ago

and it seems that you set the threshold to 0.18 to determining whether two residues interact, why? can you give a reason?

Roche78 commented 1 year ago

Feature generation scripts are already provided. One can easily perform training from scratch by following the network architecture and hyperparameters described in the manuscript. For performance assessment, we used evaluation metrics and criteria consistent with prior work for a fair performance comparison

onlyonewater commented 1 year ago

ok, thanks, I will have a try. could you give me a paper link about how to choose the threshold? which paper do you refer to?

Roche78 commented 1 year ago

You may take a look at the GraphPPIS paper available at https://academic.oup.com/bioinformatics/article/38/1/125/6366544

onlyonewater commented 1 year ago

In GraphPPIS, they use an dynamic threshold which is maximizing the F1-score, but in your code, you use a certainty threshold (0.18), https://github.com/Bhattacharya-Lab/EquiPPIS/blob/ae1627767320f8d6e8dd3ffaeba0ee105bd8c4b5/EquiPPIS.py#L151 I think it is different between your implementation and GraphPPIS.

Roche78 commented 1 year ago

It is important to note that our prediction output is independent of any thresholds -- one can easily change the threshold if needed. For the EquiPPIS.py code provided in the github repo, we used the same threshold as that found in https://github.com/biomed-AI/GraphPPIS/blob/ba01c20f5c2c4816cc31b829f20ae9f0f7c7bc79/GraphPPIS_predict.py#L287.

onlyonewater commented 1 year ago

ok, got it, thanks!!!