This model was initialized from
Bio_ClinicalBERT by adding three hidden layers after the BERT pooler layer. The model was trained on the Drug-Drug Interaction dataset extracted from
DrugBank database and
National Library of Medicine API.
It achieves the following results on the Test dataset:
Predict Drug Drug Interaction (DDI) from Chemical Structure of two drugs. The Model returns the probability of the two drugs having interaction with each other.
To construct the input, using "[SEP]" token to seperate between the two drugs, example of a properly constructed input is as following
1drug1 = "[Ca++].[O-]C([O-])=O" #Calcium Carbonate
2drug2 = "OC[C@H](O)[C@@H](O)[C@H](O)[C@H](O)CO" #Sorbitol
3correct_input = "[Ca++].[O-]C([O-])=O [SEP] OC[C@H](O)[C@@H](O)[C@H](O)[C@H](O)CO"
To avoid data leakage and able to predict DDI for new drugs, the drug1 or drug2 in the validation and the test set were not included in the training set. Their SMILES chemical structures were never exposed to the training process.
Using AWS EC2 g5.4xlarge instance. 24GB GPU.
