Table 1 Accuracy of BERT models on three independent datasets. DrugProt is the dataset containing 2788 positive articles based on DTIs (positive class) and 1215 from negative articles class. Medline is a completely negative dataset, and ChEMBL is a completely positive dataset containing DTIs
From: Using BERT to identify drug-target interactions from whole PubMed
Dataset | Articles | BERT | SciBERT | BioBERT | BioMed-RoBERTa | BlueBERT | Majority voting |
|---|---|---|---|---|---|---|---|
DrugProt | 4003 | 68 | 65.9 | 71.5 | 71.4 | 67.5 | 69.6 |
Medline | 55,056 | 99.7 | 98.6 | 75.2 | 99.9 | 100 | 100 |
ChEMBL | 876 | 89.6 | 93.2 | 91.2 | 83.4 | 88.7 | 90.3 |