Using BERT to identify drug-target interactions from whole PubMed

• DOI: 10.1186/s12859-022-04768-x
• Publication Date: 2022-06-21T13:04:28Z
• Authors: Jehad Aldahdooh; Markus Vähä-Koskela; Jing Tang; Ziaurrehman Tanoli
• Keywords: Repurposing; Drug repositioning; DrugBank; Approved drug; chEMBL

Drug-target interactions (DTIs) are critical for drug repurposing and elucidation of mechanisms, manually curated by large databases, such as ChEMBL, BindingDB, DrugBank DrugTargetCommons. However, the number articles likely constitutes only a fraction all that contain experimentally determined DTIs. Finding extracting experimental information is challenging task, there pressing need systematic approaches to assist curation To this end, we applied Bidirectional Encoder Representations from Transformers (BERT) identify articles. Because DTI data intimately depends on type assays used generate it, also aimed incorporate functions predict assay format.Our novel method identified 0.6 million (along with protein information) which not previously included in public databases. Using 10-fold cross-validation, obtained ~ 99% accuracy identifying containing quantitative drug-target profiles. The F1 micro prediction format 88%, leaves room improvement future studies.The BERT model study robust proposed pipeline can be overlooked Overall, our provides significant advancement machine-assisted extraction curation. We expect it useful addition mechanism discovery repurposing.