Table 2.
Table summarizing several aspects of the compared studies. Several publications investigated different variants of the proposed models. We present the performance of only the best models among them.
| Work | Datasets | Model | Tasks Performed | Performance |
|---|---|---|---|---|
| Elangovan et al. (2020) [39] | Processed version of the IntAct dataset with seven types of interactions | Ensemble of fine-tuned BioBERT models; no external knowledge used | Typed and untyped RE with relationship types such as phosphorylation, acetylation, etc. | Typed PPI: 0.540; untyped PPI: 0.717; metric: F1 score |
| Giles et al. (2020) [40] | Manually curated from the MEDLINE database | Fine-tuned BioBERT model; used STRING database knowledge during dataset curation | Classification problem with classes coincidental mention, positive, negative, incorrect entity recognition, and unclear | Curated data and BioBERT: 0.889; metric: F1 score |
| Su et al. (2020) [41] | Processed versions of the BioGRID, DrugBank, and IntAct datasets | Fine-tuned the BERT model integrated with LSTM and additive attention; no external knowledge used | Classification tasks on PPI (binary), DDI (multiclass), and ChemProt (multiclass) | PPI: 0.828; DDI: 0.807; ChemProt: 0.768; metric: F1 score |
| Su et al. (2021) [42] | Processed versions of the BioGRID, DrugBank, and IntAct datasets | Contrastive learning model; no external knowledge used in dataset curation or as a part of the model | Classification tasks on PPI (binary), DDI (multiclass), and ChemProt (multiclass) | PPI: 0.827; DDI: 0.829; ChemProt: 0.787; metric: F1 score |
| Wang et al. (2020) [43] | Processed versions of the BioCreative VI PPI dataset | A multitasking architecture based on BERT, BioBERT, BiLSTM, and text CNN; no external knowledge used | Document triage classification, NER (auxiliary tasks), and PPI RE (main task). | NER task: 0.936, PPI RE (exact match evaluation): 0.431; metric: F1 score |
| Zhou et al. (2019) [44] | Processed versions of the BioCreative VI PPI dataset | KAN; TransE used to integrate prior knowledge from the BioGRID and IntAct datasets on triplets to the model | PPI-RE classification task from BioCreative VI | PPI RE (exact match evaluation): 0.382 PPI RE: (HomoloGene evaluation): 0.404; metric: F1 score |