•  
  •  
 

Turkish Journal of Electrical Engineering and Computer Sciences

Author ORCID Identifier

BELKIS BÜYÜKPATPAT: 0000-0001-5953-7580

BEGÜM MUTLU: 0000-0003-1960-2143

EBRU AKÇAPINAR SEZER: 0000-0002-9287-2679

Abstract

Drug-drug interaction (DDI) classification is a critical task for ensuring drug safety and supporting clinical decision-making, particularly in polypharmacy scenarios. Most existing DDI classification approaches were found to rely primarily on chemical and biological representations of drugs, whereas the potential of drug-related textual information remained comparatively underexplored. In this study, the effectiveness of using only textual features for DDI classification was investigated. To this end, drug textual features, including descriptions, indications, mechanisms, and pharmacodynamic information, were retrieved from DrugBank for two benchmark datasets, namely DDIMDL and MDF-SA-DDI. The textual features of each interacting drug pair were combined into a single sequence using special tokens, and the resulting sequence was used as input to transformer-based models, including BERT, BioBERT, UmlsBERT, and SAPBERT, enabling joint contextual modeling of information from both drugs. A systematic evaluation of general-purpose and domain-adapted transformer models for text-based DDI classification was conducted, with particular emphasis placed on UmlsBERT and SAPBERT, which incorporate biomedical knowledge from the Unified Medical Language System. Experimental results demonstrated that the domain-adapted SAPBERT model consistently achieved superior performance across both datasets, especially when description and mechanism features were leveraged, whereas the general-purpose BERT model yielded the lowest performance. On the DDIMDL dataset, SAPBERT was observed to outperform previously reported results based on alternative drug representations, while on the MDF-SA-DDI dataset, performance comparable to state-of-the-art methods was achieved, with an approximate 3% difference in F-score. Overall, the findings showed that competitive DDI classification performance could be achieved using textual features alone and that the use of domain-adapted language models positively contributed to capturing clinically relevant drug semantics.

DOI

10.55730/1300-0632.4288

Keywords

Drug-drug interaction classification, UmlsBERT, SAPBERT

First Page

1037

Last Page

1053

Publisher

The Scientific and Technological Research Council of Türkiye (TÜBİTAK)

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Share

COinS