O02_02

Enhancing Drug-Target Interaction Prediction using Large Language Models and Low-Rank Adaptation

Rintaro YASHIRO *, Nobuyuki YASUO, Masakazu SEKIJIMA

Tokyo Institute of technology
( * E-mail: yashiro.r.ab@m.titech.ac.jp)

Prediction of Drug-Target Interaction (DTI) is crucial in drug discovery applications such as drug repositioning, hit compound discovery, and side effect investigation. Experimental methods for DTI prediction are often costly and time-consuming, leading to the development of in silico prediction approaches. Databases like DrugBank and the Therapeutic Target Database (TTD) are commonly used for these in silico predictions. DrugBank, a comprehensive resource of drug information annotated from PubMed, is essential for creating training datasets when developing DTI prediction models.
However, the manual extraction of relevant DTI information from databases like PubMed, which sees millions of new articles added annually, is impractical. Therefore, automating the extraction of DTIs has become a critical need in the field.
In this study, we fine-tuned the Llama3-8B and Llama3-70B models, state-of-the-art large-scale language models known for their exceptional natural language processing capabilities, to develop a model that automatically extracts DTI information from article data. We also evaluated the impact of different prompts on the model's performance and its ability to generalize. The results demonstrate significant improvements in both extraction accuracy and generalization capability.