Fine-Tuning Transformers Model for Predicting the Bioactivity of Alzheimer's Drug Candidates

The Process of drug discovery can take decades and cost billions of dollars until FDA approves it. Prediction of Molecular properties is one of the crucial tasks in drug discovery. Bioactivity values (e.g., pIC50 values) are molecular properties commonly used to evaluate and select potential drug candidates. To increase the efficiency in drug discovery, lots of research has been conducted by implementing machine learning to do drug candidate screening. In this study we conduct Large Language Model-based approach by fine-tuning of Transformers Model on Alzheimer's drug candidates to predict their bioactivity, where in earlier research implementation of Random Forest was the gold standard. We conduct fine-tuning on classification and regression tasks. For comparison, we compare the result with QSAR random forest regression and classification model. The experiment result shows that ChemBERTa as a language model has better performances for both regression and classification tasks to predict bioactivity compared to the QSAR Random Forest model. The classification task performed by ChemBERTa obtain auROC and auPRC scores of 0.77 and 0.81, respectively.