ScholarWorks@성균관대학교

초록

Molecular property prediction plays a pivotal role in advancing our understanding of molecular representations, serving as a key driver for progress in drug discovery. Leveraging deep learning to gain comprehensive insights into molecular properties has become increasingly critical. Recent breakthroughs in molecular property prediction have been achieved through molecular pretraining models, which utilize large-scale databases of unlabeled molecules for pretraining, followed by fine-tuning for specific downstream tasks. These models enable a deeper understanding of molecular properties. In this study, we review recent advancements in molecular property prediction using molecular pretraining models. Our focus includes molecular descriptors, the impact of pretraining dataset size, molecular characterization model architectures, and the diversity of pretraining task types. Additionally, we compare the performance of existing methods and propose future directions to enhance the effectiveness of molecular pretraining models.

키워드