P01-14

Protein Tertiary Structure Prediction with Fine-tuned AlphaFold2 for Ligand Virtual Screening

Yuki YASUMITSU *, Masahito OHUE

School of Computing, Institute of Science Tokyo
( * E-mail: yasumitsu.y.aa@m.titech.ac.jp )

Virtual Screening (VS) is a method for selecting drug candidate compounds from a large number of compounds using a computer.
Structure-Based Virtual Screening (SBVS) is a method to perform VS based on the 3D structure of proteins.
Compared to ligand-based methods, SBVS does not use known experimental information on the target protein, and thus can discover highly novel drug candidate compounds.
In general, it is known that the drug-bound holo structure is more accurate than the drug-unbound apo structure for SBVS.
The 3D structure of a target protein is necessary for SBVS, but when the 3D structure is unknown, it is necessary to predict the 3D structure. Homology modeling has been used to predict the 3D structure using homologous proteins with known structures, but the use of predicted structures based on machine learning models is now being considered.
In a previous study applying AlphaFold2, a protein conformation prediction model, to SBVS, the screening performance of the predicted structures was found to be inferior to that of the holo structure and comparable to that of the apo structure.
In this study, we propose a method for fine tuned AlphaFold2 using a dataset of holo structures to build a model that predicts a structure suitable for SBVS.
We also attempted to improve screening performance by using the holo structure in the template structure used by AlphaFold2.
Screening performance was then evaluated by performing docking simulations on the DUD-E data set.