abs

학술발표회초록보기

Inquiry on Abstract abstract@kcsnet.or.kr

Inquiry on Payment member@kcsnet.or.kr

현재 가능한 작업은 아래와 같습니다.
  • 03월 02일 17시 이후 : 초록수정 불가능, 일정확인 및 검색만 가능

129th General Meeting of Korean Chemical Society & Exposition Spatially Constrained 3D Scaffold-based Molecular Generative Modeling Toward a Controllable Drug Design

Submission Date :
2 / 28 / 2022 , 16 : 41 : 06
Abstract Number :
129022828628
Presenting Type:
Poster Presentation
Presenting Area :
Physical Chemistry
Authors :
Wonho Zhung, Woo youn Kim1,*
Department of Chemistry, Korea Advanced Institute of Science and Technology, Korea
1Department of Chemistry, KAIST, Korea
Assigned Code :
PHYS.P-221 Assigend Code Guideline
Presenting Time :
April 14 (THU) 11:00~13:00
De novo molecular design has been the subject of immense interest, as it has numerous applications. Remarkably, the scaffold-based generative modeling is highlighted as an effective way of generating desirable molecules where much of the chemical functionalities are preserved by its core structure. In drug discovery, one primary goal is to elaborate the scaffold to increase binding affinity and selectivity toward a binding pocket. Yet, encountering the surrounding 3D geometry of a given pocket remains a challenging problem in previous approaches. In this light, we propose a spatially-constrained 3D scaffold-based generative model. Specifically, we adopted a variational-autoencoder architecture, where the binding structure of a whole molecule is first embedded as a latent vector. The decoder then estimates the probability distributions of types and interatomic distances from the latent vector. Two different distance distributions --- the distance between ligand atoms and the distance between ligand and pocket atoms --- are separately modeled. During the generation, we used an auto-regressive process where the next type and coordinate are sampled from the learned distribution sequentially. As a result, our model can generate valid and diverse molecules from the given pocket and the scaffold. We demonstrated the selectivity of the generated molecules, where the molecule generated from the cognate receptor is barely bound to the non-cognate receptors. Finally, we leveraged the conditional generation scheme to control the properties of generated molecules, providing a possible framework for scaffold optimization.