文章信息/Info

Title:

Application of Generative Deep Learning in Object-Oriented Molecular Design

文章編號:

1674-3962(2025)05-0424-12

作者:

王紀峰; 汪瑩

復旦大學高分子科學系 聚合物分子工程國家重點實驗室，上海 200438

Author(s):

WANG Jifeng;  WANG Ying

State Key Laboratory of Molecular Engineering of Polymers, Department of Macromolecular Science, Fudan University, Shanghai 200438, China

關鍵詞:

分子生成; 生成式深度學習; 生成對抗網絡; 變分自動編碼器; 去噪擴散概率模型; 模型性能評估框架; 分子表示

Keywords:

molecule generation;  generative deep learning;  generative adversarial network (GAN);  variational autoencoder (VAE);  denoising diffusion probabilistic model (DDPM);  model performance evaluation framework;  molecular representation

分類號:

TQ317;TP18

DOI:

10.7502/j.issn.1674-3962.202411011

文獻標志碼:

A

摘要:

分子設計作為化學與材料科學中的一項核心任務，面臨著在龐大的化學空間中高效篩選并開發具備特定功能的分子的問題，傳統方法在效率和探索性方面存在明顯局限。近年來，生成式深度學習的興起為分子設計提供了自動化與智能化的新路徑。綜述了生成式深度學習在分子設計中的應用進展，首先對不同分子表示方法（如SMILES、分子圖和三維結構表示）進行比較，分析了各自的優缺點。隨后，綜合評估了3種主流生成式模型：生成對抗網絡（GAN）、變分自動編碼器（VAE）和去噪擴散概率模型（DDPM），并探討了生成式模型在目標導向分子設計中的應用，重點分析不同模型在分子生成質量與性質優化方面的差異。最后，基于現有技術的研究進展，提出了未來生成式模型在分子設計領域的研究方向。

Abstract:

Designing molecules with specific functions within an immense chemical space is a fundamental challenge in chemistry and materials science, as traditional methods often lack efficiency and exploratory capacity. The advent of generative deep learning has introduced automated and intelligent approaches that promise to transform molecular design. In this review, we summarizes the advancements in applying generative deep learning to molecular design. We first compare molecular representation methods, including SMILES notation, molecular graphs, and three-dimensional structural representations, highlight their respective advantages and limitations. We then critically evaluate three leading generative models: generative adversarial network (GAN), variational autoencoder (VAE), and denoising diffusion probabilistic model (DDPM), and discuss applications of generative models in object-oriented molecular design, with a particular focus on the differences among various models in terms of molecular generation quality and property optimization. Finally, we propose future research directions for leveraging generative models in molecular design, aiming to inspire further advancements in this rapidly evolving field.