Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MolGenSurvey: A Systematic Survey in Machine Learning Models for Molecule Design

Yuanqi Du, Tianfan Fu, Jimeng Sun, Shengchao Liu

TL;DR

<3-5 sentence high-level summary> MolGenSurvey addresses the problem of designing new molecules using machine learning by surveying two main ML paradigms: deep generative models (DGMs) and combinatorial optimization (COMs). It catalogs molecule representations (1D SMILES/SELFIES, 2D graphs, 3D geometry) and a spectrum of generative approaches, including autoregressive models, variational autoencoders, normalizing flows, GANs, diffusion models, and energy-based models, as well as combinatorial optimization methods; it discusses the optimization objective $p({\bm{x}})$ and the variational lower bound $L_{\text{VLB}}$ where relevant. It organizes molecule-design tasks into 1D/2D and 3D categories with generation and optimization settings, and outlines evaluation protocols and data resources, highlighting open challenges such as out-of-distribution generation, expensive oracles, and interpretability. The work provides a structured roadmap for translating ML advances into practical molecule design pipelines with real-world impact.

Abstract

Molecule design is a fundamental problem in molecular science and has critical applications in a variety of areas, such as drug discovery, material science, etc. However, due to the large searching space, it is impossible for human experts to enumerate and test all molecules in wet-lab experiments. Recently, with the rapid development of machine learning methods, especially generative methods, molecule design has achieved great progress by leveraging machine learning models to generate candidate molecules. In this paper, we systematically review the most relevant work in machine learning models for molecule design. We start with a brief review of the mainstream molecule featurization and representation methods (including 1D string, 2D graph, and 3D geometry) and general generative methods (deep generative and combinatorial optimization methods). Then we summarize all the existing molecule design problems into several venues according to the problem setup, including input, output types and goals. Finally, we conclude with the open challenges and point out future opportunities of machine learning models for molecule design in real-world applications.

MolGenSurvey: A Systematic Survey in Machine Learning Models for Molecule Design

TL;DR

<3-5 sentence high-level summary> MolGenSurvey addresses the problem of designing new molecules using machine learning by surveying two main ML paradigms: deep generative models (DGMs) and combinatorial optimization (COMs). It catalogs molecule representations (1D SMILES/SELFIES, 2D graphs, 3D geometry) and a spectrum of generative approaches, including autoregressive models, variational autoencoders, normalizing flows, GANs, diffusion models, and energy-based models, as well as combinatorial optimization methods; it discusses the optimization objective and the variational lower bound where relevant. It organizes molecule-design tasks into 1D/2D and 3D categories with generation and optimization settings, and outlines evaluation protocols and data resources, highlighting open challenges such as out-of-distribution generation, expensive oracles, and interpretability. The work provides a structured roadmap for translating ML advances into practical molecule design pipelines with real-world impact.

Abstract

Molecule design is a fundamental problem in molecular science and has critical applications in a variety of areas, such as drug discovery, material science, etc. However, due to the large searching space, it is impossible for human experts to enumerate and test all molecules in wet-lab experiments. Recently, with the rapid development of machine learning methods, especially generative methods, molecule design has achieved great progress by leveraging machine learning models to generate candidate molecules. In this paper, we systematically review the most relevant work in machine learning models for molecule design. We start with a brief review of the mainstream molecule featurization and representation methods (including 1D string, 2D graph, and 3D geometry) and general generative methods (deep generative and combinatorial optimization methods). Then we summarize all the existing molecule design problems into several venues according to the problem setup, including input, output types and goals. Finally, we conclude with the open challenges and point out future opportunities of machine learning models for molecule design in real-world applications.
Paper Structure (47 sections, 21 equations, 1 figure, 6 tables)

This paper contains 47 sections, 21 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Molecule Structure and Representation Learning
  3. 1D String Description
  4. 2D Molecular Graph
  5. 3D molecular Geometry
  6. Other Molecular Featurization Methods
  7. Summary
  8. Generative Methods
  9. Deep Generative Models (DGMs)
  10. Autoregressive Models (ARs)
  11. Variational Autoencoders (VAEs)
  12. Normalizing Flows (NFs)
  13. Generative Adversarial Networks (GANs)
  14. Diffusion models
  15. Energy-Based Models (EBMs)
  16. ...and 32 more sections

Figures (1)

  • Figure 1: Four data structures on an example molecule. The 1D SMILES string is Cc1ccccc1, and the 1D SELFIES string is [C][C][=C][C][=C][C][=C][Ring1][=Branch1]. \ref{['fig:molecule_feature_example_2D', 'fig:molecule_feature_example_3D']} are the 2D and 3D molecular Geometry respectively.