Scholarlink

De novo molecular design builds molecules from scratch, addressing the limitations of traditional virtual screening and exploring a broader chemical space for drugs. However, a major challenge in this field lies in the absence of standardized benchmarks for assessing the performance of molecular generation models. In this study, we propose MoGE, a comprehensive benchmarking framework that evaluates molecular generation models across five distinct drug generation scenarios, providing essential insights for the assessment of generative models in various drug design contexts. MoGE introduces 15 evaluation metrics, including three novel ones: Available Percentage, Percentage of Active Fragments, and Correlation between Protein Pocket and Molecular Volume (CPMV). Our study highlights the strengths and weaknesses of models across five different drug generation scenarios in various evaluation tasks. Additionally, we assess the impact of fine-tuning data volume on two transfer learning models, demonstrating that larger datasets enhance both the quality and applicability of the generated molecules. The research results provide important references for the selection and optimization of future models, promoting the standardization and application of the de novo molecular design field. To facilitate easier use of the benchmark and reduce the frequency of repeatedly running comparison methods, we have made the open-source Python code for MoGE available at the following GitHub repository: https://github.com/CSUBioGroup/MoGE .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MoGE: A Benchmark for Comprehensive Evaluation of Molecular Generation Models in De Novo Drug Design

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shiliang Zhang,
  • Huimin Zhu,
  • Renyi Zhou,
  • Min Li

摘要

De novo molecular design builds molecules from scratch, addressing the limitations of traditional virtual screening and exploring a broader chemical space for drugs. However, a major challenge in this field lies in the absence of standardized benchmarks for assessing the performance of molecular generation models. In this study, we propose MoGE, a comprehensive benchmarking framework that evaluates molecular generation models across five distinct drug generation scenarios, providing essential insights for the assessment of generative models in various drug design contexts. MoGE introduces 15 evaluation metrics, including three novel ones: Available Percentage, Percentage of Active Fragments, and Correlation between Protein Pocket and Molecular Volume (CPMV). Our study highlights the strengths and weaknesses of models across five different drug generation scenarios in various evaluation tasks. Additionally, we assess the impact of fine-tuning data volume on two transfer learning models, demonstrating that larger datasets enhance both the quality and applicability of the generated molecules. The research results provide important references for the selection and optimization of future models, promoting the standardization and application of the de novo molecular design field. To facilitate easier use of the benchmark and reduce the frequency of repeatedly running comparison methods, we have made the open-source Python code for MoGE available at the following GitHub repository: https://github.com/CSUBioGroup/MoGE .