高等学校化学学报

高等学校化学学报

• 研究论文 • 上一篇    下一篇

基于自由能微扰与机器学习协同优化的黑色素瘤抗原肽设计研究

王泰中1,2，钟青庐2，张珊2，吴毅剑2，李金辉2，庄齐2，吴松1,2,3,4，付蕾2,4
  

  1. 1.安徽医科大学第二临床医学院

    2.深圳大学附属华南医院医学人工智能研究所

    3.深圳大学附属华南医院泌尿外科 4.深圳大学电子与信息工程学院

  • 收稿日期:2026-04-16 修回日期:2026-07-02 出版日期:2026-07-06 发布日期:2026-07-06
  • 通讯作者: 付蕾 E-mail:leifu@szu.edu.cn
  • 基金资助:
     深圳市医学研究专项资金（批准号：E250200620，E250200622，E250200623）和重庆市自然科学基金创新发展联合基金（批准号：CSTB2025NSCQ-LZX0079）资助

Free Energy Perturbation-Machine Learning Collaborative Optimization for Melanoma Antigenic Peptide Design

WANG Taizhong1,2，ZHONG Qinglu2，ZHANG Shan2，WU Yijian2，LI Jinhui2，ZHUANG Qi2，WU Song1,2,3,4*，FU Lei2,4*   

  1. 1.The Second Clinical College of Anhui Medical University

    2.Institute of Medical Artificial Intelligence, South China Hospital, Medical School, Shenzhen University

    3.Department of Urology, South China Hospital, Shenzhen University 4.College of Electronics and Information Engineering, Shenzhen University

  • Received:2026-04-16 Revised:2026-07-02 Online:2026-07-06 Published:2026-07-06
  • Supported by:
    Supported by the Shenzhen Medical Research Fund, China(Nos.E250200620, E250200622, E250200623) and the Chongqing Natural Science Foundation Innovation and Development Joint Fund, China(No.CSTB2025NSCQ-LZX0079)

PDF (PC)

补充材料

1

摘要： 针对黑色素瘤抗原肽与主要组织相容性复合体结合亲和力不足，疫苗疗效受限的问题，须建立基于物理与人工智能交互的协同优化框架，以高效筛选高亲和力候选肽段. 以PRAME190-198肽段及HLA-A*11:01复合物为研究对象，利用自由能微扰（FEP）精确量化抗原的氨基酸突变对结合自由能的影响，并驱动FEPaML模型进行多轮优化学习；应用该抗原优化框架对海量抗原突变组合进行高通量预测与排序，并运用分子动力学模拟的自由能微扰计算对机器学习模型预测的抗原进行验证. 自由能微扰的精确物理数据校正后的机器学习模型FEPaML可以高效、准确地进行抗原的优化设计，并能够捕捉多位点突变的非线性组合效应；筛选出的最佳三突变体I7F-E8S-K9R的相对结合自由能低至-24.43 kJ·mol-1，其结合亲和力较野生型显著提升. 该方法实现了抗原肽亲和力的高效、精准优化，能够显著降低筛选成本与周期，为黑色素瘤个体化疫苗设计提供理论依据与技术支持.

关键词: 分子动力学, 自由能微扰, 机器学习, 抗原优化

Abstract: To address the limited efficacy of peptide vaccines caused by insufficient binding affinity between melanoma antigens and the major histocompatibility complex (MHC) proteins, this study proposes a free energy perturbation (FEP)-AI synergistic design framework for efficiently screening high-affinity candidate peptides. We focused on the PRAME190-198 peptide-HLA-A*11:01 complex, and employed FEP to accurately quantify the impact of antigenic amino acid mutations on binding free energy. The resulting precise physical data were used to calibrate the FEPaML model through multiple rounds of iterative optimization. The optimized framework was then applied to perform high-throughput prediction and ranking of large-scale combinatorial mutations, followed by validation using the FEP calculation. The FEPaML model, calibrated with high-precision physical data from FEP, enabled efficient and accurate antigen optimization design and successfully captured the synergistic effects of multi-site mutations. The identified optimal triple mutant, I7F-E8S-K9R, achieved a relative binding free energy of -24.43 kJ·mol-1, showing significantly improved binding affinity compared to the wild type. This framework enables efficient and precise optimization of antigen peptide affinity, substantially reducing screening cost and duration while providing robust theoretical and technical support for personalized melanoma vaccine design.

Key words: Molecular dynamics, Free energy perturbation, Machine learning, Antigen optimization

中图分类号: 

TrendMD: