基于图神经网络的单步逆合成集成研究

Graph Neural Network-Based Ensemble Methods for Single-Step Retrosynthesis

摘要

化学逆合成分析是从目标产物分子反向推导至可商业获取前体分子的过程，为了解决现有单步化学逆合成方法在泛化能力与预测精度方面的局限，本文提出了一种融合分子指纹的图神经网络集成预测模型，解决了单一模型扩展性和可解释性不高的问题。该方法通过构建多维分子表征并动态选择最优预测策略，以提升逆合成预测的准确性与稳定性。模型基于消息传递神经网络（Message Passing Neural Network，MPNN）、全局自注意力机制及 AttentiveFP 网络提取分子结构特征，并引入扩展连接指纹（Extended Connectivity Fingerprints，ECFP）实现局部结构信息与全局分子特征的联合建模，构建了 MPNN+FPS (Fingerprints) 与 AttentiveFP+FPS 两种融合架构，在 USPTO-50K 和天然产物数据集中进行了大量实验，本文方法在预测精度与泛化能力方面均表现出显著优势，为复杂分子与天然产物的逆合成分析提供了一种新的技术路径。

Abstract

Retrosynthetic analysis is the process of deriving commercially available precursor molecules from target product molecules. To address the limitations of existing single-step chemical retrosynthetic methods in terms of generalization ability and prediction accuracy, this paper proposes an ensemble prediction model integrating molecular fingerprints and graph neural networks, which addresses the poor scalability and interpretability of individual models. This method improves the accuracy and stability of retrosynthetic prediction by constructing multidimensional molecular representations and dynamically selecting the optimal prediction strategy. The model is based on message-passing neural networks (MPNNs), global self-attention mechanisms, and AttentiveFP networks to extract molecular structural features, and introduces extended connectivity fingerprints (ECFP) to achieve joint modeling of local structural information and global molecular features, constructing two fusion architectures: MPNN+FPS (Fingerprints) and Attentive FP+FPS. Extensive experiments on the USPTO-50K and natural-product datasets validate the effectiveness of the method. The proposed ensemble retrosynthetic prediction method demonstrates significant advantages in both prediction accuracy and generalization ability, providing a new technical approach for retrosynthetic analysis of complex molecules and natural products.

关键词

逆合成反应预测 / 图神经网络 / 分子指纹 / 产物分类 / 特征融合 / 动态策略选择

Key words

retrosynthetic reaction prediction / graph neural networks / molecular fingerprint / product classification / feature fusion / dynamic strategy selection

参考文献

[1] COLEY C W, ROGERS L, GREEN W H, et al. Computer-assisted retrosynthesis based on molecular similarity[J].ACS Central Science, 2017, 3(12):1237-1245.

[2] SEGLER M H S, WALLER M P. Neural symbolic machine learning for retrosynthesis and reaction prediction[J].Chemistry:A European Journal, 2017, 23(25):5966-5971.

[3] CHEN B, LI C, DAI H, et al. Retro*: Learning retrosynthetic planning with neural guided A* search[C]//Proceedings of the 37th International Conference on Machine Learning. Vienna, Austria: PMLR,2020:1608-1616.

[4] DAI H, LI C, COLEY C, et al. Retrosynthesis prediction with conditional graph logic network[J].Advances in Neural Information Processing Systems, 2019, 32：8870-8880.

[5] CHEN S, JUNG Y. Deep retrosynthetic reaction prediction using local reactivity and global attention[J].JACS Au, 2021, 1(10):1612-1620.

[6] ISHIDA S, TERAYAMA K, KOJIMA R, et al. AI-driven synthetic route design incorporated with retrosynthesis knowledge[J].Journal of Chemical Information and Modeling, 2022, 62(6):1357-1367.

[7] XIE S, YAN R, GUO J, et al. Retrosynthesis prediction with local template retrieval[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Washington, DC, USA: AAAI Press, 2023:5330-5338.

[8] LILLICRAP T P, SANTORO A, MARRIS L, et al. Back-propagation and the brain[J].Nature Reviews Neuroscience, 2020, 21(6):335-346.

[9] CAMACHO M T, PAEGELOW M, MAS J F, et al. Geomatic Approaches for Modeling Land Change Scenarios[M]. Cham:Springer, 2018:1-8.

[10] SCARSELLI F, GORI M, TSOI A C, et al. The graph neural network model[J].IEEE Transactions on Neural Networks, 2008, 20(1):61-80.

[11] KIPF T N, WELLING M. Semi-superivised classification with graph convolutional networks [EB/OL]. (2016-09-09). https://arXiv.org/abs/1609.02907.

[12] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017. Long Beach, CA, USA: Curran Associates Inc, 2017:5998-6008.

[13] WEININGER D. SMILES: A chemical language and information system. Introduction to methodology and encoding rules[J].Journal of Chemical Information and Computer Sciences, 1988, 28(1):31-36.

[14] LIU B, RAMSUNDAR B, KAWTHEKAR P, et al. Retrosynthetic reaction prediction using neural sequence-to-sequence models[J].ACS Central Science, 2017, 3(10):1103-1113.

[15] KARPOV P, GODIN G, TETKO I V. A transformer model for retrosynthesis[C]//International Conference on Artificial Neural Networks. Cham: Springer International Publishing, 2019:817-830.

[16] SCHWALLER P, LAINO T, GAUDIN T, et al. Molecular transformer: A model for uncertainty calibrated chemical reaction prediction[J].ACS Central Science, 2019, 5(9):1572-1583.

[17] ZHENG S, RAO J, ZHANG Z, et al. Predicting retrosynthetic reactions using self-corrected transformer neural networks[J].Journal of Chemical Information and Modeling, 2019, 60(1):47-55.

[18] TETKO I V, KARPOV P, VAN DEURSEN R, et al. State-of-the-art augmented NLP transformer models for direct and single-step retrosynthesis[J].Nature CommuniCations, 2020, 11(1):5575.

[19] WAN Y, HSIEH C Y, LIAO B, et al. Retroformer: Pushing the limits of end-to-end retrosynthesis transformer[C]//Proceedings of the 39th International Conference on Machine Learning. Baltimore, USA: PMLR, 2022:22475-22490.

[20] ZHONG Z, SONG J, FENG Z, et al. Root-aligned SMILES: A tight representation for chemical reaction prediction[J].Chemical Science, 2022, 13(31):9023-9034.

[21] FANG L, LI J, ZHAO M, et al. Single-step retrosynthesis prediction by leveraging commonly preserved substructures[J].Nature Communications, 2023, 14(1):2446.

[22] ZHANG X, MO Y, WANG W, et al. Retrosynthesis prediction enhanced by in-silico reaction data augmentation[EB/OL]. (2024-01-31). https://arxiv.org/abs/2402.00086.

[23] LEVIN I, LIU M, VOIGT C A, et al. Merging enzymatic and synthetic chemistry with computational synthesis planning[J].Nature Communications, 2022, 13(1):7747.

[24] KIM T, LEE S, KWAK Y, et al. READRetro: Natural product biosynthesis predicting with retrieval-augmented dual-view retrosynthesis[J].New Phytologist, 2024, 243(6):2512-2527.

[25] LIN M H, TU Z, COLEY C W. Improving the performance of models for one-step retrosynthesis through re-ranking[J].Journal of Cheminformatics, 2022, 14(1):15.

[26] LI J, FANG L, LOU J G. RetroRanker: leveraging reaction changes to improve retrosynthesis prediction through re-ranking[J].Journal of Cheminformatics, 2023, 15(1):58.

[27] MAZIARZ K, LIU G, MISZTELA H, et al. Chemist-aligned retrosynthesis by ensembling diverse inductive bias models[EB/OL]. (2024-12-09). https://arvix.org/abs/2412.05269.

[28] MARCO A C, MYERS A, GRAHAM S J H, et al. The USPTO patent assignment dataset: Descriptions and analysis[R]. Alexandria, VA, USA: U.S. Patent and Trademark Office,2015:1-53.

[29] GOODMAN J. Computer software review: Reaxys[J].Journal of Chemical Information and Modeling, 2009, 49(12):2897-2898.

[30] ROGERS D, HAHN M. Extended-connectivity fingerprints[J].Journal of Chemical Information and Modeling, 2010, 50(5):742-754.