基于变分编码器的纳米纤维素分子结构预测模型框架研究

doi:10.11949/0438-1157.20240334

摘要/Abstract

摘要：

纳米纤维素因其多样化的原料、制备方法以及改性方法而展现出丰富的分子结构及性能。但正因其结构多样性，在传统方法下研发周期长，研发成本高，若能从微观尺度设计分子结构则有助于大幅缩短该周期，而目前，现有的分子结构预测模型多适用于无机材料，对纳米纤维素的适应性有限。基于变分编码器搭建了纳米纤维素分子结构预测模型，针对纳米纤维素结构特点，设计了4条独有的结构生成约束。模型的结构生成准确率达到约63.0%。模型在识别部分结构方面表现优异，对主体结构识别率达到87.0%，能有效解耦纳米纤维素主体结构与改性基团结构，并在一定程度上证明了提出的模型框架对纳米纤维素及衍生材料的结构预测具有可行性，有助于相关材料的研发与制备。

关键词: 深度学习, 纳米纤维素, 结构预测, 神经网络, 模型设计

Abstract:

Nanocellulose exhibits rich molecular structures and properties due to its diverse raw materials, preparation and modification methods. However, due to its structural diversity, the research and development cycle under traditional methods is long and cost is high. If the molecular structure can be designed from the micro scale, it will help to significantly shorten the cycle. At present, the existing molecular structure prediction models are mostly suitable for inorganic materials and have limited adaptability to nanocellulose. Based on the structural characteristics of nanocellulose, four unique structure generation constraints were designed. The results show that the structure generation accuracy of the nanocellulose molecular structure prediction model built based on the variational encoder reaches approximately 63.0%. The model performs well in identifying partial structures, with a recognition rate of 87.0% for the main structure. It can effectively decouple the main structure of nanocellulose and the modified group structure, and proves to a certain extent that the model framework proposed in this study can effectively decouple the nanocellulose main structure and modified group structure. The structure prediction of cellulose and its derivative materials is feasible and helps to assist the development and preparation of related materials.

Key words: deep learning, nanocellulose, structure prediction, neural network, model design

中图分类号:

TS 71⁺2

赵武灵, 满奕. 基于变分编码器的纳米纤维素分子结构预测模型框架研究[J]. 化工学报, 2024, 75(9): 3221-3230.

Wuling ZHAO, Yi MAN. Research on framework of nanocellulose molecular structure prediction model based on variational encoder[J]. CIESC Journal, 2024, 75(9): 3221-3230.

图/表 10

参考文献 33

1	李艳丽, 邢惠萍, 李玉虎. CNC、CNF及BC对纸张加固效果的比较研究[J]. 中国造纸学报, 2021, 36(3): 81-86.
	Li Y L, Xing H P, Li Y H. A comparative study on paper strengthening performance of CNC, CNF, and BC[J]. Transactions of China Pulp and Paper, 2021, 36(3): 81-86.
2	Wei L Q, Agarwal U P, Hirth K C, et al. Chemical modification of nanocellulose with canola oil fatty acid methyl ester[J]. Carbohydrate Polymers, 2017, 169: 108-116.
3	Xie H X, Du H S, Yang X H, et al. Recent strategies in preparation of cellulose nanocrystals and cellulose nanofibrils derived from raw cellulose materials[J]. International Journal of Polymer Science, 2018, 2018: 1-25.
4	Jiménez A, Bidare P, Hassanin H, et al. Powder-based laser hybrid additive manufacturing of metals: a review[J]. The International Journal of Advanced Manufacturing Technology, 2021, 114(1): 63-96.
5	Oganov A R, Pickard C J, Zhu Q, et al. Structure prediction drives materials discovery[J]. Nature Reviews Materials, 2019, 4: 331-348.
6	Marakana P G, Dey A, Saini B. Isolation of nanocellulose from lignocellulosic biomass: synthesis, characterization, modification, and potential applications[J]. Journal of Environmental Chemical Engineering, 2021, 9(6): 106606.
7	Kumar R, Rai B, Gahlyan S, et al. A comprehensive review on production, surface modification and characterization of nanocellulose derived from biomass and its commercial applications[J]. Express Polymer Letters, 2021, 15(2): 104-120.
8	Chen Z K, Bononi F C, Sievers C A, et al. UV-visible absorption spectra of solvated molecules by quantum chemical machine learning[J]. Journal of Chemical Theory and Computation, 2022, 18(8): 4891-4902.
9	Schleder G R, Padilha A C M, Acosta C M, et al. From DFT to machine learning: recent approaches to materials science—a review[J]. Journal of Physics: Materials, 2019, 2(3): 032001.
10	Ryan K, Lengyel J, Shatruk M. Crystal structure prediction via deep learning[J]. Journal of the American Chemical Society, 2018, 140(32): 10158-10168.
11	Chen J, Chaudhari N S. Bidirectional segmented-memory recurrent neural network for protein secondary structure prediction[J]. Soft Computing, 2006, 10(4): 315-324.
12	Kadurin A, Nikolenko S, Khrabrov K, et al. druGAN: an advanced generative adversarial autoencoder model for de novo generation of new molecules with desired molecular properties in silico[J]. Molecular Pharmaceutics, 2017, 14(9): 3098-3104.
13	Kim H, Ko S, Kim B J, et al. Predicting chemical structure using reinforcement learning with a stack-augmented conditional variational autoencoder[J]. Journal of Cheminformatics, 2022, 14(1): 83.
14	Thakur V, Guleria A, Kumar S, et al. Recent advances in nanocellulose processing, functionalization and applications: a review[J]. Materials Advances, 2021, 2(6): 1872-1895.
15	Tajbakhsh N, Jeyaseelan L, Li Q, et al. Embracing imperfect datasets: a review of deep learning solutions for medical image segmentation[J]. Medical Image Analysis, 2020, 63: 101693.
16	徐飞翔, 蒋丽群, 郑安庆, 等. 碳基固体酸催化纤维素热解制备左旋葡聚糖和左旋葡萄糖酮[J]. 化工学报, 2022, 73(3): 1166-1172.
	Xu F X, Jiang L Q, Zheng A Q, et al. Carbon-based solid acid catalyzed the pyrolysis of cellulose to produce levoglucosan and levoglucosenone[J]. CIESC Journal, 2022, 73(3): 1166-1172.
17	Bradley A P. The use of the area under the ROC curve in the evaluation of machine learning algorithms[J]. Pattern Recognition, 1997, 30(7): 1145-1159.
18	付俊俊, 田彦, 陶劲松. 纳米微晶纤维素的表面基团及其改性[J]. 中国造纸, 2018, 37(1): 50-59.
	Fu J J, Tian Y, Tao J S. The surface groups and chemical modification of nanocrystalline cellulose[J]. China Pulp & Paper, 2018, 37(1): 50-59.
19	姚一军, 王鸿儒. 纤维素化学改性的研究进展[J]. 材料导报, 2018, 32(19): 3478-3488.
	Yao Y J, Wang H R. An overview on chemical modification of cellulose[J]. Materials Review, 2018, 32(19): 3478-3488.
20	陈子健, 唐艳军, 朱鹏, 等. 羧甲基纤维素的制备及其应用进展[J]. 中国造纸学报, 2022, 37(3): 144-154.
	Chen Z J, Tang Y J, Zhu P, et al. Progress in preparation and applications of carboxymethyl cellulose[J]. Transactions of China Pulp and Paper, 2022, 37(3): 144-154.
21	刘雄利, 王安, 王春平, 等. 纤维素纳米纤丝的制备和改性研究进展[J]. 中国造纸, 2020, 39(4): 74-83.
	Liu X L, Wang A, Wang C P, et al. Research progress in preparation and modification of cellulose nanofibril[J]. China Pulp & Paper, 2020, 39(4): 74-83.
22	Yang X P, Biswas S K, Han J Q, et al. Surface and interface engineering for nanocellulosic advanced materials[J]. Advanced Materials, 2021, 33(28): 2002264.
23	Lian P, Yan R H, Wu Z G, et al. Thermal performance of novel form-stable disodium hydrogen phosphate dodecahydrate-based composite phase change materials for building thermal energy storage[J]. Advanced Composites and Hybrid Materials, 2023, 6(2): 74.
24	Cai Y, Cui J, Chen M, et al. Multifunctional enhancement for highly stable and efficient perovskite solar cells[J]. Advanced Functional Materials, 2021, 31(7): 2005776.
25	Gražulis S, Chateigner D, Downs R T, et al. Crystallography open database—an open-access collection of crystal structures[J]. Journal of Applied Crystallography, 2009, 42(4): 726-729.
26	Hellenbrandt M. The inorganic crystal structure database (ICSD)—present and future[J]. Crystallography Reviews, 2004, 10(1): 17-22.
27	Rosen A S, Iyer S M, Ray D, et al. Machine learning the quantum-chemical properties of metal-organic frameworks for accelerated materials discovery[J]. Matter, 2021, 4(5): 1578-1597.
28	Pareek J, Jacob J. Data Compression and Visualization Using PCA and T-SNE[M]. Singapore: Springer Singapore, 2021: 327-337.
29	Jin W G, Barzilay R, Jaakkola T. Hierarchical generation of molecular graphs using structural motifs[C]//Proceedings of the 37th International Conference on Machine Learning. ACM, 2020: 4839-4848.
30	Jin W G, Barzilay R, Jaakkola T. Junction tree variational autoencoder for molecular graph generation[C]// Proceedings of the International Conference on Machine Learning. 2018.
31	Liu Q, Allamanis M, Brockschmidt M, et al. Constrained graph variational autoencoders for molecule design[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems. Montréal, Canada: ACM, 2018: 7806-7815.
32	Gómez-Bombarelli R, Wei J N, Duvenaud D, et al. Automatic chemical design using a data-driven continuous representation of molecules[J]. ACS Central Science, 2018, 4(2): 268-276.
33	Ochiai T, Inukai T, Akiyama M, et al. Variational autoencoder-based chemical latent space for large molecular structures with 3D complexity[J]. Communications Chemistry, 2023, 6(1): 249.

纳米纤维素主要分词序列	其余主要化学分词序列
‘C1C(C(C(C(O1)O)O)O)O’，	‘C(C(=O)O)’, ‘C(=O)C’, ‘C’, ‘[C@@]’
‘C1CCOC1’, ‘O1CCCC1’, ‘[O]1CCCC1’,	‘S(=O)(=O)O’, ‘[N@@+]’, ‘[NH2+]’
‘OC[C@H](O)[C@@H](O)[C@H](O)CO’	‘[C@@H]’, ‘[OH+]’, ‘[CH2-]’
‘C(C1[C@H](C(C(C(O1)O)O)O)O [C@H]2C(C(C(C(O2)CO)O)O)O)O’	‘[P@+]’, ‘[Cl+2]’, ‘[S@@]’, ‘[Si@]’, ‘[BH3-]’, ‘CC(O)C’, ‘#’, ‘[O-]’

纳米纤维素主要分词序列	其余主要化学分词序列
‘C1C(C(C(C(O1)O)O)O)O’，	‘C(C(=O)O)’, ‘C(=O)C’, ‘C’, ‘[C@@]’
‘C1CCOC1’, ‘O1CCCC1’, ‘[O]1CCCC1’,	‘S(=O)(=O)O’, ‘[N@@+]’, ‘[NH2+]’
‘OC[C@H](O)[C@@H](O)[C@H](O)CO’	‘[C@@H]’, ‘[OH+]’, ‘[CH2-]’
‘C(C1[C@H](C(C(C(O1)O)O)O)O [C@H]2C(C(C(C(O2)CO)O)O)O)O’	‘[P@+]’, ‘[Cl+2]’, ‘[S@@]’, ‘[Si@]’, ‘[BH3-]’, ‘CC(O)C’, ‘#’, ‘[O-]’

相关模型	二维结构重建结果准确率
HierVAE^[29]	0.799
NC-VAE（本研究）	0.630
JT-VAE^[30]	0.585
CG-VAE^[31]	0.424
CVAE^[32]	0.215

相关模型	二维结构重建结果准确率
HierVAE^[29]	0.799
NC-VAE（本研究）	0.630
JT-VAE^[30]	0.585
CG-VAE^[31]	0.424
CVAE^[32]	0.215

纳米纤维素及衍生物	模型目标	模型预测	准确率/%
磷酸化纳米纤维素	O=P(O)(O)OC[C@H]1OC[C@H](O)[C@@H](O)[C@@H]1O	O=C(C((CCOC[C@H]1OCC(OC[C@@H][C@@H]O	70.45
TEMPO氧化纳米纤维素	O=C(O)[C@H]1OC[C@H](O)[C@@H](O)[C@@H]1O	O=CO[C@H][C@H]O[C@H][C@H](O[C@@H]1O	79.49
磷酸化纳米纤维素Ⅱ	O=[P@H](O)OC[C@H]1OC[C@H](O)[C@@H](O)[C@@H]1O	C=[P@H]OO[C@H]O[C@H][C@H][C@H][C@H]C[C@H][C@H][C@H]O[C@@H]1O	58.33
纳米纤维素 3,5-二甲基苯基氨基甲酸酯	C[c:6]1[cH:1][c:2]([CH3:1])[cH:3][c:4](NC(=O)OC[C@@H]2C[C@H] (OC(=O)N[c:4]3[cH:3][c:2]([CH3:1])[cH:1][c:6](C)[cH:5]3)[C@@H] (OC(=O)N[c:4]3[cH:3][c:2]([CH3:1])[cH:1][c:6](C)[cH:5]3)[C@H](O)O2)[cH:5]O1	O[c:6]1O[c:2]O[CH3:1]O[cH:3]OONCOOOOOCO2C[C@H]([C@H][C@H]([C@H][C@H][C@H]N[c:4]3[cH:3][C@H])[c:6]()[cH:5]3)C=)[cH:3][cH:1][c:6]()[cH:5])[C@H](O1	56.85
高碘酸氧化纤维素	O=CCO[C@H](CO)[C@@H](O)C=O	O=CO([C@H]O[C@H][C@H](O	69.23