机器学习设计新型有机分子研究进展

doi:10.6023/cjoc202012037

有机化学 ›› 2021, Vol. 41 ›› Issue (7): 2666-2675.DOI: 10.6023/cjoc202012037 上一篇下一篇

综述与进展

机器学习设计新型有机分子研究进展

谭胖^a^,^b, 刘旭红^a^,^c, 谌彤童^d, 秦智慧^a^,^b, 杨涛^a, 刘晓彤^a^,^e^,^f^,^g, 刘秀磊^a^,^b^,^*()

a 北京信息科技大学北京材料基因工程高精尖创新中心北京 100101
b 北京信息科技大学数据与科学情报分析实验室北京 100101
c 北京信息科技大学网络文化与数字传播北京市重点实验室北京 100192
d 北京跟踪与通信技术研究所北京 100094
e 中国科学院煤炭化学研究所煤转化国家重点实验室太原 030001
f 中科合成油技术有限公司国家能源煤基液体燃料研发中心北京 101400
g 中国科学院大学北京 100049

收稿日期:2020-12-22 修回日期:2021-02-08 发布日期:2021-03-04
通讯作者: 刘秀磊
基金资助:
北京信息科技大学“勤信人才”培育计划; 北京信息科技大学促进高校内涵发展; 北京市教育委员会科技计划一般项目(KM202111232003); 北京市自然科学基金(4204100)

Research Progress on New Organic Molecules Design via Machine Learning

Pang Tan^a^,^b, Xuhong Liu^a^,^c, Tongtong Chen^d, Zhihui Qin^a^,^b, Tao Yang^a, Xiaotong Liu^a^,^e^,^f^,^g, Xiulei Liu^a^,^b()

a Beijing Advanced Innovation Center for Materials Genome Engineering, Beijing Information Science and Technology University, Beijing 100101
b Laboratory of Data Science and Information Studies, Beijing Information Science and Technology University, Beijing 100101
c Beijing Key Laboratory of Internet Culture and Digital Dissemination Research, Beijing Information Science and Technology University, Beijing 100192
d Beijing Institute of Tracking and Telecommunications Technology, Beijing 100094
e State Key Laboratory of Coal Conversion, Institute of Coal Chemistry, Chinese Academy of Sciences, Taiyuan 030001
f National Energy Center for Coal to Liquids, Synfuels China Co., Ltd, Beijing 101400
g University of Chinese Academy of Sciences, Beijing 100049

Received:2020-12-22 Revised:2021-02-08 Published:2021-03-04
Contact: Xiulei Liu
Supported by:
Qin Xin Talents Cultivation Program of Beijing Information Science & Technology University; Beijing University of Information Science and Technology to Promote the Development of the Connotation of Colleges and Universities; Beijing Education Commission for General Project of Science and Technology Plan(KM202111232003); Beijing Municipal Natural Science Foundation(4204100)

文章分享

摘要/Abstract

新型有机分子一直是有机化学领域的研究重点, 其在开发高性能材料方面具有重要意义. 传统的有机分子发现是一个类似于“炒菜”的试错过程, 它耗时耗能且效率相对低下. 常见的量子化学方法试图根据期望属性值筛选出合理的分子结构, 以更好地指导实验, 然而, 由于计算资源相对于算法复杂度严重不足, 精确给出实验指导在大多数情况下难以实现. 近年来机器学习的出现改变了这种情况, 训练好的模型可以快速推测出分子的属性. 更令人兴奋的是机器学习可以逆向进行分子设计, 拓宽人类的想象力, 给出其在分子设计领域的“神之一手”. 本综述首先介绍了逆向分子设计所必须的分子描述方式, 随后对几种常见的深度生成模型加以归纳, 对新型有机分子设计研究现状进行了总结, 最后探讨了新型有机分子设计所面临的挑战, 展示了笔者做出的部分探索.

关键词: 机器学习, 生成模型, 逆向分子设计, 分子描述, BASE64编码

Low-cost and high-performance materials have become more and more important in past decades. It exhibits the technology level of a country. Chemists used to find the candidate material according to property regression and quantitative structure activity relationship (QSAR). Traditional methods focus on finding new molecule from prior knowledge with trial and error experiments. They are time-consuming and low efficiency on screening molecules. The appearance of machine learning (ML) changes this embarrassing situation in two ways. One is accelerating the property prediction process to prevent wasting time on worse candidates. The other is inverse molecule design which expands the imagination of human. Lots of researches show promising results using different inverse design method such as, variational auto-encoder (VAE), generative adversarial networks (GAN), reinforcement learning (RL), and recurrent neural network (RNN). They introduce uncertainty from different level to generate new structure candidates. In any method, molecule descriptor has a great impact on the result. The descriptor converts the 3D structures in real world to a vector or a notation string to feed into all kinds of ML models. Large number of descriptors have been developed in cheminformatic, bioinformatic, quantum chemistry and natural language process (NLP). Some classical descriptors are Coulomb matrix (CM), smooth overlap of atomic positions (SOAP), weighted graph (WG), simplified molecular input line entry specification (SMILES). They show different advantages and solving problems from different aspects. CM has clear definition and good result on energy regression. SOAP is good at reflecting local environment features of an atom. However, they are easy to encode but hard to decode. That is a reason why people prefer WG and SMILES in the structure inverse design tasks. WG and SMILES express structure as a graph (an atom as a node and a bond as an edge) or string to apply massive mature GNN or NLP algorithm on them. Nowadays, most of the ML applications on chemistry and molecule science are focus on developing new model to regress properties. However, it is thought that there is still large improving space on inverse design methods and traditional descriptors. In this paper, WG and SMILES are briefly introduced firstly. Then, four generative models are presented, including VAE, GAN, RL and RNN. Further, the current progress and challenges of inverse design methods are summarized case by case. Finally, some of the author՚s understanding and explorations are given out. It is proved that SMILES with BASE64 preprocessed shows some advantages on molecular reconstruction and worth to study deeply in future.

Key words: machine learning, generative model, inverse molecule design, molecule description, BASE64 encoding

引用此文

谭胖, 刘旭红, 谌彤童, 秦智慧, 杨涛, 刘晓彤, 刘秀磊. 机器学习设计新型有机分子研究进展[J]. 有机化学, 2021, 41(7): 2666-2675.

Pang Tan, Xuhong Liu, Tongtong Chen, Zhihui Qin, Tao Yang, Xiaotong Liu, Xiulei Liu. Research Progress on New Organic Molecules Design via Machine Learning[J]. Chinese Journal of Organic Chemistry, 2021, 41(7): 2666-2675.

导出引用 EndNote|Reference Manager|ProCite|BibTeX|RefWorks

分享此文

图/表 10

图1. 机器学习设计新型有机分子流程

Figure 1. Process of designing new organic molecule by machine learning

Molecule	SMILES
Ethane	CC
Carbon dioxide	O=C=O
Acetic acid	CC(=O)O
Benzene	c1ccccc1
Phenol	Oc1ccccc1

表1. SMILES实例(三种不同的分子)

Table 1. Examples of SMILES (three different molecules)

图2. 使用One-hot编码处理SMILES的示意图

Figure 2. Schematic diagram of using One-hot coding deal with SMILES

图3. 分子图转成标记图

Figure 3. Conversion process of molecule to labeled graph

图4. (a)变分自编码器整体结构和(b)变分自编码器设计新型有机分子实例

Figure 4. (a) Structure of variational auto-encoder, and (b) an example of designing new organic molecules with variational auto-en- coder

图5. (a)生成对抗网络训练过程和(b)生成对抗网络设计新型有机分子实例

Figure 5. (a) Training process of generative adversarial networks, and (b) an example of designing new organic molecules with generative adversarial networks

图6. (a)强化学习训练过程和(b)强化学习设计新型有机分子实例

Figure 6. (a) Training process of reinforcement learning, and (b) an example of designing new organic molecules with reinforcement learning

图7. (a) CVAE模型结构图: 编码器(Encoder)用于编码SMILES, 解码器(Decoder)用于将隐藏空间(Latent space)采样出的隐藏向量重构回SMILES, 属性预测器(Property predictor)用于预测隐藏空间采样出的隐藏向量的属性值; (b) ORGAN模型结构图: 左侧为训练判别器(Discriminator)过程, 判别器被训练为一个二分类模型, 右侧为优化生成器(Generator)过程, 依据判别器误差以及特定目标(Objectives)进行优化; (c) MolGAN模型结构图: 生成器(Generator)用于从分布Z~P(Z)中采样生成分子, 判别器(Discriminator)用于判断分子来自真实数据集还是来自生成器, 奖励网络(Reward network)用于估计分子的属性值, X~Pdata(X)用于得出分子的真实属性值; (d) ImatGen模型结构图: 第一步AE用于将分子编码成指纹图谱(image fingerprint), 第二步VAE用于将指纹图谱编码为隐藏空间(Latent space)

Figure 7. (a) Structure of CVAE: encoder generates SMILES; decoder reconstructs the latent vector to SMILES; property predictor infers the property value of the input. (b) Structure of ORGAN: the left is the training process of the discriminator (a binary classification model); the right is the training process of the generator which is optimized according to the error of discriminator and objectives. (c) Structure of MolGAN: generator creates molecules from the distribution Z~P(Z); discriminator judges whether the input molecules come from real data or generator; reward network estimates the property values of molecules; X~Pdata(X) is a function of property labels. (d) Structure of ImatGen: AE encodes molecules into image fingerprint; VAE encodes image fingerprint into latent space

Model	Dataset	Descriptor	Valid/%
CVAE	ZINC	SMILES	73.9^[23]
CVAE	QM9	SMILES	79.3^[23]
JT-VAE	QM9	Graph	100^[50]

表2. 各种模型所对应的数据源、描述符、有效分子百分比

Table 2. Data sources, descriptors, valid of molecules of various models

Method	Accuracy	Accuracy without padding	LogP	QED	SAS
Without BASE64	98.14%	95.08%	0.928	0.093	0.444
BASE64	98.53%	96.33%	0.934	0.092	0.440

表3. 未经过BASE64编码的方法与经过BASE64编码的方法结果对比a

Table 3. Comparison of the results between the methods without BASE64 coding and those with BASE64 coding.

参考文献 54

[1]	(a) Notomi, M.; Naganuma, M.; Nishida, T.; Tamamura, T.; Iwamura, H.; Nojima, S.; Okamoto, M. Appl. Phys. Lett. 1991, 58,720. doi: 10.1063/1.104526
	(b) Drews, J. Science 2000, 287(5460),1960. doi: 10.1126/science.287.5460.1960
[2]	(a) Zgou, H.; Hamidi, M.; Lére-Porte,J. -P.; Serein-Spirau, F.; Bouachrine, M. Acta Phys.-Chim. Sin. 2008, 24(1),37. doi: 10.1016/S1872-1508(08)60003-0
	(b) Qian, L.; Shen, Y.; Chen, J.; Zheng, K. Acta. Phys.-Chim. Sin. 2006, 22(11),1372. doi: 10.1016/S1872-1508(06)60069-7
[3]	Langer, M.; Goeßmann, A.; Rupp, M. arXiv: 2003.12081.
[4]	(a) Bohacek, R.; McMartin, C.; Guida, W. Med. Res. Rev. 1996, 16,3. doi: 10.1002/(ISSN)1098-1128
	(b) Ruddigkeit, L.; van Deursen, R.; Blum,L. C.; Reymond,J. -L. J. Chem. Inf. Model. 2012, 52(11),2864. doi: 10.1021/ci300415d
[5]	DiMasi,J. A.; Grabowski,H. G.; Hansen,R. W. J. Health Econ. 2016, 47,20. doi: 10.1016/j.jhealeco.2016.01.012
[6]	(a) Tan, N.; Li, J.; Li, Z.; Li, X. Acta. Phys.-Chim. Sin. 2006, 22(4),397. doi: 10.1016/S1872-1508(06)60011-9
	(b) LeCun, Y.; Bengio, Y.; Hinton, G. Nature 2015, 521(7553),436. doi: 10.1038/nature14539
	(c) Eraslan, G.; Avsec, Ž.; Gagneur, J.; Theis,F. J. Nat. Rev. Genet. 2019, 20(7),389.
	(d) Burbidge, R.; Trotter, M.; Buxton, B.; Holden, S. Comput. Chem. 2001, 26(1),5. doi: 10.1016/S0097-8485(01)00094-8
	(e) Liu, X.; Zhang, T.; Yang, T.; Liu, X.; Song, X.; Yang, Y.; Li, N.; Rignanese,G. -M.; Li, Y.; Wen, X. J. Phys. Chem. A 2020, 124(42),8866. doi: 10.1021/acs.jpca.0c06319
[7]	(a) Williams,R. J.; Zipser, D. Neural Comput. 1989, 1(2),270. doi: 10.1162/neco.1989.1.2.270
	(b) Yu, S.; Su, J.; Luo, D. IAcc 2019, 7,176600.
[8]	(a) Van Houdt, G.; Mosquera, C.; Nápoles, G. Artif. Intell. Rev. 2020.
	(b) Hou, H.; Xu, I.; Chen, M.; Liu, Z.; Guo, W.; Gao, M.; Xin, Y.; Cui, L. IAcc 2020, 8,90907.
[9]	(a) Cong, I.; Choi, S.; Lukin, M. Nat. Phys. 2019, 15,1.
	(b) Soydemir, D. Int. J. Intell. Syst. Appl. Eng. 2019, 7,222. doi: 10.18201/ijisae.2019457674
[10]	Wu, Z.; Pan, S.; Chen, F.; Long, G.; Zhang, C.; Yu, P. IEEE T. Neur. Net. Lear. 2021, 32,4.
[11]	Sanchez, B.; Aspuru-Guzik, A. Science 2018, 361,360. doi: 10.1126/science.aat2663
[12]	Schwalbe-Koda, D.; Gómez-Bombarelli, R. arXiv: 1907.01632v1.
[13]	Herbst, I. Commun. Math. Phys. 1974, 35,181. doi: 10.1007/BF01646192
[14]	Arora, A. Can. Med. Assoc. J. 2020, 192,E848. doi: 10.1503/cmaj.200092
[15]	Meyer, F. Pattern Recogn. Lett. 2014, 47,72. doi: 10.1016/j.patrec.2014.02.018
[16]	(a) Karabunarliev, S.; Ivanov, J.; Mekenyan, O. Comput. Chem. 1994, 18,189. doi: 10.1016/0097-8485(94)85010-0
	(b) Lin,T. -S.; Coley, C.; Mochigase, H.; Beech, H.; Wang, W.; Wang, Z.; Woods, E.; Craig, S.; Johnson, J.; Kalow, J.; Jensen, K.; Olsen, B. ACS Cent. Sci. 2019,5.
[17]	Arsham,D. H.; Davani, D.; Yu, J. Math. Comput. Simulat. 1993, 35,493. doi: 10.1016/0378-4754(93)90067-5
[18]	O'Boyle, N. J. Cheminform. 2012, 4,22. doi: 10.1186/1758-2946-4-22
[19]	Walsh,M. O. Am. Book Rev. 2012, 33,23.
[20]	Ikebata, H.; Hongo, K.; Isomura, T.; Maezono, R.; Yoshida, R. J. Comput.-Aided Mater. Des. 2017,31.
[21]	Ertl, P.; Lewis, R.; Martin, E.; Polyakov, V. arXiv: 1712.07449v2.
[22]	(a) Heller, S.; McNaught, A.; Pletnev, I.; Stein, S.; Tchekhovskoi, D. J. Cheminform. 2015,7.
	(b) Grethe, G.; Blanke, G.; Kraut, H.; Goodman, J. J. Cheminform. 2018,10.
[23]	Gómez-Bombarelli, R.; Duvenaud, D.; Hernández-Lobato, J.; Aguilera-Iparraguirre, J.; Hirzel, T.; Adams, R.; Aspuru-Guzik, A. ACS Cent. Sci. 2016,4.
[24]	(a) Bolano, A. Sci. Trends 2018.
	(b) Meziani, A. Proc. Am. Math. Soc. 1996,124.
	(c) Sandi-Urena, S. Chem. Tea. Inter. 2019.
[25]	Lusci, A.; Pollastri, G.; Baldi, P. J. Chem. Inf. Model. 2013,53.
[26]	Duvenaud, D.; Maclaurin, D.; Aguilera-Iparraguirre, J.; Gómez- Bombarelli, R.; Hirzel, T.; Aspuru-Guzik, A.; Adams, R. Adv. Neural Inf. Process. Syst. 2015,13.
[27]	Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; Ong,S. P. Chem. Mater. 2019, 31(9),3564. doi: 10.1021/acs.chemmater.9b01294
[28]	Ramakrishnan, R.; Dral,P. O.; Rupp, M.; von Lilienfeld,O. A. Sci. Data 2014, 1(1),140022. doi: 10.1038/sdata.2014.22
[29]	RDKit: http://www.rdkit.org
[30]	OpenSMILES:http://opensmiles.org/
[31]	MolVS: https://molvs.readthedocs.io/en/latest/
[32]	Turcani, L.; Berardo, E.; Jelfs, K. J. Comput. Chem. 2018, 39,1931. doi: 10.1002/jcc.v39.23
[33]	Liu, Y.; Lin, S.; Clark, R. Proc. AAAI Con. Artif. Intel. 2020, 34,13869.
[34]	Thada, D.; Shrivastava, U.; Sharma, J.; Singh, K.; Ranadeep, M. Int. J. Inn. Res. Com. Sci. Tech. 2020,8.
[35]	(a) Hinton, G.; Zemel, R. Adv. Neural Inf. Process. Syst. 1994,6.
	(b) Soydaner, D. Neural Pro. Lett. 2020.
	(c) Ferreira, D.; Silva, S.; Abelha, A.; Machado, J. Appl. Sci. 2020, 10,5510. doi: 10.3390/app10165510
[36]	Ponti, M.; Kittler, J.; Riva, M.; de Campos, T.; Zor, C. Pattern Recogn. 2017,61.
[37]	Burda, Y.; Grosse, R.; Salakhutdinov, R. arXiv: 1509.00519v4.
[38]	Kulkarni, T.; Whitney, W.; Kohli, P.; Tenenbaum, J. arXiv:1503.03167v4.
[39]	Arjovsky, M.; Chintala, S.; Bottou, L. arXiv:1701.07875v3.
[40]	Rueschendorf, L. Probab. Theory Rel. Fields 1985, 70,117. doi: 10.1007/BF00532240
[41]	Lamberti, P.; Majtey, A. Phys. A 2003, 329,81.
[42]	Miyato, T.; Kataoka, T.; Koyama, M.; Yoshida, Y. arXiv: 1802.05957v1.
[43]	Harding, A.; Mine, H.; Osaki, S. Oper. Res. Q. (1970-1977) 1971, 22,192.
[44]	Lek, S.; Park,Y. S. Encycl. Ecol. 2008,2455.
[45]	(a) Guimaraes, G.; Sanchez, B.; Farias, P.; Aspuru-Guzik, A. 2017.
	(b) Popova, M.; Isayev, O.; Tropsha, A. Sci. Adv. 2017,4.
[46]	Kusner, M.; Paige, B.; Hernández-Lobato, J. arXiv: 1703.01925v1.
[47]	(a) Johansen, S.; Juselius, K. Oxford Bull. Econ. Statist. 1990, 52,169. doi: 10.1111/obes.1990.52.issue-2
	(b) Chow, G. Econ. Modelling 1984, 1,134. doi: 10.1016/0264-9993(84)90001-4
[48]	Dai, H.; Tian, Y.; Dai, B.; Skiena, S.; Song, L. arXiv: 1802.08786v1.
[49]	De Cao, N.; Kipf, T. aarXiv: 1805.11973v1.
[50]	Jin, W.; Barzilay, R.; Jaakkola, T. arXiv: 1802.04364.
[51]	Jørgensen, M.; Mortensen, H.; Meldgaard, S.; Kolsbjerg, E.; Jacobsen, T.; Sørensen, K.; Hammer, B. J. Chem. Phys. 2019, 151,054111. doi: 10.1063/1.5108871
[52]	Noh, J.; Kim, J.; Stein,H. S.; Sanchez-Lengeling, B.; Gregoire,J. M.; Aspuru-Guzik, A.; Jung, Y. Matter 2019, 1(5),1370. doi: 10.1016/j.matt.2019.08.017
[53]	Jain, A.; Ong,S. P.; Hautier, G.; Chen, W.; Richards,W. D.; Dacek, S.; Cholia, S.; Gunter, D.; Skinner, D.; Ceder, G.; Persson,K. A. APL Mater. 2013, 1(1),011002. doi: 10.1063/1.4812323
[54]	Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez,A. N.; Kaiser, Ł.; Polosukhin, I. In Proceedings of the 31st International Conference on Neural Information Processing Systems, Curran Associates Inc., Long Beach, California, USA, 2017, p. 6000.

机器学习设计新型有机分子研究进展

Research Progress on New Organic Molecules Design via Machine Learning

RichHTML

PDF

可视化

摘要/Abstract

引用此文

分享此文

图/表 10

参考文献 54

相关文章 1

编辑推荐

相关统计

本文评价