Transferring Graph Neural Network Models for Predicting Bond Dissociation Energy between Datasets

Yao-Yuan Huo,
Jun Jiang^,

Hefei National Laboratory for Physical Sciences at the Microscale, School of Chemistry and Materials Science, University of Science and Technology of China, Hefei 230026, China

More Information

Corresponding author: E-mail: jiangj1@ustc.edu.cn

摘要

- /
- /
Abstract: Machine learning (ML) approaches like neural networks have been widely used in chemical researches for fast estimating chemical properties. Generating ML models of good precision requires datasets of high quality, which can be difficult to obtain. In this work, we trained graph neural network (GNN) models from different datasets and verified transferring of the models to other datasets. Our result shows that cross-dataset evaluation can give less accurate but still correlative prediction results on different datasets. Errors are mainly due to systematic errors. The value range of prediction result is highly related to the range of training set. The precisions of different bonds show different distributions. C–H bond constantly gets the highest precision in the tested bonds.
- Machine learning /
- Graph neural network /
- Cross validation

HTML全文

Figure 1. Distributions of three datasets including MG, BDE-db, and ZINC.

(A) Counts of heavy atoms in molecules. (B) BDEs of bonds.

下载: 全尺寸图片幻灯片

Figure 2. Bond types compositions in three datasets.

下载: 全尺寸图片幻灯片

Figure 3. Structure of input data for a molecule.

下载: 全尺寸图片幻灯片

Figure 4. The processing structure of a message passing layer in GNN.

下载: 全尺寸图片幻灯片

Figure 5. Illustration of GNN model structure. (A) A message passing layer. (B) Process of the whole network.

下载: 全尺寸图片幻灯片

Figure 6. Correlation plots of evaluations on test set of the same dataset.

(A) MG dataset. (B) BDE-db dataset.

下载: 全尺寸图片幻灯片

Figure 7. Correlation plots of evaluations on test set of different datasets.

(A) Trained on MG dataset and evaluated on BDE-db dataset. (B) Trained on MG dataset and evaluated on ZINC dataset. (C) Trained on BDE-db dataset and evaluated on MG dataset. (D) Trained on BDE-db dataset and evaluated on ZINC dataset. The data points are colorized by category of bond types and elements in the molecule.

下载: 全尺寸图片幻灯片

Figure 8. Comparison of prediction errors of different bonds using different models.

(A) Evaluation on BDE-db test set. (B) Evaluation on single bonds of CHON molecules of MG test set. (C) Evaluation on single bonds of CHON molecules of ZINC dataset. Bond types are ordered by occurrence counts in test set.

下载: 全尺寸图片幻灯片

Figure 9. BDEs of N–N bonds in diazo group acquired with GNN prediction and DFT calculation.

(A) MG GNN model and ZINC dataset. (B) MG GNN model and MG dataset.

下载: 全尺寸图片幻灯片

Figure 10. Typical bonds with the highest GNN predicted BDE. (A) MG model on BDE-db dataset. (B) MG model on single bonds of MG dataset.

下载: 全尺寸图片幻灯片

参考文献(20)

[1]	J. Wu and X. Xu, J. Chem. Phys. 127, 214105 (2007). doi: 10.1063/1.2800018
[2]	J. Wu, Y. Zhou, and X. Xu, Int. J. Quantum Chem. 115, 1021 (2015). doi: 10.1002/qua.24919
[3]	S. Urata, A. Takada, T. Uchimaru, A. K. Chandra, and A. Sekiya, J. Fluorine Chem. 116, 163 (2002). doi: 10.1016/S0022-1139(02)00128-8
[4]	X. Qu, D. A. R. S. Latino, and J. Aires-de-Sousa, J. Cheminformatics 5, 34 (2013). doi: 10.1186/1758-2946-5-34
[5]	B. Maryasin, P. Marquetand, and N. Maulide, Angew. Chem, Int. Ed. 57, 6978 (2018). doi: 10.1002/anie.201803562
[6]	M. Szwarc, Chem. Rev. 47, 75 (1950). doi: 10.1021/cr60146a002
[7]	G. Kresse and J. Furthmüller, Comput. Mater. Sci. 6, 15 (1996). doi: 10.1016/0927-0256(96)00008-0
[8]	F. Scarselli, M. Gori, and A. C. Tsoi, IEEE Trans. Neural Networks 20, 61 (2009). doi: 10.1109/TNN.2008.2005605
[9]	P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. F. Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, C. Gülçehre, H. F. Song, A. J. Ballard, J. Gilmer, G. E. Dahl, A. Vaswani, K. R. Allen, C. Nash, V. Langston, C. Dyer, N. Heess, D. Wierstra, P. Kohli, M. Botvinick, O. Vinyals, Y. Li, and R. Pascanu, Relational Inductive Biases, Deep Learning, and Graph Networks. arXiv preprint (2018), arXiv: 1806.01261.
[10]	P. C. St. John, Y. Guan, Y. Kim, S. Kim, and R. S. Paton, Nat. Commun. 11, 2328 (2020). doi: 10.1038/s41467-020-16201-z
[11]	M. Wen, S. M. Blau, E. W. C. Spotte-Smith, S. Dwaraknath, and K. A. Persson, Chem. Sci. 12, 1858 (2021). doi: 10.1039/D0SC05251E
[12]	Y. Kim, Y. Jeong, J. Kim, E. K. Lee, W. J. Kim, and I. S. Choi, Chem. Asian J. 17, e202200269 (2022). doi: 10.1002/asia.202200269
[13]	C. W. Coley, W. Jin, L. Rogers, T. F. Jamison, T. S. Jaakkola, W. H. Green, R. Barzilay, and K. F. Jensen, Chem. Sci. 10, 370 (2019). doi: 10.1039/C8SC04228D
[14]	E. Mansimov, O. Mahmood, S. Kang, and K. Cho, Sci. Rep. 9, 20381 (2019). doi: 10.1038/s41598-019-56773-5
[15]	D. Chen, K. Gao, D. D. Nguyen, X. Chen, Y. Jiang, G. W. Wei, and F. Pan, Nat. Commun. 12, 3521 (2021). doi: 10.1038/s41467-021-23720-w
[16]	M. Dolg, Energy-Consistent Pseudopotentials of the Stuttgart/Cologne Group.
[17]	S. Grimme, J. Antony, S. Ehrlich, and H. Krieg, J. Chem. Phys. 132, 154104 (2010). doi: 10.1063/1.3382344
[18]	T. Lu and F. Chen, J. Mol. Model. 19, 5387 (2013). doi: 10.1007/s00894-013-2034-2
[19]	P. St. John, Y. Guan, Y. Kim, and S. Kim, Bde-Db: A Collection of 290 664 Homolytic Bond Dissociation Enthalpies for Small Organic Molecules, Figshare (2019).
[20]	X. Gonze, P. Ghosez, and R. Godby, Phys. Rev. Lett. 74, 4035 (1995). doi: 10.1103/PhysRevLett.74.4035