GREYC's Chemistry dataset

This page contains several archive or links towards various chemical databases of molecules. Each database concerns a specific problem (either classification or Regression).

Achiral Molecules

• Monoamine Oxydase (MAO) dataset (classification problem). The class are specified in the file dataset.ds. This dataset is composed of 68 molecules divided into two classes: 38 molecules inhibit the monoamine oxidase (antidepressant drugs) and 30 do not.

  Results bellow are obtained for each method using a leave one out procedure with a two-class SVM. This classification scheme is made for each of the 68 molecules of the dataset.

  Classification accuracy on the monoamine oxidase (MAO) dataset. Method

  	Method	Classification accuracy
  (1)	Suard et al. (2002)	80% (55/68)
  (2)	Vishwanathan et al. (2010)	82% (56/68)
  (3)	Neuhaus and Bunke (2007)	90% (61/68)
  (4)	Riesen et al. (2007)	91% (62/68)
  (5)	Normalized standard Graph Laplacian kernel	90% (61/68)
  (6)	Normalized fast Graph Laplacian kernel	90% (61/68)
  (7)	Mahé and Vert (2008)	96% (65/68)
  (8)	Gaüzère et al. (2012)	94% (64/68)

• A database of acyclic molecules with hetero atoms (acyclic ethers, peroxides, acetals and their sulfur analogues). The dataset is concerned with the determination of the boiling point using regression (see dataset.ds)

  In both results presented bellow GLK stands for graph Laplacian kernel.

  Boiling point prediction on acyclic molecule dataset using a leave one out procedure. Method Average error (ºC) RMSE (ºC) (1) Cherqaoui et al. (1994a) 3.01 5.102 (2) Neuhaus and Bunke (2007) 6.84 9.04 (3) Riesen et al. (2007) 7.39 9.94 (4) GLK (diffusion process) 7.27 9.84 (5) Suard et al. (2002) 7.28 12.37 (6) Vishwanathan et al. (2010) 14.64 18.95 (7) Mahé and Vert (2008) 5.53 9.50 (8) Gaüzère et al. (2012) 4.90 7.80 (9) Gaüzère et al. with forward selection 3.23 4.36 (10) Gaüzère et al. with backward elimination 2.63 3.70

  Boiling point prediction on acyclic molecule dataset using 90% of the dataset as train set and remaining 10% as test set. Method Average error (ºC) RMSE (ºC) (1) (2) Neuhaus and Bunke (2007) 7.78 10.27 (3) Riesen et al. (2007) 7.66 10.19 (4) GLK (Diffusion Process) 8.58 11.99 (5) Suard et al. (2002) 7.22 12.24 (6) Vishwanathan et al. (2010) 14.57 18.72 (7) Mahé and Vert (2008) 6.42 11.02 (8) Gaüzère et al. (2012) 5.27 8.10 (9) Gaüzère et al. with forward selection 4.89 7.05 (10) Gaüzère et al. with backward elimination 4.87 6.75

• The Predictive Toxicology Challenge (PTC) dataset (classification problem). Please see this page for a full documentation.
• Alkane database(regression problem) This is a purely structural database (only carbons) with a problem of regression (boiling points data in dataset.ds).

  Results displayed bellow have been obtained using several test sets composed of 10% of the database, the remaining 90% being used as training set. For each test molecules composing the test and training sets have been choosen randomly.

  Boiling point prediction on alkane dataset. RMSE stands for root mean squared error.

  	Method	Average error (ºC)	RMSE (ºC)
  (1)	Cherqaoui and Villemin (1994)	3.11	3.70
  (2)	Neuhaus and Bunke (2007)	5.42	10.01
  (3)	Riesen et al. (2007)	5.27	7.10
  (4)	Suard et al. (2002)	4.66	6.21
  (5)	Vishwanathan et al. (2010)	10.61	16.28
  (6)	Graph Laplacian kernel	10.79	16.45
  (7)	Mahé and Vert (2008)	2.41	3.48
  (8)	Gaüzère et al. (2012)	1.41	1.92

• HIV database (classification problem) contains 42,687 compounds evaluated for evidence of anti-HIV activity by DTP AIDS Antiviral Screen of the National Cancer Institute. 422 molecules are confirmed active, 1081 moderately active and 41,184 inactive.
• Polyciclic Aromatic Hydrocarbons dataset (PAH) This dataset is composed cyclic unlabeled graphs. All atoms are carbons, all bounds are aromatics. Note that few acyclic bounds connect some atoms to cycles. This is a classification problem (cancerous or not cancerous molecules)

Classification of cyclic molecules solely composed of carbons using a 10-fold cross classification

	Method	Classification accuracy
(1)	Treelet Kernel[5]	71.3
(2)	Graph Edit distance[8]	72
(3)	Cycles [16]	63
(4)	Relevant Cycle graph [6]	77.7
(5)	Relevant Cycle Hypergraph [7]	76.3
(6)	Augmented Cycle [8]	80.7

• Monoterpenoide dataset This dataset is composed different classes of monoterpenoides molecules encoded by labeled graphs. This is a classification problem with 10 classes (Molecules are grouped according to the existence of a same precursor). This dataset is thus well suited for k-mean like algorithms.
• You may also be interested by the datasets of the University of California, Irvine

Chiral Molecules

• Optical rotation of chiral molecules.

  A dataset of acyclic molecules with chiral atoms. Molecules are provide with their optical rotation for a regression problem. Each molecules is composed of one or two chiral vertices (mean number : 1.06).

  The standard deviation of optical rotation angles is equal to 38.25º with values ranging from −89º to 78º .

Optical rotation angle of molecules with stereoisomerism.

	Method	RMSE (angle)
(1)	Mahé and Vert (2008)	34.1
(2)	Gauzere B. et al, 2012	26.2
(3)	Brown J. et al. 2010	24.2
(4)	Grenier P.-A. et al, 2013	14.8

• Dataset of chiral derivative of the Vitamin D.

  This dataset is composed of 69 derivatives of the vitamin D. Molecules are provided with their biological activity for a regression problem. After normalization the standard deviation of biological activity is equal to 0.256259.

Biological activity

	Method	RMSE
(1)	Mahé and Vert (2008)	0.251
(2)	Gauzere B. et al, 2012	0.271
(3)	Brown J. et al. 2010	0.184
(4)	Grenier P.-A. et al, ICPR 2014	0.194
(5)	Grenier P.-A. et al, ICPR 2014 kernel multiplied by Gauzere B. et al, 2012 kernel	0.191
(6)	Grenier P.-A. et al, S+SPR2014	0.180

• ACE Dataset

  This dataset is composed of all the stereoisomers of the perindoprilate. As this molecule has 5 stereocenters, the dataset is composed of 32 molecules. This is a classification problem where each molecule which inhibit the angiotensin-converting enzyme (ACE) correspond to a class, and those who do not inhibit it to the other class.

  Please consult Castillo-Garit 2007 for further details on this dataset.

  Biological activity

  	Method	Accuracy
  (1)	Treelet Kernel (Gauzere B. et al, 2012)	71.875
  (2)	Brown J. et al. 2010	96.875
  (3)	Grenier P.-A. et al, ICPR 2014	87.5
  (4)	Grenier P.-A. et al, S+SPR2014	96.875
  (4)	Interaction graph 1 (or 2), Grenier P.-A. et al, GbR2015	93.75
  (5)	Interaction graph 1 +MKL, Grenier P.-A. et al, GbR2015	100

• Steroid Dataset

  This dataset is composed of 64 steroid. One molecule is withdraw from the original database because this molecule is different from every other molecules and thus induced a bias on results. Molecules are provided with their molecular rotation for a regression problem. After normalization the standard deviation of molecular rotation is 0.208116. The mean number of stereocenters is 9.16.

  Further information about this dataset may be found in Bernstein S.(1941)

  Results obtained using the Mean absolute Error(MAE) during the validation step

  Ref	Method	MAE/RMSE
  Mahé, P., Vert, J.-P., 2008.	Tree pattern kernel	0.030/0.056
  Gauzere B. et al, 2012	Treelet Kernel	0.044/0.087
  Brown J. et al. 2010	Oriented tree pattern	0.030/0.070
  Grenier P.-A. et al, ICPR 2014	Stereo kernel	0.051/0.083
  Grenier P.-A. et al, GbR 2015	Graph of interaction 3 with Treelets	0.039/0.069

References

1. Bernstein, S., Kauzmann, W. J., & Wallis, E. S. (1941). The relationship between optical rotatory power and constitution of the sterols. The Journal of Organic Chemistry, 6(2), 319-330.
2. Castillo-Garit, J. A., Marrero-Ponce, Y., Torrens, F., & Rotondo, R. (2007). Atom-based stochastic and non-stochastic 3D-chiral bilinear indices and their applications to central chirality codification . Journal of Molecular Graphics and Modelling, 26(1), 32-47.
3. Cherqaoui, D., Villemin, D., 1994. Use of a neural network to determine the boiling point of alkanes. J. Chem. Soc. Faraday Trans. 90, 97 102.
4. Cherqaoui, D., Villemin, D., Mesbah, A., Cense, J.M., Kvasnicka, V., 1994a. Use of a neural network to determine the normal boiling points of acyclic ethers, peroxides, acetals and their sulfur analogues. J. Chem. Soc. Faraday Trans. 90, 2015 2019.
5. Gaüzère, B., et al. Two new graphs kernels in chemoinformatics. Pattern Recognition Lett. (2012), http://dx.doi.org/10.1016/j.patrec.2012.03.020.
6. Gaüzère B., Brun L., Villemin D., and Brun M., “Graph kernels based on relevant patterns and cycle information for chemoinformatics,” in Proceedings of ICPR 2012, vol. 7626, November 2012, pp. 1775–1778.
7. Gaüzère B., Brun L., and Villemin D., “Relevant cycle hypergraph representation for molecules,” in Graph-Based Representations in Pattern Recognition, Lecture Notes in Computer Science. Springer Berlin Heidelberg, 2013, vol. 7877, pp. 111–120
8. Gaüzère, Benoit, Brun, Luc & Villemin, Didier. Graph kernel encoding substituents relative positioning. In Proceedings of International Conference on Pattern Recognition (ICPR), 2014 , pages 6 p. , Stockholm(SE) , Sep 2014 .
9. Grenier, Pierre-Anthony, Brun, Luc & Villemin, Didier. Treelet Kernel Incorporating Chiral Information. In 9th IAPR-TC15 International Workshop on Graph-based Representations in Pattern Recognition , Vienne, Autriche , May 2013
10. Mahé, P., Vert, J.-P., 2008. Graph kernels based on tree patterns for molecules. Machine Learn. 75 (1), 3 35.
11. Neuhaus, M., Bunke, H., 2007. Bridging the Gap between Graph Edit Distance and Kernel Machines. World Scientific Pub Co Inc..
12. dd Riesen, K., Neuhaus, M., Bunke, H., 2007. Graph embedding in vector spaces by means of prototype selection. In: Escolano, F., Vento, M. (Eds.), 6th IAPR-TC15 Internat. Workshop GbRPR 2007. IAPR TC15. Springer-Verlag, pp. 383 393.
13. Suard, F., Rakotomamonjy, A., Bensrhair, A., 2002. Kernel on bag of paths for measuring similarity of shapes. In: European Symposium on Artificial Neural Networks. pp. 355 360.
14. Vishwanathan, S., Borgwardt, K.M., Kondor, I.R., Schraudolph, N.N., 2010. Graph kernels. J. Machine Learn. Res. 11, 1201 1242.
15. J. Brown, T. Urata, T. Tamura, M. A. Arai, T. Kawa- bata, and T. Akutsu. Compound analysis via graph ker- nels incorporating chirality. Journal of Bioinformatics and Computational Biology , 8(1) :63–81, 2010
16. Grenier, Pierre-Anthony, Brun, Luc & Villemin, Didier. Incorporating Molecule's Stereisomerism within the Machine Learning Framework. In Structural, Syntactic, and Statistical Pattern Recognition - Joint IAPR International Workshop, S+SSPR 2014, Joensuu, Finland, August 20-22, 2014. Proceedings , pages 12--21 2014 .
17. Grenier, Pierre-Anthony, Brun, Luc & Villemin, Didier. A Graph Kernel incorporating molecule's stereisomerism information. In Proceedings of ICPR 2014 , pages - , Stockholm, Suede , Aug 2014 .
18. Grenier, Pierre-Anthony, Brun, Luc & Villemin, Didier. From bags to graphs of stereo subgraphs in order to predict molecule’s properties In Proceedings of GbR2015, to be published by LNCS.
19. T. Horvath, “Cyclic pattern kernels revisited,” in Proceedings of KDD’04, Springer-Verlag, Ed., vol. 3518, 2005, pp. 791 – 801.