Recursive Neural Networks for Undirected Graphs for Learning Molecular Endpoints
Accurately predicting the endpoints of chemical compounds is an important step towards drug design and molecular screening in particular.
Here we develop a recursive architecture that is capable of mapping Undirected Graphs into individual labels, and apply it to the prediction of a number of different properties of small molecules. The results we obtain are generally state-of-the-art.
The final model is completely general and may be applied not only to prediction of molecular properties, but to a vast range of problems in which the input is a graph and the output is either a single property or (with small modifications) a set of properties of the nodes.
KeywordsSupport Vector Machine Root Mean Square Error Undirected Graph Aqueous Solubility Multi Layer Perceptron
- 17.Frasconi, P.: An introduction to learning structured information. J. Chem. Inf. Comput. Sci. 1387/1998, 99 (2004)Google Scholar
- 19.Bernazzani, L., Duce, C., Micheli, A., Mollica, V., Sperduti, A., Starita, A., Tiné, M.: Predicting physical-chemical properties of compounds from molecular structures by recursive neural networks. Applied Intelligence 19(1-2), 9–25 (2003)Google Scholar
- 27.The open babel package version 2.1.1, http://www.openbabel.org/
- 32.Molecular diversity preservation international database, http://www.mdpi.org/