Incorporating Molecule’s Stereisomerism within the Machine Learning Framework

Abstract

An important field of chemoinformatics consists in the prediction of molecule's properties, and within this field, graph kernels constitute a powerful framework thanks to their ability to combine a natural encoding of molecules by graphs, with classical statistical tools. Unfortunately some molecules encoded by a same graph and differing only by the three dimensional orientation of their atoms in space have different properties. Such molecules are called stereoisomers. These latter properties can not be predicted by usual graph methods which do not encode stereoisomerism. In this paper we propose to encode the stereoisomerism property of each atom of a molecule by a local subgraph. A kernel between bags of such subgraphs provides a similarity measure incorporating stereoisomerism properties. We then propose two extensions of this kernel incorporating in each sub graph information about its surroundings.

From Bags to Graphs of Stereo Subgraphs in Order to Predict Molecule’S Properties

“…Methods which do not encode stereoisomerism information [10,3] obtain poor results as we can see in Table 3 (lines 1-2). The adaptation of the tree pattern kernel to stereoisomerism [1] and our previous kernels [7,5] (lines 3-5) improves the results over the two previous methods hence showing the insight of adding stereoisomerism information. Taking into account relationships between minimal stereo subgraphs (lines 6-8) allows us to obtain better results than our previous method [5].…”

“…The adaptation of the tree pattern kernel to stereoisomerism [1] and our previous kernels [7,5] (lines 3-5) improves the results over the two previous methods hence showing the insight of adding stereoisomerism information. Taking into account relationships between minimal stereo subgraphs (lines 6-8) allows us to obtain better results than our previous method [5]. Unlike the previous dataset, the graphs of interactions G 2 and G 3 have higher degree (≈ 2) and thus obtains good results.…”

“…In this dataset two stereocenters have a same minimal stereo subgraph, but different surrounding. The stereo kernel [7] and one of the graph of interactions (G 3 ), can not differentiate those two stereocenters, which have different influence on the property, this explains why other method ( [1,5] and the two other graphs of interactions) obtain a better accuracy. However, for our graphs of interactions, treelet of size one have a negative effect on the classification, this explains why we do not obtain better results than [1,5].…”

“…The stereo kernel [7] and one of the graph of interactions (G 3 ), can not differentiate those two stereocenters, which have different influence on the property, this explains why other method ( [1,5] and the two other graphs of interactions) obtain a better accuracy. However, for our graphs of interactions, treelet of size one have a negative effect on the classification, this explains why we do not obtain better results than [1,5]. By using a multiple kernel learning algorithm [12], we can learn a weight for each treelet, that allow us to discard treelet of size 1 and to obtain the best results with the first graph of interactions.…”