Chemical Similarity Searching with neural graph matching methods

The thesis examines three novel structural similarity methods that employ a network of simple auto-associative neural networks for storing structural information about databases of molecular graphs. This information can be used to discover similarities from a query graph to any of the graphs in the model database. The fast learning and recall ability of the neural network facilitates efficient approximate matching of graphs. The first method is based on the theory of relaxation labelling to be discussed in chapter 4. The consistencies between label pairs for different edge labels are stored within the associative memories. The discrete relaxation version initialises a set of vertex candidate labels from the set of model graphs for each of the query vertices. The support for each of the candidate labels is gauged by summing the number of consistent query neighbour vertices that have at least one candidate label which is consistent with the neighbour candidate vertex. By specifying a minimum required support, candidate labels can be eliminated which are deemed implausible if they do not achieve the minimum threshold. This algorithm is applied to different molecular data sets in chapter 5. The motivation for its application is that it is possible to search for similar molecules to a known drug in the hope that these molecules are going to exhibit similar properties in these assays. The comparison against conventional graph matching techniques reveals a slight deterioration in search effectiveness due to the approximate nature of the neural relaxation algorithm. However, it is often faster than the other methods and different types of encoding mechanisms do not cause the new method to time out as they do in some conventional methods. The second method corresponds to a novel graph clustering technique which is examined in chapter 6. Here, the approximate mean graph of a cluster of similar graphs is examined on a data set of small images and on molecular databases. Since the computation of the mean graph is computationally expensive, the neural relaxation algorithm from the previous chapter was employed as a pre-processor in order to prune the search space. The approximate mean graph is subsequently used as the input to the similarity search. The results showed a significant increase in the effectiveness in comparison to the search with single reference graphs. The third method denotes a different approach to structural similarity which is discussed in chapter 7. The graphs in the database are decomposed into sets of subgraphs which are stored in a network of binary neural networks. Re-occuring subgraphs are only stored once and are re-used whenever they are encountered. These subgraphs can be treated as features of the graphs and the number of common features between graphs signifies their degree of similarity. Assuming conditional independence of these features, it is possible to apply simple classification algorithms to discriminate between different classes of graphs.