We investigate the effectiveness of using a tree comparison based method to screen for drug candidates. Molecules are represented as trees in which ring systems are reduced to single nodes. These trees are compared to the tree of a selected known binder and the molecules are ranked according to the normalized size of their largest common subtree. The nodes of the molecular trees contains information about the atoms or ring systems they represent (e.g. charge and hydrogen donor/acceptor properties). In this way we can restrict which nodes are matched when calculating the size of the largest common subtree. Our experiments show that this results in a speed up of the calculation and an improvement in the quality of the result. We test our method on the DUD (Directory of Useful Decoys) test set and compare it to other ligand based screening methods (topological descriptors and CDK fingerprints), using the enrichment and AUC (Area Under Curve) as a performance measure. We also compare the performance with the docking program DOCK. The results of our experiments indicate that our method performs better than the topological descriptor and that it is comparable to the CDK fingerprint. The docking based method is by far the worst performing methods, which is probably due to the design of the DUD set. Our comparability to the other ligand based method is very promising and indicate that our method is well suited for further investigation and improvement.
Main Research Area:
17th Annual International Conference on Intelligent Systems for Molecular Bioligy (ISMB) & 8th European Conference on Computational Biology (ECCB), 2009