1. Predicting patterns of biological performance using chemical substructure features Diego Borges-Rivera 08/04/08

2. Introduction  cheminformatics – allow us to computationally describe similarity  synthetic chemists – describe through visual inspection  we will describe compounds by the presence of chemical substructures  we will attempt to identify sets of substructures that predict biological performance 1011101000101 0101000101101 0101000101101 1

3. Previous work  Clemons/Kahne/Wagner et al. -- disaccharide profiling in multiple cell states  found sets of substructures relevant to biological activity patterns  substructures highly specific to disaccharides 10 20 30 40 50 60 substructures

4. Biological performance profile  400 compounds, 8 assays in duplicate  tested for cell proliferation in 8 different cell lines  class labels are active (A) or inactive (I) active compound

5. What are fingerprints?  compound collection fed into commercial software  each substructure = 1 bit  the fingerprint shows which substructures are present substructure #1725 substructure #886 substructure #7017

6. Overview of cheminformatic methods  produced fingerprints  7700 total substructures  filtered set  left 2166 substructures

7. feature (substructure) selection to find predictive subsets evaluate methods for predictive value Overview of computational methods  two steps independent of each other

8. ReliefF: substructure selection +10 2166 weights Top 5 Bottom 5

9. K nearest neighbors (knn): predictive accuracy  Examples: k = 2, 5 compound being classified = ?

10. Similarity between compounds  similarity between two fingerprints  Tanimoto coefficient  this is used twice: (1) in ReliefF (2) in knn Example: Compound a: 0 0 1 Compound b: 1 0 1 Tanimoto coefficient = 1 / 2 =.5

11. Cross-validation: predictive accuracy  10 subsets  test set: one of the subsets  training set: the remaining subsets test set training set

12. Picking parameters for methods  which parameters produce the best predictive accuracies number of neighbors used in ReliefF {1, 2, 4, etc} number of neighbors used in knn {1, 2, 4, etc} number of ReliefF substructures used to predict classes in knn {1, 20, 100, etc}

13. Picking number of substructures predictive accuracy 1.0.9.8.7.6.5.4.3.2.1 0.0 1 20 all number of substructures used to predict

14. Group of substructures best able to predict

15. Future work  multi-class  different feature selection

16. Acknowledgements Computational Chemical Biology Joshua Gilbert Paul Clemons Hyman Carrinski Summer Research Program in Genomics Shawna Young Lucia Vielma Maura Silverstein