Empirical Validation of the Effectiveness of Chemical Descriptors in Data Mining Kirk Simmons DuPont Crop Protection Stine-Haskell Research Center 1090 Elkton Road Newark, DE
The Study Purpose Strategy –Methods –Metrics Results Practical Application Conclusions
Purpose Chemical Structure Conference (1996) – Holland –Data mining/similarity methodologies reported –Used numerous descriptor sets –No standard datasets –Comparisons difficult Comparative study of chemical descriptors across varied biology
Strategy Systematically evaluate descriptors within a compound dataset across multiple biological endpoints All compounds have experimentally measured endpoints Diversity of biological endpoints –In-Vitro (receptor affinity, enzyme inhibition) –In-Vivo (insect mortality) Explored nine common descriptor sets Train and then use model to forecast a validation set
Methods Four In-Vitro assays –48K compound dataset for training –Corporate database for validation Two In-Vivo assays –75-100K compound datasets –Randomly divided into training and validation subsets Recursive Partitioning - analytic method –Appropriate method for HTS data –Selected statistically conservative inputs (p-tail < 0.01)
Metrics 4-way Interaction –Analytic Method, Compound Set, Biology, and Descriptors Efficiency of analysis (Lift Chart) –Fraction of Actives found/Fraction of Dataset tested –Rewards efficiency only Effectiveness of analysis (Composite Score) –Fraction of Actives found x Efficiency –Rewards efficiency as well as completeness
Results - Training
Results - Forecasting
Averaged Results - Training
Averaged Results - Forecasting
Practical Application RP-based models using screening data on 3 targets –Activity treated as active/inactive –DiverseSolutions R BCUT descriptors RP-models used to forecast vendor compounds (1M) Selected compounds purchased/screened –Hit-rates improved 530% over training sets –New structures and improved activity
Historical Screening Results
RP-based Screening Results
Results Comparison
Conclusions Not all chemical descriptors equally effective –Whole molecule property-based less effective –Chemical feature-based appear more effective Training models effectiveness –Averaged 28% of theory –Room for 4-fold improvement Validation models effectiveness –Averaged 16% of theory –Room for 6-fold improvement
Acknowledgements Dr. Linrong Yang, FMC Corporation –Completed the work FMC Corporation –Release of the results Prof. Peter Willett, University of Sheffield Prof. Alex Tropsha, University of North Carolina Prof. Doug Hawkins, University Minnesota DuPont Corporation