Drugdesign bymachine learning: Theuseofinductive logic programming tomodelthestructure-activi ty relationships of trimethoprim analogues binding todihydrofolate reductase (arfclintegence/ee acv/prote l/active sites)

Themane rning program GOLEMfrom thefield ofinductive logic ngwasapplied tothe drugdesign problem ofmding v-activity relation- ships. Thetraining data fortheprogam were44trlmehoprim analogues andtheir observed inhibition ofEscherichia colU dihydrofobte reductase. Afurther 11compounds wereused as unseen test data. GOLEMobtained rules that werestat y moreaccurate ontherinindataandalso beiter onthetest datathana schl regression model. Importanty machinelearningyi understandable rlsthat character- ized thechemistry offavored inhibitors interms ofpoluity, flexibili, andhydrogen-bonding character. These rules agree withthestereochemistry oftheinteraction observed crystallo- graphically. Thedesign ofapotent pharmaceutical agent fromalead compound isoften based onanunderstanding ofthequan- titative structure-activi ty relationship (QSAR) inarelated