A novel genetic algorthm for neural computing to identify quantitative structure activity relationship (QSAR), named mutation- based genetic algorithm (MGA), ispresented. MGA only uses the mutation operator for local search. To enhance the efficiency of local search, the genes that represent the variables employ different time-varying mutation rates in MGA. Combining random restart technique with the local search strategy, the algorthm can give satisfactory solution in a limited time. As a typical object of the neural comuting for QSAR, a set of 74 2,4-dialmino-5-(substituted benzyl) pyrimidines that inhibit dihydrofolate reductase were used to verify the effectiveness of MGA in computings of predicting bio-activity. Cross- validation trials and the test of predicting activity demonstrated that the predictive ability of the QSAR model built with MGA is better than those provided by other methods.

Key words: MOLECULAR DESIGN, NEURONS, PHARMACEUTICAL CHEMISTRY, QUANTITATIVE STRUCTURE ACTIVITY RELATIONSHIP