A steepest descent method for optimizing the Carbó molecular similarity index was implemented and evaluated. Comparisons were made between this procedure and the extensively used simplex method. Several data sets were considered, and in each case the gradient method showed a substantial improvement in the time taken for the optimization to converge while comparable similarity values were obtained. In some cases, performance enhancements of up to an order of magnitude were observed. © 1997 by John Wiley & Sons, Inc.