Explorando o espaço de parâmetros do método semiempírico RM1 pela utilização de otimização não-linear

Abstract

Molecular modeling allows us to calculate properties of molecular compounds, being used mainly in the discovery of new drugs or in the improvement of existing prototypes. There are several ways to generate these models, with semiempirical methods being the most computationally efficient alternatives, but with an accuracy that varies a lot, depending on the approach and on the chosen or adjusted parameters. One such method is RM1 (Recife Model 1), created in 2006 as a reparameterization of AM1 (Austin Model 1), a very successful model. RM1 achieved good results, but it is important to assess whether the chosen parameterization was the best possible. In this work, the parameter space for the RM1 method was explored, using a variation of the nonlinear optimization algorithm DFP starting from different points, evaluating whether it is possible to offer a substantial improvement in its accuracy only with a reparameterization. We used, as starting points, parameters obtained from a previous work by the group, which used genetic algorithms, which at the time of their evaluation showed to be slightly better than RM1. The parameterization procedure performed in the present work was able to obtain points that only improved the bond distances, however, in general, it was not able to obtain better points than those obtained when genetic algorithms were used. We attribute this to an imbalance in the cost function used, which probably considered errors in bond distances to be preferred to be reduced to the detriment of other properties. Our conclusion is that, in order to improve the results, a broader prior study of the cost function is necessary, in order to reach a set of adequate weights to obtain a more efficient parameterization.