Matrix Generation and Pretreatments

For the chemometric analysis, the three-dimensional MIFs obtained from GRID are rearranged as one-dimensional vectors. In the GRID/PGA approach [6], one such vector is obtained for each MIF, and the vectors are used to build a two-dimensional X matrix, in which the rows are the probe-target interactions (the objects) and each column contains the variables that describe energetically these interactions at a given grid point. The process used to obtain the X matrix is illustrated in Fig. 3.1. 

A careful pretreatment is necessary to focus on the relevant variables. Often variables with low absolute values ( 0.01 kcal moT ) and those with low standard deviations ( 0.02-0.03kcal mol ) are removed in order to eliminate noise. Autoscaling is not recommended, since all the data comes from the same source (GRID probe-target interaction energies) and all the data are expressed in the same units (kcalmof ). Thus, autoscaling might introduce noise in the model. 

Usually, any positive interaction energy present in the X matrix is then set to 0 kcal mol Thus one focuses on the negative, favorable interaction energies and removes the information about small protein shape differences. Additionally, only using negative interaction energies allows a straightforward interpretation of the results. 

---