On Identifying Likely Determinants of Biological Activity in High Dimensional QSAR Problems

A new approach to identifying descriptors associated with biological activity in ‘asymmetric’ data sets has been developed for high dimensional QSAR problems. As an example we have applied this approach to a problem first described by Selwood et al. [3]. Using Cluster Significance Analysis (CSA) in an explicit predetermined sequence of steps, we identified six likely determinants of antifilarial activity among antimycin analogs characterized by 53 descriptors. These determinants were ATCH5, ATCH2, VDWVOL, S8_1DX, ATCH4 and DIPV_X. Of them, only ATCH5, ATCH4 and DIPV_X were significant (R = 0.83) in a stepwise multiple regression analysis (MRA). These results are compared with those of the original workers and others who analyzed the data by different methods. By combining these descriptors with those identified by the other groups, we discovered a still better correlation. Jointly, ATCH5, ATCH4, DIPV_X, LOGP and MOFI_Y correlate highly with antifilarial potency (R = 0.91). The problem of discovering the ‘true’ determinants of biological activity in high dimensional asymmetric data sets is discussed.