Effect of parameter variations on the effectiveness of HQSAR analyses

HQSAR is a new method for generating alignment-free quantitative structure- activity relationships. Experiments with four different datasets suggest th at the variations in PLS scores that are observed with short hologram lengt hs can be substantially removed either by taking the mean or median of the crossvalidation scores or by using very long holograms. However, because th e hashing process unnecessarily obfuscates PLS regression modelling, we sug gest that PLS should preferentially be applied to unhashed fragment bit-str ings where computationally feasible. The predictive ability of the method i s also affected by the size of the fragments that are used, although this e ffect appears to be dataset-dependent. Variations in the type, rather than the size, of the fragments utilised can also have a significant effect on b oth internal and external predictivity scores.