Crossvalidation, Bootstrapping, and Partial Least Squares Compared with Multiple Regression in Conventional QSAR Studies
RICHARD D. III CRAMER, Jeffrey D. Bunce, David E. Patterson, Ildiko E. Frank · Quantitative Structure-Activity Relationships · 1988
Abstract To better evaluate, in the context of QSAR studies, new validation techniques such as bootstrapping and crossvalidation and the new analytic technique of partial least squares (PLS), seventeen QSAR results taken from nine recent publications were reexamined using these techniques. The results indicate that bootstrapping and crossvalidation are more powerful indicators of possible chance correlation than are the classical tests based on assumed normal independent distribution of variables. Although PLS will not detect all correlations existing within a set of data, its conservative behavior is particularly valuable when the candidate physicochemical descriptors are numerous and non‐orthogonal.