A RIGOROUS QSAR ANALYSIS?

摘要

The two-dimensional plots may be interpreted as parabolic relationships between biological potency and lipophilicity. A Hansch analysis led to a formally significant QSAR equation. Its significance depends on the presence of a single observation. This is in sharp contrast with the assumptions of sampling theory. Diagnostics statistics methods have shown that there is a multicollinearity without predictive model power. It was also shown that a sample member is influential, a high-leverage point and an extra-carrier point but not an outlier. Using all compounds as training set, NLS and PLS regression did not show significance. Provided that one is interested in continuing QNB research, we have said: make a 'new series design of QNB analogs' and 'new biological tests'. Such design repair was not the aim of the study. If one is interested in examining the role of diagnostics statistics in QSAR research, we concluded: two-dimensional diagrams prove nothing but probably bring understanding features readily to the eye. Diagnostics statistics procedures are suitable for computational automation and, most importantly, probably successful in dealing with highly complex and multivariate data. Subsequent studies of more complex data are needed. Note added in proof: Quite recently, Kubinyi developed the so-called mutation-and-selection uncover-models algorithm for variable selection in QSAR. It starts from a model containing any number of randomly chosen variables. 'Random mutation' (first by addition or elimination of only one or very few random variables which were called x_(noise) in a previous publication, afterwards by simultaneous random addition or elimination at a time), led to a new model which is evaluated by an appropriate fitness function. Whether the concept of random perturbation (see the previous study) as an additional technique of diagnostic statistics is used to investigate inference robustness, or as a genetic-algorithm-like approach for selecting variables in QSAR studies, is meaningless with respect to the common underlying idea. However, it seems that Stone's critical comment to the random-perturbation concept applied to the example under discussion is indeed questionable.