Time-series behavior and autoQSAR

It is to be expected that the predictive performance of global models will detoriate over time as projects move on and produce new chemotypes not used in the generation of the model. Literature on time-dependent QSAR model behavior is stiU sparse. In a recent investigation this effect was shown for a human plasma protein binding (hPPB) model.A solution to this behavior is the use of correction libraries (see Section m.B.4.) and/or to update (rebuild) the model regularly. An automated approach to this prob-lan has been published and it is expected that this technology (autoQSAR) will become more widespread in the future.  

CHAPTER 23 Quantitative Approaches to Structure-Activity Relationships 

Predicted Measured Positive (good) Negative (bad) Total 

Sensitivity (= recall) a/(a + c) Fraction of positives correctly predicted positive true positive rate 

Specificity df b + d) Fraction of negatives correctly predicted negative true negative rate 

---