LIBRARY meta-learning technique was originally suggested by Tetk= o (Tetko IV et al. Large-scale evaluation of log P predictors: local co= rrections may compensate insufficient accuracy and need of experimentally t= esting every other compound. Chem. Biodivers. 2009 Nov;6(11):1837-1844.). The following excerpt was taken from a doctoral work by Sushko (Sushko = I., Applicability Domain of QSAR models. Doctroral work. 2011. http://mediatum.ub.tum.de/node?id=3D1004002)=
=20Sushko I, Applicability Domain of QSAR models. Excerpt=20 =20
Given a predictive model and an additional set of new experimental measu= rements (referred to as =E2=80=9Clibrary=E2=80=9D), it is possible to corre= ct the model by taking into account these measurements. The process is call= ed LIBRARY model correction (because we complement a model with a =E2=80=9C= library=E2=80=9D of experimental measurements) and is performed as follows.=
To obtain a corrected prediction for a molecule J, we calculate the orig=
inal (noncorrected) prediction given by the original model and find K molecu=
les from the =E2=80=9Clibrary=E2=80=9D {Ji, i =3D 1..K} nearest to the mole=
cule being predicted. The nearest compounds are defined by the correlation =
coefficient in space of model predictions.
Then, we calculate the expected residual for the molecule J as the avera= ge residual for the K nearest compounds according to the following expressi= on:
Finally, we correct the original prediction by subtracting the average r= esidual:
Thus, the LIBRARY correction assumes that for a new compound a model wil= l behave similarly to compounds from the =E2=80=9Clibrary=E2=80=9D. Such te= chnique is especially useful in case if retraining of the original model is= infeasible due to high computational complexity or unavailability of the o= riginal training set. The LIBRARY technique was introduced by Tetko and was= shown to significantly increase the prediction accuracy for the lipophilic= ity and distribution coefficient models.