On defining applicability domains for prediction models in chemoinformatics

AXEL J. SOTO; IGNACIO PONZONI; GUSTAVO E. VAZQUEZ

Quilmes

Congreso; Congreso Argentino de Bioinformática y Biología Computacional; 2010

Asociación Argentina de Bioinformática y Biología Computacional

QSPR (Quantitative Structure-Property Relationships) involve statistical modelling techniques that are used for predicting biological and physicochemical properties of compounds from their molecular structure (descriptors). These models allow the development of virtual screening procedures that are applied on heterogeneous chemical repositories in order to prioritize and identify promising compounds for experimental validation. These in silico procedures allow to disregard compounds before being synthesized, which implies considerably time and economic savings. High prediction accuracy of QSPR models is a crucial issue in the modern drug development process. One of its major difficulties arises from the lacking of generalization capabilities when using non-homogeneous data. In other words, an unseen compound could be out of the model applicability domain (AD), and hence its prediction is prone to be not reliable. Several works [1, 2] describe this problem and the need to be taken into account when developing QSPR models.