What is Wrong with Quantitative Structure–Property Relations Models Based on Three-Dimensional Descriptors?

Quantitative structure–property relations (QSPR)
employing
descriptors derived from the three-dimensional (3D) molecular structure
are frequently applied for property prediction in various fields of
research. However, there is no common understanding of the necessary
degree of detail to which molecular structure has to be known for
reliable descriptor evaluation, but computational methods used vary
from simplified molecular mechanics up to rigorous ab initio programs.
In order to quantify the yet unknown error due to this heterogeneity,
widely used 3D molecular descriptors from diverse fields of application
are evaluated for molecular structures computed by different computational
methods. The results clearly indicate that the widespread, exclusive
use of the most stable molecular conformation as well as too simplistic
computational methods yield systematically erroneous descriptor values
with misleading information for the inferred structure–property
relations. Thus, generating an awareness and understanding of this
fundamental problem is considered an important first step to make
3D QSPR a generally accepted property prediction method.