In Silico Prediction of Volume of Distribution in Humans. Extensive Data Set and the Exploration of Linear and Nonlinear Methods Coupled with Molecular Interaction Fields Descriptors

We present three in silico volume
of distribution at steady state
(VDss) models generated on a training set comprising 1096 compounds,
which goes well beyond the conventional drug space delineated by the
Rule of 5 or similar approaches. We have performed a careful selection
of descriptors and kept a homogeneous Molecular Interaction Field-based
descriptor set and linear (Partial Least Squares, PLS) and nonlinear
(Random Forest, RF) models. We have tested the models, which we deem
orthogonal in nature due to different descriptors and statistical
approaches, with good results. In particular we tested the RF model,
via a leave-class-out approach and by using a set of 34 additional
compounds not used for training. We report comparable results against
in vivo scaling approaches with geometric mean-fold error at or below
2 (for a set of 60 compounds with animal data available) and discuss
the predictive performance based on the ionization states of the compounds.
Lastly, we report the findings using a two-tier approach (classification
followed by regression) based on VDss ranges, in an attempt to improve
the prediction of compounds with very high VDss. We would recommend,
overall, the RF model, with 33 descriptors, as the primary choice
for VDss prediction in humans.