Decision trees versus support vector machine for classification of androgen receptor ligands

With the current concern of limiting experimental assays, increased interest now focuses on in silico models able to predict toxicity of chemicals. Endocrine disruptors cover a large number of environmental and industrial chemicals which may affect the functions of natural hormones in humans and wildlife. Structure-activity models are now increasingly used for predicting the endocrine disruption potential of chemicals. In this study, a large set of about 200 chemicals covering a broad range of structural classes was considered in order to categorize their relative binding affinity (RBA) to the androgen receptor. Classification of chemicals into four activity groups, with respect to their log RBA value, was carried out in a cascade of recursive partitioning trees, with descriptors calculated from CODESSA software and encoding topological, geometrical and quantum chemical properties. The hydrophobicity parameter (log P), Balaban index, and descriptors relying on charge distribution (maximum partial charge, nucleophilic index on oxygen atoms, charged surface area, etc.) appear to play a major role in the chemical partitioning. Separation of strongly active compounds was rather straightforward. Similarly, about 90% of the inactive compounds were identified. More intricate was the separation of active compounds into subsets of moderate and weak binders, the task requiring a more complex tree. A comparison was made with support vector machine yielding similar results.