Beyond the scope of Free-Wilson analysis: building interpretable QSAR models with machine learning algorithms

J Chem Inf Model. 2013 Jun 24;53(6):1324-36. doi: 10.1021/ci4001376. Epub 2013 Jun 12.

Abstract

A novel methodology was developed to build Free-Wilson like local QSAR models by combining R-group signatures and the SVM algorithm. Unlike Free-Wilson analysis this method is able to make predictions for compounds with R-groups not present in a training set. Eleven public data sets were chosen as test cases for comparing the performance of our new method with several other traditional modeling strategies, including Free-Wilson analysis. Our results show that the R-group signature SVM models achieve better prediction accuracy compared with Free-Wilson analysis in general. Moreover, the predictions of R-group signature models are also comparable to the models using ECFP6 fingerprints and signatures for the whole compound. Most importantly, R-group contributions to the SVM model can be obtained by calculating the gradient for R-group signatures. For most of the studied data sets, a significant correlation with that of a corresponding Free-Wilson analysis is shown. These results suggest that the R-group contribution can be used to interpret bioactivity data and highlight that the R-group signature based SVM modeling method is as interpretable as Free-Wilson analysis. Hence the signature SVM model can be a useful modeling tool for any drug discovery project.

MeSH terms

  • Algorithms
  • Artificial Intelligence*
  • Drug Discovery
  • Pharmaceutical Preparations / chemistry
  • Pharmacology
  • Quantitative Structure-Activity Relationship*

Substances

  • Pharmaceutical Preparations