Composite multi-parameter ranking of real and virtual compounds for design of MC4R agonists: Renaissance of the Free-Wilson methodology

Abstract

Drug design is a multi-parameter task present in the analysis of experimental data for synthesized compounds and in the prediction of new compounds with desired properties. This article describes the implementation of a binned scoring and composite ranking scheme for 11 experimental parameters that were identified as key drivers in the MC4R project. The composite ranking scheme was implemented in an AstraZeneca tool for analysis of project data, thereby providing an immediate re-ranking as new experimental data was added. The automated ranking also highlighted compounds overlooked by the project team. The successful implementation of a composite ranking on experimental data led to the development of an equivalent virtual score, which was based on Free-Wilson models of the parameters from the experimental ranking. The individual Free-Wilson models showed good to high predictive power with a correlation coefficient between 0.45 and 0.97 based on the external test set. The virtual ranking adds value to the selection of compounds for synthesis but error propagation must be controlled. The experimental ranking approach adds significant value, is parameter independent and can be tuned and applied to any drug discovery project.

Beyond the Scope of Free-Wilson Analysis: Building Interpretable QSAR Models with Machine Learning Algorithms

“…The basic idea in the Free-Wilson approach is that the biological activity of a molecule can be described as the sum of the activity contributions of its specific substructures (parent core and the corresponding substituents (R-groups)), and the major advantage of this method is its interpretability. Although the Free-Wilson methodology has provided fruitful transparent models − for a range of experimental observations, it is hampered by the limitations in prediction scope. The training set needs to be properly designed to be able to predict a modeled property in an enumerated library for all R-group combinations, and naturally it does not provide explorative prediction for novel R-groups which are not present in the training set.…”

Beyond the Scope of Free-Wilson Analysis. 2: Can Distance Encoded R-Group Fingerprints Provide Interpretable Nonlinear Models?

“…The Free-Wilson approach uses the substituents themselves as descriptors and can therefore only be used to predict within the chemical space defined by the fragments in the training set. Although limited in its prediction scope, the Free-Wilson approach, and the modified Fujita-Ban version , have provided fruitful transparent models − for a range of experimental observations. Several methods have been developed in recent years applying molecular fingerprints for building fragment-based QSAR models. , Molecular fingerprints are representations of chemical structures originally designed to assist in chemical database substructure searching, similarity searching, nearest neighbor analysis, clustering, and classifications.…”