An automated framework for QSAR model building.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• In-silico quantitative structure-activity relationship (QSAR) models are crucial for screening large compound databases.
• Growing chemical data necessitates computationally efficient, automated QSAR tools for researchers lacking machine learning expertise.
• A fully automated QSAR modeling platform is a valuable addition to the scientific community.

Purpose of the Study:

• To develop a fully automated and customizable QSAR modeling framework.
• To streamline the entire QSAR modeling process from data preparation to validation.
• To provide an accessible tool for researchers regardless of their machine learning expertise.

Main Methods:

• Automated data curation, dataset characteristic evaluation, and variable selection.
• Development of a modelability score to assess data feasibility for modeling.
• Testing the framework on thirty diverse QSAR problems.
• Implementation of an efficient feature selection methodology.

Main Results:

• The automated workflow successfully built reliable QSAR models for thirty different problems.
• Feature selection reduced redundant data by 62-99% and decreased prediction error by 19%.
• Models built with feature selection showed a 49% increase in percentage of variance explained (PVE) compared to those without.
• Models with a modelability score above 0.6 achieved an average PVE of 0.71.

Conclusions:

• A highly customizable, fully automated QSAR modeling framework was developed.
• The workflow requires no advanced parameterization or user expertise in machine learning.
• The framework efficiently develops reliable QSAR models, even for challenging datasets, by incorporating data modelability estimation and effective variable selection.