How wrong can we get? A review of machine learning approaches and error bars

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning applications in drug discovery

Background:

• Ligand-based virtual screening is crucial for identifying potential drug candidates.
• Numerous machine learning methods exist, but their application in virtual screening requires careful consideration.

Purpose of the Study:

• To introduce and compare three nonlinear machine learning methods for ligand-based virtual screening: support vector regression, Gaussian process models, and decision trees.
• To explain how confidence estimates (error bars) can be derived from these models.
• To discuss essential aspects of model building, evaluation, and practical application.

Main Methods:

• Focus on three nonlinear machine learning techniques: support vector regression, Gaussian process models, and decision trees.
• Detailed explanations of each method's principles and intuitive understanding.
• Methodologies for model selection, performance evaluation, and verification of error bar quality.

Main Results:

• Provides an introduction to support vector regression, Gaussian process models, and decision trees for virtual screening.
• Explains the generation and verification of confidence estimates (error bars) for these models.
• Discusses practical considerations for implementing these machine learning methods.

Conclusions:

• The study offers a practical guide to applying specific nonlinear machine learning methods in ligand-based virtual screening.
• Highlights the importance of confidence estimates for robust model evaluation.
• Aims to facilitate the effective use of these computational tools in drug discovery.