Assessing synthetic accessibility of chemical compounds using machine learning methods.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• De novo drug design generates numerous potential drug candidates.
• Many generated compounds are synthetically infeasible, limiting their practical value.
• Existing synthetic accessibility tools are slow, processing only a few molecules per minute.

Purpose of the Study:

• To develop rapid computational methods for predicting synthetic accessibility of chemical compounds.
• To improve the efficiency of drug discovery pipelines by filtering out synthetically challenging molecules early.

Main Methods:

• Utilized support vector machines (SVM) operating on molecular descriptors.
• Developed two distinct SVM approaches: RSsvm for reaction-specific accessibility and DRsvm for general accessibility.
• Trained RSsvm on compounds assessed via retrosynthetic analysis; trained DRsvm on compounds selected from a diverse library based on similarity.

Main Results:

• RSsvm achieved a 0.952 receiver operator characteristic (ROC) score in cross-validation for reaction-specific synthesis prediction.
• DRsvm achieved a 0.888 ROC score on an independent dataset for general synthesis prediction.
• Both implementations can process thousands of compounds per minute, significantly faster than existing methods.

Conclusions:

• The developed SVM-based approaches provide rapid and accurate predictions of synthetic accessibility.
• These methods can significantly accelerate the drug discovery process by enabling faster evaluation of potential drug candidates.
• The computational tools offer valuable solutions for identifying synthetically feasible molecules in large chemical libraries.