Nonpher: computational method for design of hard-to-synthesize structures

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Machine learning models in cheminformatics require diverse training data, including challenging examples.
• Assessing synthetic feasibility, a key molecular property, is computationally complex.
• Existing methods for generating hard-to-synthesize compounds have limitations.

Purpose of the Study:

• To introduce Nonpher, a novel computational method for generating hard-to-synthesize virtual molecular libraries.
• To optimize molecular morphing for creating compounds with appropriate synthetic difficulty.
• To evaluate Nonpher's effectiveness against established methods like SAscore and dense region (DR).

Main Methods:

• Nonpher utilizes a molecular morphing algorithm for iterative structure generation.
• The algorithm introduces simple structural modifications (atom/bond addition/removal).
• Morphing is optimized to produce molecules that are challenging but not overly complex to synthesize.

Main Results:

• Nonpher successfully generated a virtual library of hard-to-synthesize compounds.
• A random forest classifier trained on Nonpher data outperformed models trained on SAscore and DR data.
• This indicates Nonpher's utility in creating more informative datasets for machine learning.

Conclusions:

• Nonpher provides an effective computational approach for constructing challenging molecular libraries.
• The method enhances the training of machine learning models for tasks like property prediction.
• Nonpher represents a valuable tool for advancing cheminformatics research and drug discovery.