Evolutionary design of molecules based on deep learning and a genetic algorithm.

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Evolutionary design accelerates molecular structure modification for targeted properties.
• A key challenge is rapidly evolving molecules while ensuring chemical validity.

Purpose of the Study:

• To develop a novel evolutionary design method that overcomes the chemical validity challenge.
• To leverage deep learning for guiding molecular evolution and property prediction.

Main Methods:

• Utilized deep learning models to extract knowledge from material databases.
• Employed genetic algorithms with mutation and crossover on Morgan fingerprint vectors.
• Used a recurrent neural network to reconstruct valid molecular structures from fingerprints.
• Integrated deep neural networks for efficient molecular property prediction.

Main Results:

• Successfully evolved molecular structures with maintained chemical validity.
• Demonstrated efficient and versatile molecular evaluations through repeated application of the method.
• Performed four design tasks focused on modifying light-absorbing wavelengths of organic molecules.

Conclusions:

• The proposed method effectively combines evolutionary algorithms and deep learning for accelerated molecular design.
• This approach enhances the ability to design molecules with specific properties, such as desired light absorption characteristics.