Evolutionary design of molecules based on deep learning and genetic algorithm
- Scientific Reports
The development of high-performance organic functional materials is becoming more important with the advent of new technologies. However, molecular discovery is still a time-consuming step based on costly trial-and-error processes, thus many researches have been conducted to improve the efficiency. Among various approaches, evolutionary design has been gaining a lot of attention as one of the useful tools to accelerate the design process by automatically modifying molecular structures to target directions. But this useful methodology also has a practical challenge: how to rapidly evolve the molecules while maintaining their chemical validity. To address this issue, we develop a deep-learning-based evolutionary design in which the deep-learning models extract the inherent knowledge in the materials database and effectively guide the evolutionary design. In the proposed method, the fingerprint vectors of seed molecules are evolved by mutation and crossover of genetic algorithm, and the evolved fingerprints are reconstructed into actual molecular structures through the recurrent neural network while maintaining the chemical validity. And by predicting their molecular properties using the deep neural network models, more versatile but efficient molecular evaluation is conducted. Three design tasks to change the maximum light-absorbing wavelengths of organic molecules are performed with PubChem library. And the fully data-driven workflow successfully optimizes the molecular structures without incorporating any prior chemical knowledge.
- Sci Rep 11, 17304 (2021)