Inverse molecule design represents a next frontier in drug and materials development due to its ability to readily design novel molecules with chosen properties. Despite enormous progress, inverse molecule design still faces two main challenges, i.e., reaching high accuracy between targeted and calculated properties of generated molecules and extrapolation beyond the training samples. To overcome these gaps, we propose a novel reinforced regressional and conditional generative adversarial network (RRCGAN) to design molecules with desired heat capacity. A reinforcement center was introduced to reward chemically valid molecules whose properties are close to the targeted ones. To bias RRCGAN model toward higher values, a transferred model was finetuned by high-value samples. To increase the accuracy of the transferred model and extrapolate outside the training range, active learning framework was proposed to choose molecules with high heat capacity and high uncertainty to be evaluated and fed back to the transferred model. RRCGAN was successfully demonstrated to generate molecules with 13% higher heat capacity compared to the maximum value of training molecules. As validated by DFT, around 90% of samples have relative error within 20% of targeted heat capacity. RRCGAN is a general framework and can be directly expended to other properties and modified for multi-objective properties. The on-demand designed molecules having desired physicochemical properties can be screened by chemists and synthesized by autonomous robots.

Inverse generation of molecules towards biased properties via a deep generative model and transfer learning

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.