De novo molecule design towards biased properties via a deep generative framework and iterative transfer learning

Abstract

De novo design of molecules with targeted properties represents a new frontier in molecule development. Despite enormous progress, two main challenges remain: (i) generating novel molecules conditioned on targeted, continuous property values; (ii) obtaining molecules with property values beyond the range in the training data. To tackle these challenges, we propose a reinforced regressional and conditional generative adversarial network (RRCGAN) to generate chemically valid molecules with targeted HOMO–LUMO energy gap (ΔE_H–L) as a proof-of-concept study. As validated by density functional theory (DFT) calculation, 75% of the generated molecules have a relative error (RE) of <20% of the targeted ΔE_H–L values. To bias the generation toward the ΔE_H–L values beyond the range of the original training molecules, transfer learning was applied to iteratively retrain the RRCGAN model. After just two iterations, the mean ΔE_H–L of the generated molecules increases to 8.7 eV from the mean value of 5.9 eV shown in the initial training dataset. Qualitative and quantitative analyses reveal that the model has successfully captured the underlying structure–property relationship, which agrees well with the established physical and chemical rules. These results present a trustworthy, purely data-driven methodology for the highly efficient generation of novel molecules with different targeted properties.