Generation of Tailorable Multi-Property Molecules via Multimodal Spectral Fusion Evolution

Molecular generation stands at the forefront of intelligent molecular design, promising transformative advances in the discovery of molecules meeting druglikeness or catalysis requirements. However, existing generative models typically focus on optimizing a single target property at a time, hindered by the intrinsic difficulty of simultaneously encoding and evolving multiple structural features. Herein, we developed MUltimodal Spectral fusion Evolution (MUSE)—a spectroscopy-guided generative framework that unifies multimodal spectroscopic descriptors as a universal chemical language for multi-property molecular creation. A Spectra-to-Molecules (S2M) model, as one of MUSE's core modules, achieved high success rate in molecular structure generation by training on a self-built density functional theory (DFT) levelmultimodal spectral database. Leveraging the continuously variable nature of spectral descriptors, MUSE achieved synergistic integration of structural, vibrational, and electronic information by combining complementary spectroscopic modalities. Through evolutionary algorithms that constrain the generation direction of multiproperty molecules, the framework enables an up to 700-fold increase in the enrichment of rare quadruple-property candidates. This work provides a scalable pathway for multiobjective discovery in chemistry and materials science.