A property–agnostic framework for scalable molecular inverse design via quantum annealing

Abstract

Technologies for designing molecules with deTechnologies for designing molecules with desired properties have the potential to drive innovation across a wide range of fields. Molecular inverse design typically involves three key tasks: chemical latent space representation, property prediction, and molecule generation. While deep learning models trained on large molecule-property datasets can address all three tasks within a unified framework, they often require substantial property-specific retraining when targeting new molecular properties, limiting scalability. In this work, we propose an algorithmic framework that integrates machine learning and quantum annealing by explicitly decoupling chemical latent space representation, property prediction, and molecule generation. By freezing deep molecular representations learned from large datasets and separating molecule generation from deep learning, the proposed method enables inverse design for new target properties at a low training cost, requiring only lightweight model adaptation. Using quantum annealing, 98% of the generated molecules were novel and exhibited properties close to the desired targets, indicating efficient exploration beyond the training distribution. Moreover, the molecular generation rate was approximately six times faster than that of classical optimization algorithms. These results demonstrate that modularizing molecular inverse design into complementary learning and optimization components provides a scalable and effective alternative to monolithic deep learning-based approaches.