Improving Fidelity and Diversity in Chemical Language Transformers for Inverse Molecular Design

Rapid, sustainable redesign of large functional molecules demands efficient exploration of vast chemical spaces. Chemical language models (CLMs), especially transformers, learn long-range structure-property relations and enable swift, batched candidate generation after training. However, inverse molecular design is often ill-posed─many structures can meet a target─and conditioned generation often decodes to invalid or off-spec molecules. To address this challenge, we propose a novel CLM-based inverse design framework that optimizes latent representations toward desired target properties. Our approach introduces a round-trip fidelity metric to quantify and diagnose decoder-induced latent-space drift, which we mitigate via postdecoding re-ranking and predictor-guided minimal-edit repair to correct invalid structures. To demonstrate the framework, we target the surfactant critical micelle concentration (CMC) and compare large pretrained CLMs, our lightweight CLM, a fragment-based genetic algorithm, and a prompt-conditioned ChatGPT baseline. We observe that our framework yields a high proportion of valid and diverse molecules (∼90%) while maintaining a target property error close to 1%. Moreover, interpretability analysis confirms that the designed molecular structures adhere to established physical design rules, highlighting the framework's ability to extract physical insights for molecular design. Therefore, the current framework provides an efficient and broadly applicable solution to the inverse design of novel functional molecules.