Invalid SMILES are beneficial rather than detrimental to chemical language models

Abstract

Generative machine learning models have attracted intense interest for their ability to sample novel molecules with desired chemical or biological properties. Among these, language models trained on SMILES (Simplified Molecular-Input Line-Entry System) representations have been subject to the most extensive experimental validation and have been widely adopted. However, these models have what is perceived to be a major limitation: some fraction of the SMILES strings that they generate are invalid, meaning that they cannot be decoded to a chemical structure. This perceived shortcoming has motivated a remarkably broad spectrum of work designed to mitigate the generation of invalid SMILES or correct them post hoc. Here I provide causal evidence that the ability to produce invalid outputs is not harmful but is instead beneficial to chemical language models. I show that the generation of invalid outputs provides a self-corrective mechanism that filters low-likelihood samples from the language model output. Conversely, enforcing valid outputs produces structural biases in the generated molecules, impairing distribution learning and limiting generalization to unseen chemical space. Together, these results refute the prevailing assumption that invalid SMILES are a shortcoming of chemical language models and reframe them as a feature, not a bug.