Property‐Conditioned Transformer Language Model for De Novo Molecular Generation Using Self‐Referencing Embedded Strings

Transformer-based molecular language models have demonstrated strong capabilities for de novo molecular generation by modeling chemical structures as token sequences. However, controllable generation of molecules with desired physicochemical properties remains challenging without reinforcement learning (RL) or post-generation filtering. In this study, we propose a property-conditioned autoregressive transformer framework for molecular generation using the SELFIES representation. Continuous molecular descriptors, including logP, quantitative estimate of drug-likeness, and molecular weight (MolWt), were normalized and embedded via a multilayer perceptron and injected into token embeddings to enable global conditioning during sequence modeling. The model was trained end-to-end using the maximum likelihood with cross-entropy loss. Experiments conducted on the combined subsets of QM9, ZINC, and MOSES datasets demonstrated high validity (100%), strong uniqueness, and substantial novelty in the generated molecules. Furthermore, the generated property distributions closely aligned with the training data statistics. The proposed framework provides a computationally efficient and stable approach for controllable molecular generation without requiring RL or graph-based modeling.