Steering Vector Fields for Property-Controlled Molecular Generation with Chemical Language Models

Abstract

Chemical language models have recently become a powerful tool for the de novo generation of drug-like molecules represented as SMILES strings. A central challenge is steering generation toward compounds with favorable properties such as solubility and absorption. To this end, we investigate inference time control of generative chemical language models using activation steering. Using contrastive activation addition, we seek to improve three relevant properties: molecular size, aqueous solubility (log S), and lipophilicity (log P) without changing the model weights. We compare two interventions: a single global vector which is added to the activation in the last transformer layer, and a novel vector field where the addition vector is computed as a function of the current hidden state. Across multiple protein targets and two pre-trained models, the global steering vector yields desired results in just over half of our experiments, while the vector field achieves larger shifts at the expense of a decrease in the validity rate.