Multi-objective molecular optimization is a critical process in drug discovery that aims to modify a given molecule to simultaneously improve multiple properties while maintaining a similar overall structure. This process often requires iterative modifications guided by computational or experimental feedback. Recent large language model LLM-based methods have begun to explore this iterative approach; however, these methods are not explicitly trained to leverage such feedback, while existing approaches primarily focus on single-turn molecular optimization. To address this gap, the authors introduce Multi-turn Feedback Guided MFG molecular optimization, a reinforcement learning framework that trains an LLM to progressively optimize molecules using feedback from previous turns.
MFG introduces a feedback vector capturing both property improvements and source molecule similarity, alongside a feedback-aligned reward function that promotes progressive improvement across turns. The framework is applied over GeLLM³O, a state-of-the-art LLM model for multi-objective molecular optimization. The authors evaluate GeLLM³O-MFG on ten multi-objective molecular optimization tasks. The results demonstrate that GeLLM³O-MFG achieves the highest average similarity-constrained success rate 81.50%, outperforming the base model GeLLM³O 70.81%.
Unlike prior LLM-based methods that are not explicitly trained to leverage feedback, MFG trains the model to progressively improve molecules across multiple turns. The feedback vector explicitly captures two signals: how much each property improved relative to the source molecule, and how similar the modified molecule remains to the original. This dual signal prevents the model from making excessive changes that sacrifice structural integrity while still achieving property gains. The feedback-aligned reward function shapes the training objective to value progressive improvement, encouraging the model to make sensible intermediate steps rather than attempting large, risky modifications in a single turn.
Across the ten tasks, GeLLM³O-MFG consistently surpasses the base model. The 81.50% average similarity-constrained success rate represents a substantial improvement over 70.81%, indicating that the feedback-guided training effectively balances property optimization with molecular similarity. The paper presents results across diverse optimization targets commonly encountered in drug discovery, showing that the approach generalizes well beyond a single property or molecule type.
The authors note that MFG’s training protocol can be integrated with existing LLM fine-tuning pipelines. The feedback vector and reward function are modular components that can be adapted to different molecular representation schemes and property sets. By explicitly modeling the trade-off between improvement and similarity, MFG provides a more principled approach to iterative molecular optimization than methods that rely on ad-hoc feedback handling or single-turn optimization pressures.