Back to AI Research

AI Research

MFG trains molecular optimization to use feedback across several turns

Key Takeaways

  • A reinforcement-learning framework rewards progressive property improvements while retaining similarity to a starting molecule.
  • ## Molecular optimization is an iterative task
  • [Multi-turn Feedback Guided optimization](https://arxiv.org/abs/2610.18851), or MFG, studies modifying a starting molecule to improve several properties while preserving a similar overall structure.
  • The authors focus on how a language model uses feedback from one proposed change to guide the next.
  • Their motivation is that an iterative interface does not automatically make a model effective at iterative optimization.

Molecular optimization is an iterative task

Multi-turn Feedback Guided optimization, or MFG, studies modifying a starting molecule to improve several properties while preserving a similar overall structure. The authors focus on how a language model uses feedback from one proposed change to guide the next.
Their motivation is that an iterative interface does not automatically make a model effective at iterative optimization. Existing approaches can generate a new proposal after feedback without having been explicitly trained to use that feedback over several turns.

Reward progress and preserve the starting structure

MFG supplies a feedback vector that records property improvements and similarity to the source molecule. A feedback-aligned reward function encourages progressive improvement across turns. The method applies reinforcement learning to the GeLLM3O base model for this setting.
The researchers evaluate ten multi-objective optimization tasks. They report an average similarity-constrained success rate of 81.50% for the adapted model, compared with 70.81% for the base model. This metric combines the experiment's success requirements and structural-similarity constraint; it should not be interpreted as a measured rate of successful drug development.
The contribution is explicitly training a model to use an optimization history, rather than judging only a single proposed modification in isolation.

Inspect the trajectory behind the final candidate

Franklin analysis: a useful evaluation should retain every proposed molecule, its computed feedback and the resulting revision. This reveals whether one property improved by sacrificing another or by moving too far from the starting structure.
Compare methods under the same feedback and evaluation budget, and report the properties and similarity threshold used to define success. Computational optimization results identify candidates for further investigation; they do not establish clinical effectiveness or safety. The paper supplies evidence that feedback-aware training can help the reported optimization tasks, with the trajectory providing a practical way to examine how that improvement occurs.

Comments