MFG: Multi-turn Feedback Guided Molecular Optimization

Trieu Nguyen; Frazier N. Baker; Ruoxi Gao; Daniel Adu-Ampratwum; Xia Ning

cs.AIOct 13, 2026

MFG: Multi-turn Feedback Guided Molecular Optimization

Authors: Trieu Nguyen, Frazier N. Baker, Ruoxi Gao, Daniel Adu-Ampratwum, Xia Ning

Organizations: Department of Computer Science and Engineering, The Ohio State University · Division of Medicinal Chemistry and Pharmacognosy, The Ohio State University · Department of Biomedical Informatics, The Ohio State University · Translational Data Analytics Institute, The Ohio State University

Abstract

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, we 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. We apply MFG over $\text{GeLLM}^{3}\text{O}$ , a state-of-the-art LLM model for multi-objective molecular optimization, evaluate $\text{GeLLM}^{3}\text{O-MFG}$ on ten multi-objective molecular optimization tasks. The results demonstrate that $\text{GeLLM}^{3}\text{O-MFG}$ achieves the highest average similarity-constrained success rate ( $81.50%$ ), outperforming the base model $\text{GeLLM}^{3}\text{O}$ ( $70.81%$ ).

Figures & tables

Figure 1: Overall figure of Multi-turn Feedback Guided (MFG) $\mathop{\text{MOMO}}\limits$ .

General-purpose LLMs
Model	Setting	BDP			BDQ			BPQ			DPQ			BDPQ
Model	Setting	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$
Qwen3.5-9B	ST	32.80	0.76	0.98	18.80	0.73	0.54	90.40	0.75	0.58	15.60	0.72	0.51	11.00	0.75	0.54
Qwen3.5-9B	MT	36.40	0.72	0.95	21.60	0.70	0.60	90.40	0.75	0.58	19.80	0.69	0.36	12.20	0.76	0.51
Ministral-3-8B	ST	27.60	0.71	0.74	24.00	0.67	0.45	64.80	0.68	0.52	12.80	0.68	0.28	8.60	0.72	0.56
Ministral-3-8B	MT	37.60	0.68	0.70	24.60	0.66	0.69	66.00	0.65	0.65	13.60	0.68	0.36	11.60	0.70	0.47

Table 1: Overall Performance on IND Tasks

General-purpose LLMs
Model	Setting	MPQ			BDMQ			BHMQ			BMPQ			HMPQ
Model	Setting	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$
Qwen3.5-9B	ST	75.80	0.69	0.41	17.00	0.76	0.41	72.03	0.73	1.03	75.39	0.71	0.47	81.25	0.73	0.54
Qwen3.5-9B	MT	75.80	0.69	0.41	23.40	0.71	0.43	77.12	0.69	1.03	75.39	0.71	0.47	81.25	0.69	0.68
Ministral-3-8B	ST	47.80	0.65	0.37	15.40	0.66	0.52	51.69	0.66	1.07	47.12	0.65	0.47	34.38	0.67	0.59
Ministral-3-8B	MT	49.60	0.64	0.39	23.20	0.64	0.76	57.63	0.63	1.03	56.54	0.65	0.49	40.62	0.63	0.69

Table 2: Overall Performance on OOD Tasks

Model	Setting	BDP			BDQ			BPQ			DPQ			BDPQ
Model	Setting	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$
GeLLM 3 O-RL	ST	79.00	0.61	3.67	70.60	0.63	4.94	90.60	0.60	1.31	47.20	0.62	2.00	41.80	0.62	2.28
GeLLM 3 O-RL	MT	79.40	0.58	4.18	80.60	0.59	5.67	90.60	0.60	1.31	47.20	0.62	2.00	49.20	0.59	3.14
GeLLM 3 O-MFG	ST	83.20	0.62	3.62	73.00	0.63	4.30	89.20	0.62	1.25	48.40	0.63	1.97	44.00	0.61	2.50
GeLLM 3 O-MFG	MT	84.20	0.61	4.24	86.20	0.61	5.32	91.20	0.60	1.32	54.60	0.60	2.49	53.40	0.59	3.19
		$\texttt{SR}_{0.6}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.6}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.6}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.6}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.6}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$

Table 3: Ablation Study

Figure 2: An optimization case on BDPQ with $T=2$ . (a) The source molecule. (b) The best optimized molecule at turn $t=1$ . (c) The best optimized molecule at turn $t=2$ . The modified fragments are highlighted.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

General-purpose LLMs
Model	$T$	BDP			BDQ			BPQ			DPQ			BDPQ
Model	$T$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$
Qwen3.5-9B	1	32.80	0.76	0.98	18.80	0.73	0.54	90.40	0.75	0.58	15.60	0.72	0.51	11.00	0.75	0.54
	2	36.40	0.72	0.95	21.60	0.70	0.60	88.60	0.71	0.65	19.80	0.69	0.36	12.20	0.76	0.51
	3	35.40	0.71	0.97	19.80	0.70	0.44	81.80	0.71	0.61	16.60	0.70	0.35	11.60	0.72	0.51
	4	34.60	0.70	1.04	19.80	0.71	0.51	75.80	0.71	0.62	15.40	0.69	0.33	11.80	0.71	0.60

Appendix

Table A.2 : Overall Performance on IND Tasks Across Turn Settings

General-purpose LLMs
Model	$T$	MPQ			BDMQ			BHMQ			BMPQ			HMPQ
Model	$T$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$	$\texttt{SR}_{0.5}^{\uparrow}$	$\texttt{Sim}^{\uparrow}$	$\texttt{RI}^{\uparrow}$
Qwen3.5-9B	1	75.80	0.69	0.41	17.00	0.76	0.41	72.03	0.73	1.03	75.39	0.71	0.47	81.25	0.73	0.54
	2	68.00	0.66	0.46	23.40	0.71	0.43	77.12	0.69	1.03	73.82	0.67	0.54	81.25	0.69	0.68
	3	63.20	0.66	0.43	21.20	0.69	0.41	73.73	0.69	1.03	62.83	0.66	0.55	72.92	0.68	0.73
	4	60.20	0.66	0.45	21.40	0.69	0.42	70.34	0.70	0.98	59.16	0.67	0.52	70.83	0.67	0.74

Appendix

Table A.3 : Overall Performance on OOD Tasks Across Turn Settings

Artifact	Source	License Type	Accessibility
MuMOInstruct	https://huggingface.co/datasets/NingLab/MuMOInstruct	CC BY 4.0	Dataset
Llama-3.1-8B-Instruct	https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct	Llama 3.1 Community License	Checkpoint
GeLLM 3 O-P(6) Llama	https://huggingface.co/NingLab/GeLLMO-P6-Llama	CC BY-NC 4.0	Checkpoint
Qwen3.5-9B	https://huggingface.co/Qwen/Qwen3.5-9B	Apache 2.0	Checkpoint
Ministral-3-8B-Instruct-2512-BF16	https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-BF16	Apache 2.0	Checkpoint
LlaSMol-Mistral-7B	https://huggingface.co/osunlp/LlaSMol-Mistral-7B	CC BY 4.0	LoRA Adapter

Appendix

Table A.4 : Licenses and sources of artifacts used in our study.

Molecular Optimization

Molecular optimization computationally designs or edits chemical structures to improve desired properties while respecting chemical and practical constraints.

See all 103 papers

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs
Apr 24, 2026Rui Gao, Youngseung Jeon, Swastik Roy +2RL for Language Models Molecular Optimization
University of California, Los Angeles · Amazon
Small Molecule Optimization with Large Language Models
Jul 26, 2024Philipp Guevorguian, Menua Bedrosian, Tigran Fahradyan +3Molecular Optimization Evolutionary Optimization
YerevaNN research lab, Yerevan, Armenia.
M3OS: A Monte Carlo Graph Search-Orchestrated Multi-Agent LLM System for Evidence-Traced Molecular Optimization
Sep 28, 2026Junjie Wang, Yaowei Jin, Ruohui Tang +7Molecular Optimization Multi-Agent LLM Systems
Lingang Laboratory, Shanghai, 200031, China. · School of Information Science and Technology, ShanghaiTech University, Shanghai, 201210, China. · Global Institute of Future Technology, Shanghai Jiao Tong University, Shanghai, 200240, China.

Apr 24, 2026J/K move · Enter open · S save

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs

Rui Gao, Youngseung Jeon, Swastik Roy, Morteza Ziyadi+1

University of California, Los Angeles · Amazon

RL for Language Models Molecular Optimization Large Language Model-Guided Optimization Multi-Objective Language Model Alignment Drug Discovery

Open paper AskPDF arXiv

Large language models (LLMs) show promise for molecular optimization, but aligning them with selective and competing drug-design constraints remains challenging. We propose C-Moral, a reinforcement learning post-training framework for controllable multi-objective molecular optimization. C-Moral combines group-based relative optimization, property score alignment for heterogeneous objectives, and bottleneck-sensitive non-linear reward aggregation to improve stability across competing molecular properties. Experiments on C-MuMOInstruct and S

^2

-Bench MolOpt show that C-Moral achieves the best performance among compared methods on both benchmarks. On C-MuMOInstruct, C-Moral achieves the best Success Optimized Rate (SOR) of 48.9% on in-domain tasks and 39.5% on out-of-domain tasks while preserving scaffold similarity. On S

^2

-Bench MolOpt, it also achieves the strongest results across LogP, MR, and QED optimization tasks. These results suggest that C-Moral is an effective way to align molecular LLMs with continuous and constrained molecular design objectives. Our code and models are publicly available at https://github.com/Rwigie/C-MORAL.

Abstract

Molecular Optimization

Explore similar work

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs

Small Molecule Optimization with Large Language Models

M3OS: A Monte Carlo Graph Search-Orchestrated Multi-Agent LLM System for Evidence-Traced Molecular Optimization