Deciphering Scientific Reasoning Steps from Outcome Data for Molecule Optimization

Zequn Liu; Kehan Wu; Shufang Xie; Zekun Guo; Wei Zhang; Tao Qin; Renhe Liu; Yingce Xia

q-bio.BMMar 13, 2026

Deciphering Scientific Reasoning Steps from Outcome Data for Molecule Optimization

Authors: Zequn Liu, Kehan Wu, Shufang Xie, Zekun Guo, Wei Zhang, Tao Qin, Renhe Liu, Yingce Xia

Organizations: Zhongguancun Academy · School of Computer Science and Technology, University of Science and Technology of China · Global Health Drug Discovery Institute

Abstract

Emerging reasoning models hold promise for automating scientific discovery. However, their training is hindered by a critical supervision gap: experimental outcomes are abundant, whereas intermediate reasoning steps are rarely documented at scale. To bridge this gap, we propose DESRO, a framework for deciphering scientific reasoning from outcomes. By analyzing shared patterns and key differences within grouped data, a large language model (LLM) can recover the underlying logic. We instantiate this framework in molecule optimization, a pivotal stage in drug discovery that traditionally relies on the iterative reasoning of medicinal chemists. Across 2.3 million molecular property records, our framework infers optimization rationales by grouping molecules with shared fragments, then using an LLM to analyze how structural variations correlate with property differences. Based on the derived data, we train a model that conducts molecule optimization through an interpretable reasoning process. DESRO achieves the highest success rates on 15 out of 18 tasks, spanning both single- and multi-property optimization of bioactivity and ADMET properties. The reasoning process enables robust generalization to out-of-distribution scenarios, including novel property combinations, unseen biological targets, and unseen properties defined solely by natural language descriptions. In retrospective case studies under strict temporal splits, the model autonomously reconstructs expert-level lead optimization trajectories. Additionally, our framework extends beyond molecule optimization to reaction ligand selection. Our results establish deciphering reasoning steps from outcome data as a viable paradigm for enabling scientific reasoning, providing a scalable approach to accelerate scientific discovery.

Figures & tables

Figure 1: Figure 1. Overview of DESRO. a, The pipeline of DESRO for molecule optimization. Large-scale molecular property datasets are processed using fragment mining to identify common fragments. o1-mini is then employed to induce reasoning steps, including property analyses and optimization rules, from these molecule groups, resulted in a reasoning-annotated training dataset. b, Construction of multi-property optimization examples. Molecule pairs that appear across multiple property domains are identified, and their corresponding instructions, analyses, rules, and modifications are aggregated to enable joint multi-property optimization. c, The multi-task training framework. A LLaMA-3-8B model is fine-tuned using a multi-task approach, where it learns to predict the remaining reasoning components and modification steps based on varied input configurations. d, The inference workflow of the reasoning model. Given an initial molecule and a natural language optimization instruction, the model first generates an explicit reasoning trajectory, followed by the corresponding optimized molecular structure.

Figure 2: Figure 2. Evaluation of molecule optimization on ADMET properties and bioactivity. a,b, Bar plots comparing DESRO and competing methods in single-property optimization tasks for ADMET properties ( a ) and bioactivity targets ( b ) in terms of Success Rate. c, Case study for bioactivity optimization for the DRD2 target. The original molecule is modified based on the inferred rule regarding electron-withdrawing groups to improve binding affinity. d, Bar plots comparing the success rates of DESRO and baselines across multi-property optimization tasks. e, t-SNE plot visualizing the chemical space of original molecules (stars) and generated candidates (circles) for the BPQ (BBBP, pLogP and QED) task. Points are colored by the average score of the target properties, where lighter colors indicate a better average performance across all objectives. f, Scatter plot showing the property distribution of generated molecules for a selected starting scaffold in the BPQ task in terms of QED and BBBP, colored by pLogP values. The dashed box details the sub-rules applied by the model to satisfy multiple property constraints simultaneously.

Figure 3: Figure 3. Evaluation of generalization to out-of-domain (OOD) molecule optimization tasks. a, Bar plots comparing the success rates of DESRO and baselines on OOD multi-property optimization tasks. b, Box plots showing the docking score distribution for molecules optimized for ClpP2, a protein target not encountered during training. The green dashed line indicates the docking score of the input molecule; lower scores indicate better binding affinity. c, Case study of OOD multi-property optimization. The model successfully optimizes a molecule on the OOD multi-property optimization by applying learned sub-rules for single properties. d, The inference workflow for optimizing entirely unseen properties. The model leverages expert-provided property analyses to generate optimization rules and modifications without prior training on those specific properties. e, Bar plots comparing the success rates on three unseen properties: Oral Bioavailability, Clinical Toxicity (ClinTox), and drug sensitivity on the SBC-5 cell line.

Figure 4: Figure 4. Evaluation of lead optimization in real-world drug discovery scenarios. a, Case study on the optimization of MALT1 inhibitors. DESRO successfully guides optimization from an initial lead to the clinical candidate SGR-1505. b,d, Histograms illustrating the physicochemical property distributions of the molecular libraries generated by DESRO for the MALT1 ( b ) and CBL-B ( d ) optimization tasks. The distributions cover LogP, QED, and Synthetic Accessibility (SA) Score. The vertical dashed lines represent the property values of the input molecules. c, Case study on the optimization of CBL-B inhibitors. The model navigates the chemical space from a hit compound to the lead molecule ”Compound 10” and subsequently identifies a structurally distinct analog with superior potency.

Figure 5: Figure 5. Extension to reaction ligand selection. a, Ligand selection for a heteroaryl chloride (6-chloroquinoline), where DESRO infers (based on controlled comparisons within the dataset) that oxidative addition is a plausible barrier and prioritizes a bulky, electron-rich monodentate ligand (SPhos). b, Ligand selection for a heteroaryl triflate (6-triflatequinoline) in THF, where DESRO emphasizes catalyst speciation/stability considerations and selects a chelating bidentate ligand (dppf). These examples illustrate condition-dependent reasoning for ligand choice.

Category	Property	Predictor / Scorer	Reliability
In-Domain Properties
Bioactivity	DRD2	tdc.Oracle [ 93 ]	AUROC= $0.99$
	JNK3		AUC = 0.983
	GSK3 $\beta$		AUC = 0.984
ADMET	BBBP	ADMET-AI Platform [ 95 ]	AUROC = $0.869\pm 0.027$
	HIA		AUROC= $0.981\pm 0.002$

Table 1: Summary of property predictors and their reported reliability.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S2: Data distribution of the 197k+ molecule optimization reasoning processes. The pie chart categorizes the collected optimization paths by their target properties, spanning various single-property tasks (e.g., Bioactivity, QED, ADMET) and multi-property combinations.

Figure S4: Bar plot comparing DESRO and Prompt-MolOpt on Mutagenicity and hERG in terms of Success Rate.

Figure S5: Bar plot comparing DESRO and baselines on BDQ (increase BBBP, DRD2 and QED) in terms of Success Rate.

Figure S6: a - c . Scatter plots illustrating the property distribution of generated molecules for three selected starting scaffolds in the BPQ task in terms of QED and BBBP, colored by pLogP values.

Figure S7: Bar plot comparing the success rates of DESRO and baselines on the OOD multi-property optimization tasks, BDMQ (increase BBBP, DRD2, QED, and decrease mutagenicity).

Figure S8: a,b, Histograms illustrating the molecular weight distributions of the molecular libraries generated by DESRO for the MALT1 ( a ) and CBL-B ( b ) optimization tasks. The vertical dashed lines represent the property values of the input molecules.

Property	Data source	$\delta$
Bioactivity	PubChem	$y_{i}^{\prime}$
Bioactivity (DRD2)	GeLLMO	0.2
Bioactivity (JNK3)	Synthetic	0.1
Bioactivity (GSK3 $\beta$ )	Synthetic	0.1
BBBP	Prompt-MolOpt	$y_{i}^{\prime}>0.5$ and $y_{i}<=0.5$
BBBP	GeLLMO	0.2

Appendix

Table S1: Property thresholds for training properties.

Template
Bioactivity
Design a molecule by optimizing {mol1} to enhance its binding affinity with {target_name}.
Given {target_name}, suggest modifications to {mol1} to create a new molecule.
How can {mol1} be modified to better interact with {target_name}?
To enhance the binding of {mol1} with {target_name}, how can it be optimized?
What changes can be made to {mol1} to improve its affinity for {target_name}?

Appendix

Table S2: Instruction templates for bioactivity optimization

Template
BBBP
Enhance the BBB penetration capability of {mol1}.
Boost the ability of {mol1} to cross the BBB.
Enhance the ability of {mol1} to cross the blood-brain barrier.
Water solubility
Boost the solubility of {mol1} in water.

Appendix

Table S3: Instruction templates for ADMET properties and multi-property combination

Scientific Reasoning in Language Models

Scientific reasoning in language models studies how AI systems use evidence, domain knowledge, and tools to solve problems and support discovery.

See all 144 papers

MolEvolve: LLM-Guided Evolutionary Search for Interpretable Molecular Optimization
Mar 25, 2026Xiangsen Chen, Ruilong Wu, Yanyan Lan +2Molecular Optimization Large Language Model-Guided Optimization
Hong Kong Polytechnic University · Hong Kong University of Science and Technology(Guangzhou) · Tsinghua University +1
Logos: An evolvable reasoning engine for rational molecular design
Mar 10, 2026Haibin Wen, Zhe Zhao, Fanfu Wang +4Molecule Design AI for Science
City University of Hong Kong, Hong Kong 999077, China · Lanzhou University, Lanzhou 730000, China · Riltide Medicines, Beijing, China +1
A Multitask Large Reasoning Model for Molecular Science
Mar 13, 2026Pengfei Liu, Shuang Ge, Xiaobo Wang +6Scientific Reasoning Multi-Task Learning
Guangzhou Key Laboratory of Forensic Multi-Omics for Precision Identification, Southern Medical University, Guangzhou, Guangdong, 510515, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, Guangdong 518000, China · Guangdong Provincial Key Laboratory of Psychoactive Substances Monitoring and Safety, Anti-Drug Technology Center of Guangdong Province (National Anti-Drug Laboratory Guangdong Regional Center), Guangzhou, 510230, China +4

Mar 25, 2026J/K move · Enter open · S save

MolEvolve: LLM-Guided Evolutionary Search for Interpretable Molecular Optimization

Xiangsen Chen, Ruilong Wu, Yanyan Lan, Ting Ma+1

Hong Kong Polytechnic University · Hong Kong University of Science and Technology(Guangzhou) · Tsinghua University+1

Molecular Optimization Large Language Model-Guided Optimization Molecular Property Prediction

Open paper AskPDF arXiv

Despite deep learning's success in chemistry, its impact is hindered by a lack of interpretability and an inability to resolve activity cliffs, where minor structural nuances trigger drastic property shifts. Current representation learning, bound by the similarity principle, often fails to capture these structural-activity discontinuities. To address this, we introduce MolEvolve, an evolutionary framework that reformulates molecular discovery as an autonomous, look-ahead planning problem. Unlike traditional methods that depend on human-engineered features or rigid prior knowledge, MolEvolve leverages a Large Language Model (LLM) to actively explore and evolve a library of executable chemical symbolic operations. By utilizing the LLM to cold start and an Monte Carlo Tree Search (MCTS) engine for test-time planning with external tools (e.g. RDKit), the system self-discovers optimal trajectories autonomously. This process evolves transparent reasoning chains that translate complex structural transformations into actionable, human-readable chemical insights. Experimental results demonstrate that MolEvolve's autonomous search not only evolves transparent, human-readable chemical insights, but also outperforms baselines in both property prediction and molecule optimization tasks.

Abstract

Scientific Reasoning in Language Models

Explore similar work

MolEvolve: LLM-Guided Evolutionary Search for Interpretable Molecular Optimization

Logos: An evolvable reasoning engine for rational molecular design

A Multitask Large Reasoning Model for Molecular Science