RetroGEF: Dynamic Graph Edit Flow for Single-Step Retrosynthesis

Xiaozhuang Song; Xuemin Chen; Xinjian Zhao; Yaoyao Xu; Tianshu Yu

cs.LGSep 29, 2026

RetroGEF: Dynamic Graph Edit Flow for Single-Step Retrosynthesis

Authors: Xiaozhuang Song, Xuemin Chen, Xinjian Zhao, Yaoyao Xu, Tianshu Yu

Organizations: The Chinese University of Hong Kong, Shenzhen · Shanghai AI Laboratory

Abstract

Retrosynthesis enables the discovery of viable synthetic routes to target molecules. It plays a central role in modern drug discovery and materials design. Retrosynthesis involves molecular graph transformations that can change both connectivity and graph size. These transformations may introduce reactant components absent from the target while revising the product-derived structure. To model these transformations, we propose RetroGEF, a flow-based generative model for single-step retrosynthesis. Starting from the target molecule, it constructs possible reactants by adding atoms and changing bonds in the molecular graph. RetroGEF models molecular transformations and changes in graph size within the same generative process, rather than relying on a fixed-size graph canvas. It learns this process directly from product--reactant pairs without requiring a prescribed edit order. Experiments on representative retrosynthesis benchmarks demonstrate that RetroGEF achieves state-of-the-art performance.

Figures & tables

Panel of Figure 1 — Figure 1: RetroGEF : product-relative complete edits, bridge supervision, and generation. Paired endpoints supervise edit rates (dashed arrows); the shared network observes $(P,X_{t},t)$ . Each generated edit updates the graph before the next rate prediction. States are schematic.

Panel of Figure 2 — Figure 2: Edit order and complete graph updates. A: Two correction orders reach the same graph, with total bridge intensity $2\to 1\to 0$ . B: Location and attributes jointly define one update. Blue: product atoms; orange: added atoms.

Panel of Figure 3 — Figure 3: Edit-rate prediction. Intensity and complete-edit probability determine each rate. Dashed links align product atoms; gray links inside the Transformer denote feature interactions, not molecular bonds.

Method	Top-1	Top-3	Top-5	Top-10
RetroSim	32.8	–	–	56.1
LocalRetro	39.1	53.3	58.4	63.7
GLN	39.3	–	–	63.7
RetroPrime	44.1	59.1	62.8	68.5
R-SMILES	48.9	66.6	72.0	76.4
NAG2G	49.7	64.6	69.3	74.0

Table 1: USPTO-Full test results. Top-

k

accuracy (%).

	Pred.		Oracle
Graph changes	RetroDiT	RetroGEF	RetroDiT	RetroGEF
$\leq 5$	65.1	67.6	78.7	81.9
6–10	45.1	48.8	57.3	63.2
11–20	32.6	36.8	43.8	52.1
$>20$	25.7	29.5	33.6	42.5

Table 3: Transformation complexity on USPTO-Full. Top-1 accuracy (%) on 90,598 shared test products.

Added atoms	RetroGEF				RetroDiT
Added atoms	1	3	5	10	1	3	5	10
0	84.8	91.5	93.0	93.6	71.9	85.4	88.2	90.5
1–2	79.2	90.6	92.0	92.7	76.4	88.6	90.8	92.4
3–5	56.0	71.1	74.7	77.0	49.4	67.7	72.5	76.6
6–10	54.4	68.1	70.9	73.1	47.1	63.8	68.6	72.4
11–20	33.4	45.0	47.8	50.0	26.0	40.1	45.6	50.4

Table 4: USPTO-Full by added atoms. Oracle test accuracy (%) under the common decoding budget.

Figure 4: Accuracy across all ten USPTO-50K reaction classes. Shared nonuniform radial scale (%), labeled in the first panel. FG: functional group.

USPTO-50K
	Oracle		Pred.
Variant	Top-1	Top-10	Top-1	Top-10
Reference	72.3	97.0	62.4	91.9
Constant addition rate	72.1	93.0	61.4	88.1
Uniform time sampling	71.6	95.8	61.2	90.5
No edge-feature updates	73.8	96.4	62.4	91.5

Table 5: Training and architecture variants. Test accuracy (%) on USPTO-50K and USPTO-Full.

Figure 5: USPTO-Full Oracle Top-1 predictions.

Figure 6: A sampled ten-edit construction path for a USPTO-50K protection reaction. The two rows list edits 0–4 and 5–10; the final panel is the complete reactant set.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Edit type	Location	Attributes	State change
Edit type	Location	Attributes	$\Delta\|V\|$	$\Delta\|E\|$
Attached atom addition	existing atom	$\mathbf{m},\mathbf{e}$	$+1$	$+1$
Isolated atom addition	graph	$\mathbf{m}$	$+1$	$0$
Generated-atom deletion	most recent generated leaf/isolated $u$	none	$-1$	$-d(u)$
Bond addition	absent unordered pair	$\mathbf{e}$	$0$	$+1$
Bond deletion	existing bond	none	$0$	$-1$

Appendix

Table 8: Complete graph edits and their state effects. A selected edit contains all fields needed to construct its deterministic successor. Here

\mathbf{m}

and

\mathbf{e}

are complete atom and bond attribute tuples, and

d(u)

is the current degree of a deleted atom. All edit types preserve unrelated graph records.

Dataset	Variant	Oracle				Pred.
		1	3	5	10	1	3	5	10
USPTO-50K	Reference	72.3	92.9	96.0	97.0	62.4	82.8	88.6	91.9
	Constant addition rate	72.1	89.0	91.3	93.0	61.4	80.9	85.1	88.1
	Uniform time sampling	71.6	91.5	94.5	95.8	61.2	82.3	87.4	90.5
	No edge-feature updates	73.8	92.6	95.5	96.4	62.4	83.8	88.4	91.5
USPTO-Full	Reference	68.6	80.7	82.8	84.1	54.3	70.1	74.1	77.1

Appendix

Table 9: Training and architecture variants. Test Top-

1/3/5/10

accuracy (%).

Condition	Partial hints	Recovery	Top- $k$
Condition	Partial hints	Recovery	1	3	5	10
Pred.	Reference		62.4	82.8	88.6	91.9
	$\checkmark$	–	62.8	83.6	88.0	90.4
	–	$\checkmark$	60.2	81.1	85.4	88.0
	$\checkmark$	$\checkmark$	63.3	82.8	88.2	90.2
Oracle	Reference		72.3	92.9	96.0	97.0

Appendix

Table 10: Supervision comparison on USPTO-50K. Top-

k

accuracy (%): reference on test, training variants on validation. Column maxima are bold.

Reaction class	$n$	Method	Pred.				Oracle
Reaction class	$n$	Method	Top-1	Top-3	Top-5	Top-10	Top-1	Top-3	Top-5	Top-10
Alkyl./aryl.	1,516	RetroDiT	65.1	85.4	90.6	94.3	73.7	92.2	96.4	98.3
Alkyl./aryl.	1,516	RetroGEF	67.5	85.4	90.8	93.6	74.5	94.2	97.6	98.0
Acylation	1,188	RetroDiT	74.9	91.9	94.6	96.2	82.2	97.4	98.2	98.7
Acylation	1,188	RetroGEF	75.4	92.8	95.8	97.0	81.4	97.0	98.0	98.1
C–C formation	511	RetroDiT	41.9	63.2	69.9	75.0	55.4	80.2	87.1	92.0

Appendix

Table 11: USPTO-50K accuracy by reaction class. Top-

k

accuracy (%) on 4,930 paired products; bold denotes the larger value in each pair. FG denotes functional-group interconversion.

Panel of Figure 7 — Figure 7: Reactant construction on USPTO-Full under Oracle conditioning. Added atoms are orange; the panels illustrate fragment recovery, coupling partner recovery, and bond-order revision.

Figure 8: Complete trajectories for USPTO-50K classes 1–6. Orange marks added atoms and their bonds; blue marks changes to inherited atoms or bonds. Every arrow represents one edit.

Figure 9: Complete trajectories for USPTO-50K classes 7–10. The oxidation example shows all 14 edits, including ten atom additions.

Figure 10: Complete trajectories across USPTO-Full graph-edit families. The three paths contain 14, nine, and four edits; each path is shown on its recorded inference grid.

Explore similar work

RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction
Mar 13, 2026Hanbum Ko, Chanhui Lee, Ye Rin Kim +4Reaction Synthesis
Department of Artificial Intelligence, Korea University · Materials Intelligence Lab, LG AI Research · Department of Statistics, Korea University
Representation-Guided Discrete Molecular Graph Retrosynthesis
May 23, 2026Jiahai Huang, Anjie Qiao, Zhen Wang +2Chemical Structures Synthesis
Sun Yat-sen University Guangzhou, China · University of Science and Technology of China Hefei, China
Project Ariadne: Prompt-Conditioned Route Generation for Synthesis Planning
Jun 23, 2026Anton Morgunov, Victor S. BatistaMolecular Generation Synthesis
Yale University

Mar 13, 2026, cs.LGJ/K move · Enter open · S save

RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction

Hanbum Ko, Chanhui Lee, Ye Rin Kim, Rodrigo Hormazabal+3

Department of Artificial Intelligence, Korea University · Materials Intelligence Lab, LG AI Research · Department of Statistics, Korea University

Reaction Synthesis LLM Reasoning Strategies Language Modeling

Open paper AskPDF arXiv

Retrosynthesis prediction aims to identify reactants that can synthesize a given product molecule. Although molecular large language models (LLMs) have recently shown promising results, most existing methods either generate reactants directly or provide only generic product-level analysis, without explicitly reasoning about bond-disconnection strategies that justify specific reactant choices. This paper proposes RetroReasoner, a retrosynthetic reasoning model that captures chemists' strategic disconnection-based thinking. RetroReasoner is trained with supervised fine-tuning and reinforcement learning. For supervised fine-tuning, SyntheticRetro generates structured disconnection rationales paired with reactant predictions. For reinforcement learning, a round-trip reward evaluates predicted reactants by passing them through a forward synthesis model and rewarding predictions that reconstruct the original product. RetroReasoner can also be applied to multi-step retrosynthetic planning by incorporating it into a parallelized Monte Carlo tree search framework, reducing search time while increasing the number and diversity of valid synthetic pathways. Experimental results show that RetroReasoner outperforms prior baselines, including not only molecular LLMs but also retrosynthesis-specific expert models, and generates a broader range of feasible reactant proposals, especially for challenging reaction instances. The code is available at https://github.com/KU-AGI/RetroReasoner.

Abstract

Explore similar work

RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction

Representation-Guided Discrete Molecular Graph Retrosynthesis

Project Ariadne: Prompt-Conditioned Route Generation for Synthesis Planning