Verification-Guided Retrosynthesis: Reaction Feedback and Evidence Timing

Hongxuan Shi; Hankz Hankui Zhuo

cs.AIOct 28, 2026

Verification-Guided Retrosynthesis: Reaction Feedback and Evidence Timing

Authors: Hongxuan Shi, Hankz Hankui Zhuo

Organizations: School of Artificial Intelligence, Nanjing University, China

Abstract

Retrosynthesis planning seeks synthesis routes from purchasable starting materials to a target molecule. Search over predicted reactions can connect a target to available materials, yet an infeasible reaction can invalidate the resulting route. Evaluating route quality only after search has finished cannot correct errors made during search. We introduce Verification-Guided Reflow (VGR), incorporating reaction verification as a decision action in AND/OR search. Verification results are represented as binary verdicts; queries are scheduled from the search structure, and acquired verdicts exclude reactions or adjust search priorities. We define verified solvability as the fraction of targets for which a complete route accepted stepwise by the verifier is found. On Retro190, feeding verification results into search increases verified solvability from 51.1% to 62.6% relative to route evaluation without feedback. Further controlled experiments show that earlier verification and immediate prioritization of accepted reactions each improve route discovery, yet combining them performs worse than either change alone. We therefore propose early pruning and delayed guidance: rejected reactions are pruned immediately, while accepted reactions gain search priority when processed as part of a complete candidate route. In a separate comparison with the verification scheduling rule fixed, this strategy increases verified solvability from 61.1% with immediate guidance to 71.1%, using fewer proposal calls and unique verifier queries.

Figures & tables

Positive release	$r_{t}(R)$	Effect on search
neutral	$0$	retain native priority
immediate	$1$	release credit upon acquisition
root-delayed	$h_{t}^{\mathrm{root}}(R)$	release while processing a closed root candidate

Table 1: Positive-evidence release controls. Acquisition and negative activation are separate from the release rule.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Policy	Acquisition	Negative acts	Positive release	Release variable	$\alpha_{\min}$
post-hoc	exposed closed route	never in search	never in search	$0$	–
closure-prune	root proof closure	immediately	neutral	$0$	1
closure-guide	root proof closure	immediately	at acquisition	1	$[0,1)$
PI-neutral	after first downstream call	immediately	never	$0$	–
PI-immediate	after first downstream call	immediately	at acquisition	1	$[0,1)$
PI-root-delayed	after first downstream call	immediately	root proof closure	$h_{t}^{\mathrm{root}}(R)$	$[0,1)$

Appendix

Table 5: Exact mapping of reported policies to acquisition and activation clocks.

Component	Identity and transformation	Search use
Retro190 targets	Public 190-target list distributed with Retro-fallback [ 15 ] ; canonicalized in the released order	primary mechanism and policy comparison
Pistachio-derived targets	100 DESP Pistachio Hard pairs in source order; canonical targets retained; required starts excluded	target-only transfer with the same stock
Inventory	23,081,583 unique canonical eMolecules entries; zero target-in-stock hits	terminal purchasability test
Primary search	template MLP, top-50 applicable sanitized and deduplicated outcomes; learned retro* value; best-first retro*; depth guard 50	principal environment
Transfer search	RetroChimeraDeNovo R-SMILES 2 USPTO-FULL; five root augmentations, beam 50, official parsing and rank aggregation, final top-50; constant-zero value	configuration-level stress test
Verifier	ReactionT5v2 forward USPTO-MIT checkpoint; blind reactants; five-beam decoding; canonical non-stereochemical product match	frozen binary feedback and delivery certificate

Appendix

Table 6: Reproducibility card for the principal datasets and model configurations.

Policy	Search unique	Delivery checks	Cached total unique
PI-neutral	46,089	838	46,089
PI-immediate	49,428	901	49,428
PI-root-delayed	43,974	969	43,974
closure	3,872	889	3,872

Appendix

Table 7: Query accounting in the uncapped release-timing run.

Policy	Topological	Certified	$\Sigma N$	Search unique	Node visits
PI-neutral	165	126	44,471	46,089	299,700
PI-immediate- $\alpha=.25$	152	116	48,758	49,428	144,272
PI-root-delayed- $\alpha=.25$	165	135	42,579	43,974	308,057
closure acquisition- $\alpha=.25$	172	128	47,627	3,872	281,593
soft maturity- $\beta=1$	168	137	42,376	43,651	262,703
soft maturity- $\beta=2$	165	135	42,310	43,553	280,331

Appendix

Table 8: Continuous structural-maturity controls on Retro190.

Policy	Topological	Certified	$\Sigma N$	Search unique	Node visits
post-hoc	82	61	22,804	0	129,969
neutral	77	64	18,652	18,837	120,860
immediate- $\alpha=.25$	79	66	17,859	17,941	59,503
root-delayed- $\alpha=.25$	77	66	18,053	18,243	131,613
soft maturity- $\beta=1$	80	67	17,331	17,496	113,247
soft maturity- $\beta=2$	79	69	17,513	17,758	122,458

Appendix

Table 9: Continuous structural-maturity controls on the 100-target Pistachio-derived target-only set.

$M_{\mathrm{search}}$	PI-neutral	PI-immediate	PI-root-delayed	Closure
100	118	112	113	126
500	126	115	135	128
uncapped	126	116	135	128

Appendix

Table 10: Certified targets under shared target-local unique search-query ceilings.

Policy	Certified	Topological	$\Sigma N$	AUSC@500	Feedback events	Cached unique	Positive maturation
post-hoc	97	172	57,365	.397	0	3,502	–
closure- $\alpha=.25$	121	172	48,970	.473	4,715	3,790	–
early-prune	121	165	46,416	.497	46,067	47,803	–
PI-root-delayed- $\alpha=.25$	135	165	42,579	.537	56,169	43,974	28.0%
FU-root-delayed- $\alpha=.25$	144	148	34,580	.621	416,570	99,399	1.19%

Appendix

Table 11: Observed search–verification frontier under the reported acquisition sweep.

PUCT policy	Certified	Topological	$\Sigma N$	AUSC-N@500
PI-neutral	132	151	39,318	.569
PI-immediate- $\alpha=.25$	132	148	38,778	.573
PI-root-delayed- $\alpha=.25$	136	151	37,465	.590
PI-immediate- $\alpha=0$	125	143	39,492	.564
PI-root-delayed- $\alpha=0$	137	151	37,009	.593

Appendix

Table 12: Full-190 PUCT coverage and one-step-call results.

PUCT policy	$\Sigma$ simulations	AUSC-sim@100k	$\Sigma$ nodes
PI-neutral	926,369	.687	9,595,796
PI-immediate- $\alpha=.25$	1,126,527	.685	18,520,797
PI-root-delayed- $\alpha=.25$	866,633	.708	8,725,386
PI-immediate- $\alpha=0$	1,315,355	.652	22,687,335
PI-root-delayed- $\alpha=0$	855,568	.714	8,611,078

Appendix

Table 13: Full-190 PUCT simulation and graph results.

Retrosynthesis

See all 25 papers

Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback
Oct 10, 2026Zixun Lan, Zhenfu Liu, Yi Li +6Retrosynthesis Reinforcement Learning
Xi’an Jiaotong-Liverpool University, Suzhou 215000, China · Sichuan University of Science & Engineering, Zigong 643000, China · Fudan University, Shanghai 200433, China +1
RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning
Oct 12, 2026Ye Rin Kim, Yongjun Jeong, Seongjun Yun +4Retrosynthesis LLM Planning
Korea University, Department of Artificial Intelligence · Korea University, Department of Statistics
Project Ariadne: Prompt-Conditioned Route Generation for Synthesis Planning
Jun 23, 2026Anton Morgunov, Victor S. BatistaAutoregressive Generation Retrosynthesis
Yale University

Oct 10, 2026J/K move · Enter open · S save

Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback

Zixun Lan, Zhenfu Liu, Yi Li, Xinheng Wang+5

Xi’an Jiaotong-Liverpool University, Suzhou 215000, China · Sichuan University of Science & Engineering, Zigong 643000, China · Fudan University, Shanghai 200433, China+1

Retrosynthesis Reinforcement Learning

Open paper AskPDF arXiv

Single-step retrosynthesis is a one-to-many inverse problem, yet most models are trained and evaluated against a single recorded reactant set. Reference-based training gives no explicit positive signal for unrecorded alternatives, and exact-match evaluation rejects non-matching predictions regardless of chemical plausibility. We propose Retrosynthesis via Reinforcement Learning from Forward Feedback (RetroRLFF), which optimizes reactant generation using both reference agreement and forward consistency with the target product. Starting from a supervised decoder-only Transformer, RetroRLFF generates complete reactant candidates and uses a frozen forward reaction model to predict products. A sequence-level reward combines reference agreement and forward consistency while penalizing invalid outputs. Direct Advantage-Based Policy Optimization updates the policy without human preference annotations or a preference-trained reward model. On USPTO-50K, RetroRLFF outperforms all 14 baselines in candidate-level round-trip accuracy at k = 1, 3, 5, and 10. With approximately 50 million parameters, it achieves 90.1% top-1 round-trip accuracy, exceeding R-SMILES, the strongest competing baseline on this metric, by 8.5 percentage points. Compared with its supervised policy, round-trip accuracy increases from 83.1% to 90.1%, while exact-match accuracy is maintained at 57.9% versus 57.5%. In an anonymized expert assessment of 131 randomly selected test cases, organic synthesis experts prefer RetroRLFF over EditRetro in 67.9% of comparisons. These results support forward feedback as a training signal for improving forward consistency while retaining top-1 reference recovery. Expert assessment provides complementary evidence of synthetic feasibility on the selected cases, although neither forward consistency nor expert preference establishes experimental feasibility.

Abstract

Retrosynthesis

Explore similar work

Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback

RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning

Project Ariadne: Prompt-Conditioned Route Generation for Synthesis Planning