Rules & precedent
Human-coded transformations and analog retrieval. Interpretable, but coverage depends on the rule base.
LHASA · RetroSim
Bayes Pharma · Synthesis Intelligence
Retrosynthesis has moved beyond one model, one route.
Explore how expert systems, template models, sequence transformers, graph editors, ensembles, flow models and chemistry LLMs differ—and where each earns a place in modern synthesis planning.
Science-first: reported benchmark numbers are not treated as directly comparable across datasets, splits or evaluation protocols.
20models & planners mapped
7architectural families
3planning layers
1rule: no single leaderboard tells the whole story
The landscape
Every generation solved a different bottleneck.
Retrosynthesis systems differ in what they represent, how they propose disconnections, whether they use explicit reaction precedent, and whether they can plan full routes. Newer does not automatically mean better for every chemistry problem.
Learn which rule fits
Neural networks rank templates instead of applying every possible transformation.
NeuralSym · GLN · LocalRetroChemistry as translation
SMILES-to-SMILES generation unlocked flexible template-free prediction at scale.
Molecular Transformer · Chemformer · R-SMILESReason on molecular structure
Graph models encode bonds, atoms and reaction centers directly instead of inferring them from text alone.
Graph2Edits · NAG2GDiversity + convergence
Ensembles and flow models explicitly exploit complementary biases and diverse feasible proposals.
RetroChimera · Retro SynFlowChemistry-specialist LLMs
Large models add broader chemical priors, Top-K generation and increasingly explicit reasoning—but still need chemistry validation.
RetroDFM-R · Insilico SSRS · C3LMInteractive model atlas
Compare the model, not just the score.
Search by model, architecture or use case. Add up to four models to the comparison tray to inspect their scientific differences side-by-side.
Direct comparison
What actually changes between models?
Representation and search behavior matter as much as headline accuracy. The table below highlights the dimensions that affect real synthesis-planning behavior.
| Dimension | Template / precedent | Sequence / Transformer | Graph / edit | Flow / ensemble | Chemistry LLM |
|---|---|---|---|---|---|
| Primary representation | Reaction templates / fingerprints | SMILES tokens | Molecular graph, atoms, bonds | Multiple model spaces or stochastic paths | Tokens + learned chemical priors |
| Interpretability | High | Medium | High–medium | Medium | Variable |
| Uncommon reaction coverage | Limited by template library | Can generalize beyond templates | Strong structural generalization potential | Designed for breadth / diversity | Potentially broad; OOD reliability must be tested |
| Common failure mode | Template coverage gap | Invalid / implausible sequence generation | Incorrect edit sequence / attachment | Diverse but chemically weak proposals | Plausible-sounding hallucination |
| Best scientific role | Precedent-aware anchor | Fast flexible challenger | Structural / reaction-center challenger | Diversity & consensus | Reasoning / broad proposal challenger |
| Needs downstream validation? | Yes | Yes | Yes | Yes | Absolutely |
Model selection
Which family is strongest for which job?
The best practical stack is role-based. Use a model where its inductive bias helps, then challenge it with a fundamentally different method.
Known chemistry & precedent
Favor explicit templates, analog retrieval and systems trained on broad reaction databases.
RetroSim · GLN · LocalRetro · ASKCOSComplex structural disconnections
Graph-first models make atom/bond changes visible and preserve molecular topology.
Graph2Edits · NAG2GDiverse hypothesis generation
Use models designed to explore more than one plausible precursor set.
Retro SynFlow · C3LM · Insilico SSRSReasoned explanation
LLM-style systems can add explicit rationale and broader chemical context, but require verification.
RetroDFM-R · C3LMRobust single-step ranking
Use complementary inductive biases rather than trusting one architecture.
RetroChimeraFull route planning
Single-step models need a search layer, stock definition, stopping criteria and route ranking.
AiZynthFinder · ASKCOS · Retro*BayesARC concept
The Retrosynthesis Convergence Engine.
No single retrosynthesis model becomes “truth.” Independent proposal engines compete, chemistry checks prune weak routes, and route-level evidence stays inspectable through human review.
STARTTarget moleculeSMILES · structure · programme context
↓
LAYER 1Independent proposal enginesDifferent biases · different chemistry
RetroChimeraensembleNAG2GgraphSynFlowflowRetroDFM-Rreasoning LLMC3LM / SSRSchemistry LLMLocalRetrotemplate
↓
↓
LAYER 2Multistep searchRetro* · MCTS · AiZynthFinder / ASKCOS-style planners
↓
LAYER 3Route-level ranking
stepscoststockhazardstereoprecedentyield riskscale-up
↓
✓
FINAL AUTHORITYQualified chemist reviewPredicted routes remain hypotheses until expert and experimental evidence supports them.
A Top-1 number is not a synthesis strategy.
Retrosynthesis is intrinsically one-to-many. Multiple precursor sets can be chemically valid even when only one appears in a benchmark record.
Dataset matters
USPTO-50K, USPTO-FULL, Pistachio, CAS-derived and proprietary corpora create very different difficulty and coverage.
Split matters
Random, temporal, reaction-class and out-of-distribution splits answer different generalization questions.
Ground truth is incomplete
Exact-match evaluation can punish chemically valid alternatives that differ from the recorded literature route.
Planning ≠ one step
A strong one-step precursor model can still perform poorly inside a full search tree if it generates weak branches or poor calibration.
Evidence trail
Primary sources behind the atlas.
Model cards link to papers, official repositories or current provider pages where available. “New” means publication/release recency, not automatic superiority.
RetroChimeraMicrosoft Research · 2025
Retro SynFlowDiscrete flow matching · 2025
RetroDFM-RReasoning LLM + RL · 2025
C3LMPlausibility-aware chemistry LLM · 2026
InsilicoMMAI-Chem-SSRS 1.1Microsoft Foundry model card · 2026
NAG2GJACS Au · 2024
Graph2EditsNature Communications · 2023
AiZynthFinderAstraZeneca Molecular AI
ASKCOSMIT synthesis planning platform