Strong specialist era
Sequence, graph and template systems dominated single-step prediction, while planners connected them into route search.
Molecular TransformerLocalRetroGraph2EditsNAG2GRetro*AiZynthFinder
RETROSYNTHESIS MODEL INTELLIGENCE
From one predicted reaction to competing chemical hypotheses.
Retrosynthesis is no longer one architecture, one score or one route. Compare how expert rules, templates, sequence models, graph models, ensembles, flows, chemistry LLMs and route planners see the same synthesis problem differently.
Evidence boundaryBenchmark numbers are not treated as directly comparable across different datasets, splits and evaluation protocols.
22models & planners mapped
8architecture families
5systems from 2025–2026
1rule: no leaderboard is synthesis truth
THE 2024 → 2026 SHIFT
Retrosynthesis is becoming a convergence problem.
This is not a clean replacement of older methods. Established models remain valuable anchors. The change is that newer systems increasingly add diversity, explicit reasoning and complementary inductive biases rather than optimizing only one exact-match answer.
Diversity + reasoning + convergence
Ensembles, flow models and chemistry-specialist LLMs broaden the reaction space and make disagreement between methods scientifically useful.
RetroChimeraRetro SynFlowRetroDFM-RC3LMInsilico SSRSModern planners
MODEL FAMILIES
Six complementary ways to see the chemistry.
The practical question is not “which family wins?” but which inductive bias is useful for the task—and what independent method should challenge it.
Rules & precedent
Anchor the search in known transformations and traceable chemical precedent.
LHASA · RetroSim · GLN · LocalRetroSequence models
Learn reaction patterns as molecular-language transformations and generate fast challengers.
Molecular Transformer · Chemformer · R-SMILESGraph models
Represent atoms, bonds and reaction centres directly for structurally explicit disconnections.
RetroXpert · Graph2Edits · NAG2GEnsembles & flows
Increase diversity and exploit disagreement between complementary inductive biases.
RetroChimera · Retro SynFlowChemistry LLMs
Generate broader strategic alternatives and explicit rationale while remaining validation-dependent.
RetroDFM-R · C3LM · Insilico SSRSRoute planners
Turn one-step proposals into multistep searches terminating in defined starting-material stock.
Retro* · AiZynthFinder · ASKCOSINTERACTIVE MODEL ATLAS
Compare the model—not just the score.
Search and filter the curated retrosynthesis landscape. Select up to four systems to compare representation, template dependence, search behaviour, strengths, limitations and access side-by-side.
Architecture family
Planning layer
22of 22 systems shown
C3LM
Chemistry Constraint-Consistent LM with plausibility- and novelty-aware Top-K training.
Planning layerSingle-step
RepresentationChemistry-specialised LLM
Template modeTemplate-free
Search behaviourTop-K generation + reward shaping
Best atOOD plausibility-focused generation and complementary reaction-space exploration.
WatchVery recent research; prospective validation and independent replication remain important.
InsilicoMMAI-Chem-SSRS 1.1
2.6B chemistry-specialist LLM for diverse single-step retrosynthesis.
Planning layerSingle-step
RepresentationSMILES-specialised LLM
Template modeTemplate-free
Search behaviourTop-K text generation
Best atCompact specialist model aimed at chemically plausible, diverse disconnections rather than one exact answer.
WatchProvider benchmark methodology is not directly comparable with classical exact-match leaderboards.
Retro SynFlow
Discrete flow matching from synthons toward diverse precursor sets.
Planning layerSingle-step
RepresentationGraph / discrete states + flow matching
Template modeTemplate-free
Search behaviourStochastic flow + forward-model steering
Best atDiverse reactant generation with inference-time steering toward feasible chemistry.
WatchResearch-stage; steering quality depends on reward and forward-model calibration.
RetroChimera
Learned ensemble of complementary retrosynthesis inductive biases.
Planning layerSingle-step
RepresentationMultiple submodels / learned reconciliation
Template modeMixed ensemble
Search behaviourCandidate fusion + learned ranking
Best atRobustness under distribution shift and chemist-aligned ranking through model complementarity.
WatchLower-ranked suggestions can hallucinate and still require independent chemistry verification.
RetroDFM-R
Reasoning-driven chemistry LLM trained with chemically verifiable RL rewards.
Planning layerSingle-step
RepresentationLarge language model / molecular text
Template modeTemplate-free
Search behaviourReasoning + autoregressive generation
Best atExplicit reasoning, broad chemical priors and interpretable retrosynthetic rationale.
WatchReasoning fluency is not proof of chemical feasibility.
NAG2G
Node-aligned graph-to-graph Transformer preserving structural information.
Planning layerSingle-step
Representation2D graph + 3D conformation + node alignment
Template modeTemplate-free
Search behaviourAutoregressive graph generation
Best atStructural modelling without reducing the problem to SMILES alone.
WatchGraph generation, atom alignment and conformer preparation add complexity.
Graph2Edits
Autoregressive molecular graph editing inspired by arrow-pushing logic.
Planning layerSingle-step
RepresentationMolecular graph + edit sequence
Template modeSemi-template / graph edits
Search behaviourSequential graph edits
Best atInterpretable atom/bond transformations and arbitrary-length edit sequences.
WatchWrong early edits can cascade; single-step accuracy does not imply route quality.
R-SMILES
Root-aligned SMILES reduce product/reactant sequence mismatch.
Planning layerSingle-step
RepresentationAligned SMILES tokens
Template modeTemplate-free
Search behaviourSequence generation
Best atImproves sequence correspondence so capacity focuses on reaction changes.
WatchAlignment choices and canonicalisation still influence behaviour.
Retroformer
Reaction-centre-aware Transformer for template-free retrosynthesis.
Planning layerSingle-step
RepresentationTransformer with molecular structural context
Template modeTemplate-free
Search behaviourAutoregressive generation
Best atBridges sequence generation with explicit reaction-centre information.
WatchBehaviour still depends on atom mapping and preprocessing choices.
Chemformer
BART-style molecular language model pretrained on SMILES.
Planning layerSingle-step
RepresentationSMILES language model
Template modeTemplate-free
Search behaviourAutoregressive sequence generation
Best atStrong pretrained molecular representation across reaction tasks.
WatchPretraining does not remove chemistry-validation or OOD concerns.
LocalRetro
Local atom/bond reaction templates predicted with graph attention.
Planning layerSingle-step
RepresentationGraph attention + local templates
Template modeLocal template-based
Search behaviourLocal edit / template ranking
Best atEfficient local reaction modelling with explicit transformation semantics.
WatchUnseen local patterns and template extraction quality remain limiting.
MHNreact
Modern Hopfield networks retrieve and rank reaction templates.
Planning layerSingle-step
RepresentationTemplate embeddings + molecular representation
Template modeTemplate-based
Search behaviourAssociative template retrieval
Best atStrong template generalisation and efficient associative retrieval.
WatchStill bounded by the learned template set and reaction-record quality.
AiZynthFinder
Open-source retrosynthetic planner with MCTS and pluggable policies.
Planning layerPlanner
RepresentationReaction tree + expansion policy + stock
Template modeDefault template-guided; extensible
Search behaviourMCTS and other algorithms
Best atPractical route finding to purchasable precursors with configurable stock and scoring.
WatchResults depend strongly on policy data, stock files, filters and route scoring.
GLN
Graph Logic Network combines molecular graphs with reaction-template logic.
Planning layerSingle-step
RepresentationMolecular graph + templates
Template modeTemplate-based
Search behaviourTemplate scoring / expansion
Best atPrecedent-aware prediction with strong structural context.
WatchTemplate vocabulary remains a hard boundary on possible transformations.
Retro*
Neural-guided A* search for multistep retrosynthetic planning.
Planning layerPlanner
RepresentationSearch graph + learned value / one-step model
Template modeDepends on expansion model
Search behaviourA* / best-first route search
Best atEfficient multistep search when paired with a calibrated expansion model.
WatchQuality inherits one-step model, stock definition and search-value errors.
RetroXpert
Predict reaction centres, split into synthons, then complete reactants.
Planning layerSingle-step
RepresentationGraph + sequence completion
Template modeSemi-template
Search behaviourTwo-stage one-step generation
Best atChemically intuitive reaction-centre decomposition.
WatchReaction-centre errors propagate into synthon completion.
SCROP
Transformer retrosynthesis with an explicit structure-correction step.
Planning layerSingle-step
RepresentationSMILES sequence
Template modeTemplate-free
Search behaviourBeam generation + correction
Best atReduces invalid reactant SMILES from early template-free sequence models.
WatchStill inherits sequence-representation and exact-match evaluation limitations.
Molecular Transformer
Reaction prediction and retrosynthesis as molecular translation.
Planning layerSingle-step
RepresentationSMILES tokens / Transformer
Template modeTemplate-free
Search behaviourBeam search over sequences
Best atFlexible sequence modelling and a durable reaction-prediction baseline.
WatchSequence validity is not chemical validity; token ordering can hide structural context.
ASKCOS
Synthesis-planning platform combining retrosynthesis, forward models and route search.
Planning layerPlatform
RepresentationMultiple one-step models + search controllers
Template modeMultiple approaches
Search behaviourMCTS / Retro* / interactive planning
Best atEnd-to-end synthesis planning with forward and route-level utilities.
WatchOutputs remain model-dependent and require chemical review.
NeuralSym / Segler–Waller
Neural networks learn which retrosynthetic rule fits a target.
Planning layerSingle-step
RepresentationFingerprints + explicit reaction rules
Template modeTemplate-based
Search behaviourTemplate ranking; later paired with tree search
Best atTransparent reaction rules with learned context relevance.
WatchCannot propose chemistry absent from the extracted template library.
RetroSim
Similarity-driven retrosynthesis using reaction precedents.
Planning layerSingle-step
RepresentationMolecular fingerprints + reaction templates
Template modeTemplate / analog retrieval
Search behaviourOne-step ranking; can feed a planner
Best atLiterature-like chemistry and interpretable precedent retrieval.
WatchPerformance follows template and analog coverage.
LHASA / expert CASP
Human-coded synthesis knowledge and strategic transforms.
Planning layerPlanner
RepresentationExpert rules + reaction transforms
Template modeExplicit expert rules
Search behaviourRule-based route exploration
Best atInterpretability, traceable synthetic logic and precedent-rich planning.
WatchCoverage and maintenance depend heavily on human-curated knowledge.
CONVERGENCE LOGIC
No single model becomes “truth.”
BioAtlas keeps the model evidence and limitations visible. BayesARC can then use complementary proposal engines, chemistry checks and route-level evidence before human review.
STARTTarget moleculeStructure · context · constraints
Rules / precedentSequenceGraphEnsemble / flowChemistry LLMRoute planner
REACTION HYPOTHESIS COREAgreement and disagreement stay visible.Multiple valid disconnections can coexist. Duplicate proposals are not counted as independent evidence.
Canonicalise + deduplicate
One reaction proposed six ways is still one hypothesis.
Forward-check
Test whether the proposed precursors can plausibly regenerate the product.
Precedent + novelty
Separate known transformations from unsupported or genuinely novel suggestions.
Chemical plausibility
Inspect selectivity, stereochemistry, protecting groups, conditions and obvious liabilities.
Multistep search
Connect one-step hypotheses to routes terminating in defined starting materials.
Route qualification
Compare step count, availability, cost, hazard, scalability and route robustness before review.
Chemist review required.AI proposes and challenges. Experts decide what deserves experimental follow-up.
BENCHMARK GOVERNANCE
Top-1 accuracy ≠ synthesis success.
A higher exact-match score can be useful evidence within one controlled benchmark, but it is not a universal measure of route feasibility or synthesis success.
- 01Different datasets and splits can change headline accuracy.
- 02Retrosynthesis is one-to-many: several valid precursor sets may exist for one product.
- 03Single-step exact match does not measure full-route practicality.
- 04A benchmark-correct route can still be poor on cost, safety, selectivity, scale or precursor availability.
MODEL INTELLIGENCE → SYNTHESIS DECISION
Use BioAtlas to understand the models. Use BayesARC to qualify the routes.
Keep exact versions, sources, benchmark context, limitations and disagreement visible before a model output is allowed to influence a synthesis decision.