ResearchGate

Seediscussions,stats,andauthorprofilesforthispublicationat:

/publication/391707103

Bridgingchemistryandarti?cialintelligencebyareactiondescription

language

ArticleinNatureMachineIntelligence·May2025DOI:10.1038/s42256-025-01032-8

CITATIONS0

READS13

12authors,including:

JiachengXiong

ShanghaiInstituteofMateriaMedica,ChineseAcademyofSciences20PUBLICATIONS439CITATIONS

SEEPROFILE

WeiZhang

ShanghaiInstituteofMateriaMedica

21PUBLICATIONS101CITATIONS

SEEPROFILE

FuZunyun

ShanghaiInstituteofMateriaMedica

38PUBLICATIONS822CITATIONS

SEEPROFILE

XiangtaiKong

ScienceforLifeLaboratory

20PUBLICATIONS92CITATIONS

SEEPROFILE

Allcontentfollowingthispagewasuploadedby

MingyueZheng

on16May2025.Theuserhasrequestedenhancementofthedownloadedfile.

Publishedonline:xxxxxxxx

naturemachineintelligence

Article

/10.1038/s42256-025-01032-8

Bridgingchemistryandartificialintelligencebyareactiondescriptionlanguage

Received:15May2024

Accepted:4April2025

Checkforupdates

JiachengXiong1,2,6,WeiZhang1,2,6,YinquanWang1,3,JiataoHuang4,YuqiShi1,2,MingyanXu1,2,ManjiaLi1,ZunyunFu1,XiangtaiKong1,2,YitianWang1,2,ZhaopingXiong

5&MingyueZheng1,2

Withthefast-paceddevelopmentofartificialintelligence,largelanguagemodelsareincreasinglyusedtotacklevariousscientificchallenges.

Acriticalstepinthisprocessisconvertingdomain-specificdataintoa

sequenceoftokensforlanguagemodelling.Inchemistry,moleculesareoftenrepresentedbymolecularlinearnotations,andchemicalreactionsaredepictedassequencepairsofreactantsandproducts.However,thisapproachdoesnotcaptureatomicandbondchangesduringreactions.Here,wepresentReactSeq,areactiondescriptionlanguagethatdefinesmoleculareditingoperationsforstep-by-stepchemicaltransformation.BasedonReactSeq,languagemodelsforretrosynthesispredictionmayconsistentlyexcelinallbenchmarktests,anddemonstratepromising

emergentabilitiesinthehuman-in-the-loopandexplainableartificial

intelligence.Moreover,ReactSeqhasallowedustoobtainuniversalandreliablerepresentationsofchemicalreactions,whichenablenavigationofthereactionspaceandaidintherecommendationofexperimentalproceduresandpredictionofreactionyields.WeforeseethatReactSeqcanserveasabridgetonarrowthegapbetweenchemistryandartificialintelligence.

Artificialintelligencetechnologies,representedbylargelanguagemodels(LMs),haveachievedunprecedentedbreakthroughsinnaturallanguageprocessing,influencingthemodelofscientificresearch

1

,

2

.Inthelifesciencesdomain,LMsarenowusedtominehiddeninforma-tionfromproteinandgenesequences,achievingremarkableresults.NotableexamplesincludeESM,whichinterpretstheproteinfunctionsandstructuresfromtheirsequences

3

,

4

,andGeneformer,whichpredictsgenefunctionsandinteractions

5

.Inchemistryandpharmaceuticals,theimportantconceptofchemicalLMs(CLMs),whichhandlechemicalmoleculesandreactions,hasalsoemerged

6

–

8

.

Unlikenaturallanguages,proteinsandgenes,chemicalmole-culeslackinherentsequentialrepresentations.CLMscapitalizeonchemist-definedmolecularlinearnotationstolearnandgenerate

molecularstructures.Themostcommonlyusedmolecularlinearnotationisthesimplifiedmolecularinputlineentrysystem(SMILES)

9

.Recently,toenhancetheperformanceofCLMsinspecifictasks,somenewmolecularlinearnotationsweredesigned.Forinstance,SELFIESwasdevelopedtohelpCLMsproducevalidmolecularstructures

10

,andPSMILESwasintroducedtofacilitatethelearningofpolymerrepresentations

11

.

However,thesemolecularlinearnotationsarealldesignedtodescribethestaticstructuresofchemicalmolecules.Theycannotexplicitlydescribethecrucialaspectofchemistry,namelytheprocessofatomandbondchangesinmoleculesduringachemicalreaction

12

.ThissignificantlyrestrictstheapplicationofLMsinchemicalreactionpredictionandrepresentation.CurrentLMsforchemicalreaction

1DrugDiscoveryandDesignCenter,StateKeyLaboratoryofDrugResearch,ShanghaiInstituteofMateriaMedica,ChineseAcademyofSciences,

Shanghai,China.2UniversityofChineseAcademyofSciences,Beijing,China.3DepartmentofMedicinalChemistry,SchoolofPharmacy,FudanUniversity,Shanghai,China.4SchoolofPhysicalScienceandTechnology,ShanghaiTechUniversity,Shanghai,China.5ProtonUnfoldTechnologyCo.Ltd,Suzhou,

China.6Theseauthorscontributedequally:JiachengXiong,WeiZhang.e-mail:

myzheng@

NatureMachineIntelligence

Article

/10.1038/s42256-025-01032-8

NatureMachineIntelligence

Previouslanguagemodelforreactionprediction(SMILEStoSMILES)

C1(C)=C2CCCCC2=NN1C1=CC(OC(C)C)=C(Cl)C=C1F

LM

Il

O

ClF

O

NH2

+

ONH

CC(C1CCCCC1=O)=O.

CC(OC1=CC(NN)=C(C=C1Cl)F)C

uLackofinteractivityuPoorinreactionrepresentationuLackofinterpretability

Ourproposedmethod(SMILEStoReactSeq)

F

F

22Attach

O

Break

79

2120

N

、N

6

5

1118Cl19

/

8N10

Cl

\廠16

/\

A

N

1

1213

O1514\

7Break

Reduce

O

314

O

LM

2

Attach

17

_!!<[O:1]><><[O:1]>

C1(C)_C2CCCCC2!NN!1C1=CC(OC(C)C)=C(Cl)C=C1F<[O:1]><><[O:1]>

4

3

7

5

6

10

O

1

2

8

O9

CC(C1CCCCC1=O)=O

C1(C)=C2CCCCC2=NN1C1=CC(OC(C)C)=C(Cl)C=C1F

Editingoperation

Molecule

Promptencoding

embedding

decoding

C1(C)=C2CCCCC2=NN1!C1=CC(OC(C)C)=C(Cl)C=C1F

uHuman-in-the-loopuBetterreactionrepresentationuExplainablereasoning

Fig.1|Overviewofthiswork.AcomparisonbetweenthepreviousLMforreactionpredictionbasedonSMILESandourproposedmethodbasedonReactSeq.

prediction,involvingforwardandretrosynthesisprediction,typi-callydirectlytranslatethelinearnotationsofproductsandreactantsintoeachother,whichhasbeenconsistentlycriticizedforlackinginterpretabilityandinteractivity(Fig.

1

,top)

13

–

15

.Recently,Wangetal.

16

andThakkaretal.

17

decomposedthetransformationfromproducttoreactantintotwosequentialsteps:first,translatingtheproductintosynthonsusingatransformer,andthentranslatingthesesynthonsintoreactantswithanothertransformer.Whilethistwo-stagedesignimprovesthemodel’sinterpretabilityandinteractivity,italsoincreasesthemodel’scomplexityandcompromisesitsend-to-endproperties.Additionally,duetothelimitationsofSMILESsyntax,thesemethodscanonlyindicatetheatomsinvolvedinthereactionwithoutdetailingtheirspecificchanges.Furthermore,whilepretrainedLMsexcelinrepresentationlearningofvarioussequencedata

18

,

19

,similaradvance-mentsforchemicalreactionsarenotablylacking.ExistingCLMscanlearnatomtransformationfromunmappedreactiondata

20

.However,generatingmeaningfulvectorrepresentationsforthistransformationprocessremainschallenging.Currentself-supervisedreactionrepre-sentationsstillstruggletoeffectivelycapturethesimilaritiesbetweendifferentreactions

21

.

Therefore,toadvancetheapplicationofLMsinchemistry,devel-opingnewlanguagesfordescribingchemicalreactionsisnecessary.Ideally,thislanguageshouldmakethereactionpredictionsmoreaccu-rateandinterpretable,enablingclarificationofthetransformationprocessofatomsandbonds.Thepredictionofthetransformationprocessshouldbecontrollable,allowingchemiststoguideLMswiththeirknowledge

22

.Moreover,thislanguageshouldenableLMstogen-eratebetterreactionrepresentationsfordiversedownstreamtasks.

Inthiswork,weintroduceareactiondescriptionlanguagenamedReactSeq,designedtomeettheaforementionedobjectives(Fig.

1

,bottom).Inspiredbyretrosynthesisprocess,ReactSeqdefinesboththeproductstructureandthemoleculareditingoperations(MEOs)

requiredtotransformitbackintoreactantmolecules.TheseMEOsincludethebreakingandchangingofchemicalbonds,alterationsinatomiccharges,andtheattachmentofleavinggroups(LGs),amongothers(Fig.

2a

).InaReactSeq-basedretrosynthesisLM,thereactantisnotgeneratedtoken-by-tokenfromscratch.Instead,itistransformedfromtheproductmoleculethroughtheseMEOs.Thisensurespreciseatommappingbetweenthepredictedreactantsandtheproducts,enhancingthemodel’sinterpretability.UsingReactSeq,avanillatrans-formercanachievestate-of-the-artperformanceinretrosynthesisprediction.Moreover,ReactSeqfeaturesexplicittokensdenotingMEOs,enablingtheencodingofhumaninstructions.Ourresultsshowthathumanexperts’promptscansignificantlyenhancethemodel’sperformanceandevenguideitinexploringnewreactions.Inaddition,theembeddingsofthoseMEOtokensprovideauniversalandreliablereactionrepresentation.Theseself-supervisedrepresentationscannaturallydistinguishbetweenreactiontypesandevaluatetheirsimi-larity,facilitatingsimilarreactionretrieval,experimentalprocedurerecommendationandreactionyieldprediction.

OverviewofReactSeq

ReactSeqconsistsoftwoparts:aheaderandatail(Fig.

2a

).Theheaderincludesthestructuraldetailsofatargetmoleculeandinformationonchangestoitsatomsandbonds,describinghowtotransformitintothecorrespondingsynthons.ThetailincludesthestructuresoftheLGsandtheirconnectionpositionswiththesynthons,describinghowtocompletethesynthonsintoreactants.InstandardSMILES,tokensfordoubleandtriplebondsarevisible,whiletokensforsinglebondsarehidden.However,thehiddentokenscanbespecifiedusingSMILESwithexplicitbonds(Fig.

2b

).ByreplacingthesebondtokensinSMILESwithMEOtokens(forexample,usinganexclamationmark‘!’torepre-sentbreakingabond),weobtaintheheaderofReactSeqthatrecordsthechangesandbreaksinchemicalbonds.Sometargetmoleculesin

Article

/10.1038/s42256-025-01032-8

a

C1(Cl)=NC=C(CBr)C(Cl)=C1

C1=C2CCC(O)C2=CC=C1F

C1=CC(OCC(=O)O)=CN=C1

CC(=O)C1=CC=CC=C1

HeaderofReactSeqTailofReactSeq

C1(Cl)=NC=C(C!Br)C(Cl)=C1<><C1(=O)CCC(=O)[N:1]1>

C1=C2CCC(;O)C2=CC=C1F

F

Changetodoublebond

OH

C1=CC(OCC(=O)[~OH])=CN=C1<[CH3:1]>

Atta8

Attachmentpoint

C[sC](_O)C1=CC=CC=C1<><>

O

CChangetosinglebondChangetoS-configuration

Breakbond

andconnecttoleavinggroup

Changebond

Directly

connectto

leavinggroup

Changebondandchangechirality

b

SMILES

C1(Cl)=NC=C(CBr)C(Cl)=C1

Cl

2

N

3

4

110

9

8

7Cl

5

6

Br

SMILESwith

explicitbonds

C1(-Cl)=N-C=C(-C-Br)-C(-Cl)=C-112345678910

Fig.2|IllustrationofReactSeq.a,SeveralrepresentativeexamplesofReactSeqdescribingMEOsinretrosynthesis,includingbondbreaking,bondchanging,

connectingtoLGsandchiralitychange.b,VisualizationofhiddentokensforsinglebondsinSMILES.

retrosynthesisdonotinvolvebreakingorchangingofbondsbetweenheavyatoms.Instead,theyaredirectlyconnectedtoLGs.Inthesecases,weinitiallyconverttheatomtokentotheexplicithydrogenmode,suchaschangingOto[OH],andthenaddacorrespondingMEOtoken(~)toit.Furthermore,changesinchirality,chargeandcis–transisomerismarealsodefinedinReactSeq.

ToobtainthetailofReactSeq,first,theatomsinthetargetmol-eculesthatcouldconnecttoLGsareidentified,knownasattachmentpoints.TheseincludeatomsdirectlyconnectedtoLGsorinvolvedinbondbreakingorreduction.TheLGsofeachattachmentpointareenclosedinanglebracketsandsortedbasedontheatomicindexesoftheirconnectedattachmentpoints.Followingthesesteps,astandardheader-to-tailReactSeqisobtained,maintaininghighalignmentwiththeSMILESofthetargetmolecule.ArelatedworktoReactSeqisthecondensedgraphofreaction(CGR),whichrepresentschemicalreac-tionsaspseudo-moleculesandcangeneratelinearnotationsforthesepseudo-molecules

23

.CGRdefinesoperationsforatomicandbondchangesbutfailstoprovideoperationsforcompletingLGs,whichrestrictsitscapacitytodepictunbalancedreactions.Additionally,thelinearrepresentationofCGRcannotensurealignmentwitheitherthereactantorproductSMILES.FurtherdetailsaboutReactSeqareprovidedintheMethodssection.

ReactSeqimprovesretrosynthesispredictionperformance

TodemonstratetheapplicationofReactSeq,wefirstuseditforret-rosynthesispredictionusingavanillatransformerwithoutanyaddi-tionalmodifications.Table

1

presentsacomprehensivecomparisonofourproposedmethodsandothermethodsontheUSPTO-50kdataset.

Ourmodeloutperformedallothers,regardlessofwhetherreactiontypesweregiven.

Ourmodelusesatwo-stageretrosynthesisreasoningstrategythatfirstidentifiesthereactioncentreandthencompletesthesyn-thons,correspondingtotheheaderandtailcomponentsofReactSeq,respectively.Whilemanygraphedit-basedmethods(forexample,G2Gs(ref.

24

),RetroXpert

25

,GraphRetro

26

)andthesequence-basedmethodRetroPrimealsousethisstrategy,theyoftenunderperformintopk(k≥3)accuracy.Thismightstemfromtheiruseofdifferentmodelsforeachstage’stasks,whichdisruptsthecontinuityoftheinformationflowbetweenthetwotasks,leadingtotheaccumu-lationoferrors.Incontrast,ReactSeq’sheader-to-tailstructureconsolidatesthedescriptionsoftworetrosynthesisstagesintoonesequence,enablingsequentialprocessingofthesetwotaskswithanend-to-endmodel,thusachievingstate-of-the-arttopkaccuracy.Furthermore,currentgraphedit-basedmethodssuchasGraph2Edits(ref.

27

),GraphRetro

26

andRetroExplainer

28

formulatesynthoncom-pletionasaclassificationproblem,selectingLGsfromapredefinedvocabulary.ThislimitstheirabilitytogeneratenewLGsandexplorenewreactions.Conversely,ourmodelgeneratesSMILESofLGstokenbytoken,allowingforthegenerationofnewLGsabsentfromthetrainingset,offeringgreaterflexibilitytodiscovernewchemicaltransformations.However,thisalsointroducestheriskofgenerat-ingchemicallyinfeasibleLGs,asdemonstratedinSupplementaryFig.1.WhilethreepredictedLGsappearchemicallyplausible,withanalogousreactionshavingbeenreported,thelastone—thedifluo-romethanesulfonicacidgroup—hasnotbeenobservedinexistingreactions.Thishighlightstheneedformorecarefulevaluationofnewpredictionsbeforeexecution.

Article

/10.1038/s42256-025-01032-8

NatureMachineIntelligence

Table1|TopkaccuracyofourproposedmethodandothermodelsonUSPTO-50kdataset

Class

Model

Topkaccuracy(%)

Reactionclassunknown

3510k=1

Reactionclassknown

k=1

3

5

10

Template-basedmethods

Retrosim

52

37.3

54.7

63.3

74.1

52.9

73.8

81.2

88.1

Neuralsym

53

44.4

65.3

72.4

78.9

55.3

76.0

81.4

85.1

GLN

54

52.5

69.0

75.6

83.7

64.2

79.1

85.2

90.0

LocalRetro

44

54.2

76.8

80.4

90.3

-

-

-

-

Graphedit-basedmethods

G2Gs(ref.

24

)

48.9

67.6

72.5

75.5

61.0

81.3

86.0

88.7

RetroXpert

25

50.4

61.1

62.3

63.4

62.1

75.8

78.5

80.9

MEGAN

55

48.1

70.7

78.4

86.1

60.7

82.0

87.5

91.6

GraphRetro

26

53.7

68.3

72.2

75.5

63.9

81.5

85.2

88.1

G2Retro(ref.

56

)

53.9

74.6

80.7

86.6

63.1

84.2

88.5

91.7

Graph2Edits(ref.

27

)

55.1

77.3

83.4

89.4

67.1

87.5

91.5

93.8

RetroExplainer

28

57.7

79.2

84.8

91.4

66.8

88.0

92.5

95.8

NAG2G(ref.

57

)

55.1

76.9

83.4

89.9

67.2

86.4

90.5

93.8

RetroCaptioner

58

54.3

76.3

82.6

88.1

67.2

86.0

90.3

93.4

Sequence-basedmethods

SCROP

59

43.7

60.0

65.2

68.7

59.0

74.8

78.1

81.1

Aug.Transformer

45

53.2

-

80.5

85.2

-

-

-

-

Dual-TF

60

53.6

70.7

74.6

77.0

65.7

81.9

84.7

85.9

BARTSmiles

50

55.6

-

74.2

80.9

-

-

-

-

RetroPrime

16

51.4

70.8

74.0

76.1

64.8

81.6

85.0

86.9

Chemformer

47

54.3

-

62.3

63.0

-

-

-

-

R-SMILES

46

56.3

79.2

86.2

91.0

-

-

-

-

ReactSeq

58.9

80.5

86.4

91.4

68.5

89.2

93.1

95.9

Note:thehyphenrepresentsthatthecorrespondingresultsarenotreportedandboldrepresentsthebestresult.TheresultsforLocalRetrocomefromtheauthor’smostrecentupdateonGitHub.

Wefurtheranalysedourmodel’sperformanceacrossvariousreac-tiontypes.Forrarereactiontypessuchascyclizationandfunctionalgrouptransformations,themodelmaintainedstrongperformance,achievingtoptenaccuraciesof84.6and91.3%,respectively(Supplemen-taryFig.2).However,theaccuracysignificantlydecreasedto65.6%forreactionsinvolvingstereochemicalchanges(SupplementaryTable1).Thisdiscrepancylikelystemsfromknowledgetransferability.Cyclizationandfunctionalgrouptransformationreactions,althoughunderrepre-sentedinourdataset,involveMEOssuchasbondbreakingandattachingLGs,whicharecommoninotherreactions.Thisallowsthemodeltoeffec-tivelytransferknowledgefrommorefrequentreactionstopredicttheserarertypes.Incontrast,stereochemicalchangesinvolveuniquerulesthatcannotbeinferredfromreactionswithoutsuchtransformations.

WealsoevaluatedourmethodonthelargerUSPTO-MITdataset,whereitdemonstratedsuperiorperformancewithaccuracyratesof60.5,78.5,83.3and87.6%forthetopone,three,fiveandtenpredictions,respectively(SupplementaryTable2).Theseresultshighlightourmethod’sapplicabilitytolarge-scaledatasets.Moreover,tovalidatethegeneralizabilityofourmodel,weconductedanexternalevalua-tionusingELN,areal-worldreactiondataset

29

.Ourmodelachievedstate-of-the-artperformanceonthisdataset(SupplementaryTable3),emphasizingitsrobustgeneralizationcapabilities.AblationstudiesaboutourReactSeqlanguageandtrainingstrategyareprovidedintheSupplementaryInformationC.1.

ReactSeqenablesinterpretableretrosynthesisprediction

Conventionalsequence-basedretrosynthesismethodsdirectlyconvertproductSMILESintoreactantSMILES,failingtodescribethespecific

transformationprocessfromproducttoreactants.ReactSeqaddressesthislimitationbydividingtheretrosynthesispredictionintotwophases:identifyingthereactioncentreandcompletingthesynthons.Thistwo-stagemoleculareditingapproachsimulatestheretrosyn-theticanalysisworkflow,aligningbetterwithhumanexpertintuitioncomparedtotheapproachinspiredbyspecificreactionmechanismsandofferinggreatergenerality

28

.

SupplementaryTable4showcasestheperformanceofourmodelandothermethodsinthetwostagesofretrosynthesis.Ourmodelachieved73.1%top-oneaccuracyinidentifyingreactioncentresand77.6%insynthoncompletion,significantlysurpassingpreviousmeth-ods.Amongthesetwo-stageretrosynthesismethods,RetroFormerisalsosequence-based.However,limitedbySMILESsyntax,themethodidentifiesonlyatomsinvolvedinbondbreakingwhenidentifyingthereactioncentre,failingtocapturethechangesofotheratomsandbonds.Duringthesynthoncompletionstage,RetroFormerdirectlytranslatessynthonSMILESintoreactantSMILES,notclarifyingtheprocessofattachingLGs.Somesequence-basedmethodsusetheatten-tionweightstoindicatereactioncentresandperformatomicmapping.However,theexplanationsprovidedbytheattentionmechanismdonotguaranteeconsistencywiththeactualtransformationsperformedbythemodel

30

,

31

.Incontrast,ReactSeqenablesLMstoaccuratelytrackthechangesofatomsandbondsthroughoutthereactionprocesswithoutanymodificationstothemodelarchitecture,offeringamorestream-linedandreliablesolutionforinterpretableretrosynthesisprediction.

Figure

3a

presentstheprocessofgeneratingReactSequsingbeamsearchbyourretrosynthesismodel,wherethetotalprobabilityistheproductoftheprobabilitiespredictedateachstep.Notably,thetotalpredictionprobabilityismainlyinfluencedbyMEOtokens,

Article

/10.1038/s42256-025-01032-8

NatureMachineIntelligence

Reactioncentreidentification

Synthoncompletion

C(=O)C1=CC=C(Cl)S1)N1C=CN=C1<>

Cl:1]>

OH:1]>

NH

P1=0.977

CC=C(Cl)O

Br:1]><>

NH

[O:1]>

P1=0.022

Rank-5

*Cl

P2=0.637

*OH

P2=0.358

*Cl

P2=0.485

*Br

P2=0.394

*O

P2=0.003

a

12

Cl11

10S1398

6O7

5HN

4

3

21

15N1416N18

17

Product

log(totalprobability)

0

?1

?2

?3

?4

?5

?6

?7

N!

CC(CC

<[

O

*

SCl

N

*

N

(=O)C1=

S

Cl

1)!N1C

C

S

=CN=C1<

[Cl:1]><>

Rank-1Rank-2Rank-3Rank-4

*

NN

*

C

0369121518212427303336394245

Steps

O

SCl

Cl

N

N

H2N

t

Rank-1predictionP=0.622

O

SCl

OH

N

N

H2N

t

Rank-2predictionP=0.350

O

O

SCl

SCl

Br

Cl

N

H

N

H

HN

HN

N

N

Rank-4predictionPt=0.009

t

Rank-3predictionP=0.011

O

SCl

O

N

N

H2N

t

Rank-5predictionP=0.002

b

600

500

Count

400

300

200

100

0

Retrosynthesisprediction

1,500

Correct

Incorrect

1,200

Count

900

600

300

0

1.00.80.60.40.20

Confidence

ReactioncentreidentificationCorrect

2,000

Incorrect

1,500

Count

1,000

500

0

1.00.80.60.40.20

Confidence

Synthoncompletion

Correct

Incorrect

1.00.80.60.40.20

Confidence

Fig.3|InterpretableretrosynthesispredictionwithReactSeq.a,Presentation

oftheinferenceprocessesofourmethod.P1andP2refertothepredictedprobabilitiesforreactioncentreidentificationandsynthoncompletion,

respectively,whilePtreferstothetotalprobabilityforfinalpredictionresult.b,Therelationshipbetweenmodel’spredictiveconfidencemeasuredby

predictiveprobabilityanditsaccuracyacrossdifferenttasks.

whichrepresentthedynamictransformationofdecomposingprod-uctmoleculesandcompletingsynthons.Theremainingtokens,usedtodescribethestaticmolecularstructurearepredictedconsistentlystable.Incontrast,thetotalpredictionprobabilityoftheSMILES-basedmodelisinfluencedbymanytokens,suggestingamoreintricatedecision-makingprocess(SupplementaryFig.3).Furthermore,thepredictionprobabilitiesfortheheaderandtailtokensinReactSeqallowcalculationofmodel’sconfidenceinitspredictionsateachstage.Thereisacleartrendwheretheaccuracyofpredictionsimprovesasthemodel’sconfidenceincreases(Fig.

3b

).

ReactSeqenablesprompt-basedreactionprediction

Inretrosynthesisprediction,humanexpertsoftenhaveinsightsregard-ingthelocationandtypeofreactionthatshouldoccur.WebelievethatincorporatingthisexpertknowledgethroughpromptscanguideLMs

togeneratemoreaccuratepredictions.Akeychallengeofthisprocessislinguisticallyencodingthediversehumanprompts.Thakkaretal.achievedthisbytaggingatomsinSMILES

17

.However,theirpromptswererestrictedtoindicatingbond-breakingpositions,failingtoencodeothertypesofhumanprompt.Conversely,ReactSeqdefinestokensrep-resentingvariousMEOs,enablingamorecomprehensiveandrefinedencodingofhumanprompts.

Todemonstratethis,wetrainedapromptlearningmodelusingReactSeq.AsshowninFig.

4a

,themodeliscapableofprocessingvari-oustypesofmoleculareditingpromptencodedbyReactSeq,andisguidedtoperformspecificreactiontransformations.Wefurthertestedthisprompt-basedlearningstrategyontheUSPTO-50kdataset,achiev-ing96.6%top-oneaccuracyinidentifyingreactioncentresand74.9%top-oneaccuracyinpredictingfinalreactants,significantlyoutper-formingthemodelwithoutprompts(Fig.

4b

).Itiscrucialtopointouttheseresultswereobtainedusingfullyaccuratehumanprompts,which

Article

/10.1038/s42256-025-01032-8

NatureMachineIntelligence

a

N

N

WhatifI…

N

OH

F

N

Predictionwithprompt1

N

BrAttach

N

NBreak

F

OH

N

N

N

N

H

Br

OH

N

F

Q1:breakthisbond?

Prompt1:N#CC1=CC=C(C(C(O)CC2=CC=C(F)C=C2)!N2C=NC=N2)C=C1

Q2:changethisbondtodoublebond?

Prompt2:N#CC1=CC=C(C(C(;O)CC2=CC=C(F)C=C2)N2C=NC=N2)C=C1

Q3:changethisbondtosinglebond?

Prompt3:N_CC1=CC=C(C(C(O)CC2=CC=C(F)C=C2)N2C=NC=N2)C=C1

N

Predictionwithprompt2

\\Mg+

N

N

N··Attach

OH

NBreak

F

Changetodoublebond

N

N

+Mg

OF

N

Predictionwithprompt3

N

N

N

O

Attach

OH

F

N

Changetosinglebond

b

100

Accuracy(%)

80

60

40

20

0

SynthonpredictionReactantprediction

99.297.2

96.6

95.8

98.599.192.9

74.9

Top-1Top-3Top-5Top-10

c

NO

Prompt

Break

N\

N

O

Output

NCl/Br/INCl/Br

HO

BHO

+Mg

N

O

Suzukicoupling

Grignardreaction

NCl/Br/I

N

O/

N

O

N+