Paper deep dive
Multi-Agent Closed-Loop Reasoning for Organic Structure Elucidation from Multimodal Spectra
Bingsen Xue, Zhuojun Jiang, Jianhao Zhang, Mingcheng Gu, Yizhe Yuan, Yongtai Zhuo, Yifan Zhang, Li Wang, Ya Su, Yue Yuan, Jiang Liu, Xueqian Kong, Cheng Jin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/18/2026, 4:21:39 AM
Summary
The paper introduces MACROS, a multi-agent system for automated organic structure elucidation from multimodal spectroscopic data (1H NMR, 13C NMR, HSQC, IR). MACROS employs a closed-loop reasoning mechanism involving drafting, rethinking, and refining agents to emulate expert hypothesis-testing. Trained on 100M simulated and 1.6M experimental spectra-molecule pairs, it demonstrates zero-shot generalization to real-world samples, recovering textbook spectroscopic correlations and exhibiting emergent chemical intuition, such as a ring-first parsing preference. The system achieves sixfold faster and 40% more accurate elucidation compared to existing methods.
Entities (25)
Relation Signals (26)
MACROS → processes → 13C NMR
confidence 98% · routinespectroscopicdata,including¹HNMR,¹³CNMR,heteronuclear singlequantumcoherence(HSQC)NMR,andFourier-transforminfrared(IR)spectra
MACROS → processes → HSQC
confidence 98% · routinespectroscopicdata,including¹HNMR,¹³CNMR,heteronuclear singlequantumcoherence(HSQC)NMR,andFourier-transforminfrared(IR)spectra
MACROS → processes → IR
confidence 98% · routinespectroscopicdata,including¹HNMR,¹³CNMR,heteronuclear singlequantumcoherence(HSQC)NMR,andFourier-transforminfrared(IR)spectra
MACROS → processes → 1H NMR
confidence 98% · routinespectroscopicdata,including¹HNMR,¹³CNMR,heteronuclear singlequantumcoherence(HSQC)NMR,andFourier-transforminfrared(IR)spectra
YiFan Zhang → affiliatedwith → Chinese Academy of Sciences
confidence 95% · Yifan Zhang 2 ,LiWang 2 ... 2 CASKeyLaboratoryofGreenProcessandEngineering,InstituteofProcessEngineering,Chinese AcademyofSciences
Li Wang → affiliatedwith → Chinese Academy of Sciences
confidence 95% · Yifan Zhang 2 ,LiWang 2 ... 2 CASKeyLaboratoryofGreenProcessandEngineering,InstituteofProcessEngineering,Chinese AcademyofSciences
Ya Su → affiliatedwith → Fudan University
confidence 95% · YaSu 3 ... 3 DepartmentofNeurology,NationalCentreforNeurologicalDisorders,NationalClinicalResearchCentre forAgingandMedicine,HuashanHospital,FudanUniversity
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Following the molecular discovery and synthesis revolutions, scalable automated structure elucidation from routine spectroscopic data remains an outstanding challenge. Despite decades of computational efforts, no existing system achieved reliable reasoning over unseen spectra. Here, we propose MACROS, a multi-agent system automating structure elucidation by emulating expert iterative hypothesis-testing. Trained on 100M simulated and 1.6M experimental spectra-molecule pairs, it natively supports arbitrary combinations of routine spectroscopic techniques. It achieves unprecedented zero-shot generalization to diverse real-world samples, correctly identifying synthetic compounds, natural products and metabolites above 500 Da with 1D NMR. Remarkably, MACROS spontaneously recovers textbook spectroscopic correlations from unassigned data and exhibits emergent chemical intuition such as a ring-first parsing preference, learning fundamental chemical principles rather than memorizing database patterns. MACROS augments chemists via collaboration to deliver sixfold faster, 40% more accurate elucidation. MACROS establishes a scalable foundation for fully automated structure elucidation, and catalyzes accelerated molecular discovery toward autonomous laboratories.
Tags
Links
- Source: https://arxiv.org/abs/2608.14720v1
- Canonical: https://arxiv.org/abs/2608.14720v1
Trouble viewing inline? Open PDF directly →
Full Text
45,717 characters extracted from source content.
Expand or collapse full text
Multi-AgentClosed-LoopReasoningforOrganicStructureElucidationfrom MultimodalSpectra BingsenXue #,1 ,ZhuojunJiang #,2 ,JianhaoZhang 1 ,MingchengGu 1 ,YizheYuan 1 ,YongtaiZhuo 1 ,Yifan Zhang 2 ,LiWang 2 ,YaSu 3 ,YueYuan 4 ,JiangLiu 5 ,XueqianKong 6 ,ChengJin* ,1,7,8,9 1 InstituteofMedicalRobotics,SchoolofBiomedicalEngineering,ShanghaiJiaoTongUniversity,Shanghai 200240,China 2 CASKeyLaboratoryofGreenProcessandEngineering,InstituteofProcessEngineering,Chinese AcademyofSciences,Beijing100190,China 3 DepartmentofNeurology,NationalCentreforNeurologicalDisorders,NationalClinicalResearchCentre forAgingandMedicine,HuashanHospital,FudanUniversity,Shanghai200240,China 4 DepartmentofMicrobiologyLaboratory,MinhangDistrictCenterforDiseaseControl,Shanghai200240, China 5 DepartmentofElectricalandComputerEngineering,JohnsHopkinsUniversity,Baltimore,USA 6 SchoolofChemistryandChemicalEngineering,ShanghaiJiaoTongUniversity,Shanghai200240,China 7 BeijingAndingHospitalCapitalMedicalUniversity,Beijing,China 8 NationalClinicalResearchCenterforKidneyDiseases,Beijing,China 9 InstituteofDigitalMedicine,ShanghaiJiaoTongUniversity,Shanghai,China # Theseauthorscontributedequallytothiswork. * Correspondence:chengjin520@sjtu.edu.cn(C.J.). Summary: Followingthemoleculardiscoveryandsynthesisrevolutions,scalableautomatedstructureelucidationfrom routinespectroscopicdataremainsanoutstandingchallenge.Despitedecadesofcomputationalefforts,no existingsystemachievedreliablereasoningoverunseenspectra.Here,weproposeMACROS,amulti-agent systemautomatingstructureelucidationbyemulatingexpertiterativehypothesis-testing.Trainedon100M simulatedand1.6Mexperimentalspectra-moleculepairs,itnativelysupportsarbitrarycombinationsof routinespectroscopictechniques.Itachievesunprecedentedzero-shotgeneralizationtodiversereal-world samples,correctlyidentifyingsyntheticcompounds,naturalproductsandmetabolitesabove500Dawith1D NMR.Remarkably,MACROSspontaneouslyrecoverstextbookspectroscopiccorrelationsfromunassigned dataandexhibitsemergentchemicalintuitionsuchasaring-firstparsingpreference,learningfundamental chemicalprinciplesratherthanmemorizingdatabasepatterns.MACROSaugmentschemistsvia collaborationtodeliversixfoldfaster,40%moreaccurateelucidation.MACROSestablishesascalable foundationforfullyautomatedstructureelucidation,andcatalyzesacceleratedmoleculardiscoverytoward autonomouslaboratories. Keywords:Structureelucidation,Multi-agentsystem,Closed-loopreasoning,Spectroscopy Introduction Molecularstructuredeterminationisafoundationalstepinchemical,biological,andbiomedicaldiscovery, underpinningnaturalproductresearch,drugdevelopment,studiesofbiomolecularfunction,andsynthetic chemistry 1–3 .Preciselyresolvingandquantifyingbiomolecularstructuresinthecomplexintracellular environmentisalsoafundamentalchallengeincellbiology,whichlinksphenotypetomolecularstateand function 4,5 .High-throughputsynthesis,automatedexperimentation,andmodernspectroscopyhave dramaticallyexpandedtheproductionandstructuralcharacterizationofsmallmolecules,metabolites,and biomolecularligands.Yetscalableandfullyautomateddenovostructureelucidationfromspectroscopic dataremainsanoutstandingchallenge.Evenforexperiencedspectroscopistsandconventional computationalapproaches,moleculeswithhighmolecularweight,complexarchitectures,ordense stereochemistryremainexceptionallydifficulttoresolve.Morefundamentally,thechallengeextendsbeyond automatingexistingworkflowstowardbuildinganinterpretableelucidationprocesscapableofresolving larger,morecomplex,andstructurallyambiguousmoleculeswithgreaterprecisionandrobustness 6,7 .Each spectroscopicmodalityencodesadistinctdimensionofmolecularinformation,includinglocalchemical environment,atomicconnectivity,stereochemicalconfiguration,andfunctional-groupcomposition 8,9 . Accuratestructuredeterminationthereforerequiresintegratingthesecomplementarysignalsthrough sequential,hypothesis-drivenreasoning,whichhasprovendifficulttoautomate 10,11 .Approachesthat achieveastepchangeinbothautonomousandinterpretableelucidationwillthusbetransformativeforthe nextgenerationofbiomedicalandpharmaceuticaldiscovery. Inreal-worldlaboratoryworkflows,routinespectroscopicdata,including¹HNMR,¹³CNMR,heteronuclear singlequantumcoherence(HSQC)NMR,andFourier-transforminfrared(IR)spectra,serveastheprimary basisforstructuralcharacterizationacrosschemicalandbiologicalcontexts 12–14 .Massspectroscopy(MS) andheteronuclearNMR( 11 B, 19 F, 31 P,etc.)canprovideuniquesupplementaryinformationforformulaor heteronuclearvalidation.Forreactionsorbiologicalsamplesthatproduceexpectedcompounds,structural confirmationcanbeachievedsimplybymatchingspectralpeaksagainstreferencestandards.Bycontrast, unexpectedoruncharacterizedmoleculesrequirerigorousdenovostructuralelucidation.Spectraldatabase retrievaliswidelyappliedtoidentifymoleculeswithcomparablespectralprofiles 15,16 ,yetsuchmethodsfail torecognizecompoundsthatlieoutsideexistinglibraries.Insuchcases,onlydenovostructureelucidation withoutanylibraryreferencesisfeasible,whichisfundamentallydistinctfromlibrary-basedretrieval. Humanexpertsperformdenovostructureelucidationthroughiterativehypothesistesting:inferring candidatesubstructuresfromchemicalshiftsandcouplingpatterns,crossvalidatingthesehypothesesacross multiplespectralmodalities,andrefiningstructuralassignmentsbyeliminatinginconsistentisomers 17 . Currentgenerativeapproachesforautomatingdenovoelucidationfallintotwocategories,yetbothremain inadequateforreal-worldapplications.Spectra-guidedmoleculargenerationmodels,whentrainedondata fromspecificsources,achievestrongperformanceonfixedmodalitieswithintheirtrainingdistribution 18,19 , butlacktheflexibilityrequiredtohandlereal-worldspectralvariability.Repurposedlargelanguagemodels (LLMs)andvision-languagemodels(VLMs)excelatgeneralchemistrytasksviapromptengineering 20–22 , buttheirperformancedegradessignificantlyonstructureelucidation 23,24 .Evenwithmassivescaleand pretrainingonextensivechemicalandspectralcorpora,thesemodelsarefundamentallydesignedfor linguisticorvisual-semanticreasoning,notspectroscopicanalysis.Structureelucidationfrommultimodal spectra,bycontrast,isahighlysymbolic,logic-driventaskthatreliesonspectral-molecularpattern matching,multi-stepdeduction,anditerativevalidation 25 .Crucially,neitherapproachincorporatesthe explicititerativereasoningworkflowthathumanspectroscopistsemployforcomplexspectra:hypothesis generation,cross-validationacrossmodalities,andstep-by-steprefinement.Effectivesolutionstherefore requiregenerativefoundationmodelsdesignedfromthegroundupforspectroscopicreasoning:(i) self-supervisedpretrainingonlarge,unpairedspectralcorporatoacquireintrinsicphysicochemical knowledge,(i)inherentlyflexiblearchitecturesthattoleratearbitraryormissingmodalities,and(i)explicit closed-loopreasoningmechanismsthatemulatetheiterativehypothesis-testingworkflowofexpert chemists. Here,weproposeMACROS(Multi-AgentClosed-LoopReasoningforOrganicStructure),apurpose-built multi-agentsystemformultimodalspectroscopicstructureelucidationfromroutinespectra:¹HNMR,¹³C NMR,HSQC,andIR.Unlikerepurposedgeneral-purposefoundationmodels,MACROSwasdesignedfrom firstprinciplesforspectroscopicreasoning.Modality-specificagentswerepretrainedinaself-supervised manneronlarge-scalesimulatedspectra,integratedintoahierarchicalframeworkandjointlytrainedonover 100millionspectra-structurepairs,thenfine-tunedonunassignedexperimentalspectratoadaptto real-worldspectralvariability.Atitscore,MACROSimplementsanexplicitmulti-agentclosed-loop reasoningmechanismthatemulatestheiterativeworkflowofexpertspectroscopists.Spec2Molproposes candidatestructures,Mol2Specsimulatesspectraandrankscandidatesbymulti-criterionsimilarity,while PromptSpec2Molrefinescandidatesovermultiplecyclesusinghigh-rankingoutputsasguidance.This closed-loopschemeenablestest-timeperformancescaling,allowingprogressiveaccuracygainsduring inference.Theflexibleagentworkflowreadilyaccommodatesadditionalinformationsources(e.g.,reactant structures)andfacilitatesfutureextensionstohigher-dimensionalNMRandotherspectroscopicmodalities. Forinstance,MS-derivedformulasorfragments,annotatedbyspecializedtools(e.g.,SIRIUS 26 ,BUDDY 27 ), areacceptedasmolecularpromptsforindirectincorporation,leveragingmatureMSsolutionsand circumventingthescarcityofpairedMS-NMRdatasets.Notably,attention-basedinterpretationrecovers textbook-likechemicalshift–functionalgrouprelationships(e.g.,aromaticprotonsat¹H6.5–8.5ppm, carbonylcarbonsat¹³C195–225ppm)directlyfromunassignedspectra,demonstratingthatthemodel adherestoestablishedspectroscopicprinciples.Withoutexplicitstructuralrulesencoded,MACROSexhibits emergentchemicalintuition,exemplifiedbyaring-firstpreferenceforrigidcyclicscaffoldsatarate exceeding40%.MACROSdemonstratedstrongzero-shotgeneralizationacrossdiversereal-world molecularclasses.Byfine-tuningtowarddomainorstructuralprior,MACROSyieldsspecialized performancefordistinctchemicalandbiologicalfields,includingsyntheticcompounds,complexnatural products,andhumanmetabolites.Atom-levelconfidencescoresalongreasoningtrajectoriessupported interpretablevisualizationofmolecularstructures,enablingintuitivehuman-AIcollaboration.MACROS establishesascalablefoundationforfullyautomatedspectroscopicstructureelucidationandcatalyzes acceleratedmoleculardiscoverytowardautonomousresearch. Results MACROSgeneralizesacrossdiversesourcesfordenovostructureelucidation Developingfrommulti-sourcedatasetsof>100millionspectrum-moleculepairs(ExtendedDataFigure 1a),MACROSdemonstratedhighaccuracyinmolecularparsingforarbitrarycombinationsofroutine spectra( 1 H, 13 C,HSQCNMRandIR)withoutretrievalandanymolecularpriors(e.g.formulas,reactants) (Figure1).Onrandomlyselectedtestset(n=560)fromUSPTOdataset 28 ,MACROSwasabletopredict accuratemolecularstructurewith0.88(95%CI0.86-0.89)fingerprintsimilarityandover0.95recallof commonfunctionalgroups(Figure1b-d).Specifically,weobservedhighimprovementofaccuracyafterthe closed-loopprocessofrethinkingandrefining(Figure1b).Rethinkingscoresstronglycorrelatewith ground-truthmolecularsimilarity.Thecombinedrethinkingscore(IRspectralsimilarityandclipped molecularconfidence)yieldsaconcordanceindex(c-index)of0.71,correctlyrankingthetruestructure withinthetop25%in52%ofcasesandwithinthetop10%in31%ofcases(Figure1c).MACROSwas robusttonoisyandincompletespectra(ExtendedDataFigure1b)andscalabletomoleculeswithlarger heavy-atomnumberthantrainingdata.Finally,themodelcanprovidebothatom-levelandmolecular-level confidencescoresforhumaninspectionandmodelrethinking(ExtendedDataFigure1c). ThehighaccuracyachievedontheUSPTO-derivedtestsetextendsconsistentlytospectrafromindependent sources(ExtendedDataFigure1d–g).OntheQM9Sdataset 29 ,MACROSachievedafingerprintsimilarity of0.55(95%CI0.53–0.57)fortheinitialdraftand0.71(95%CI0.69–0.74)afterclosed-looprefinement andselection.Comparableperformancewasobservedonthelarge-scalesimulatedSimPubChemdataset 30 (0.57(95%CI0.55–0.59)initial;0.67(95%CI0.65–0.69)refined).Allresultsreportedbelowwere obtainedunderstrictzero-shotconditions(notest-timefine-tuningormodality-specificadaptation; ExtendedDataFig.1e–f).OntheNMRMINDdataset 31 ,MACROSattainedafingerprintsimilarityof0.47 (95%CI0.45–0.49)fordraft,risingto0.56(95%CI0.54–0.58)afterclosed-looprefinement.Onthe NMRGymdataset 32 ,MACROSachievedafingerprintsimilarityof0.47(95%CI0.45–0.49)fordraft, risingto0.64(95%CI0.62–0.67)afterclosed-looprefinement.Theseresultsconfirmedthattheaccuracy androbustnessofMACROSgeneralizeeffectivelyfromtheprimarytrainingdistributiontodiversesources. Closed-loopmulti-agentreasoningenhancesstructureelucidationatinference MACROSachievedtest-timescalingandgreatlyimprovedtheaccuracyofmolecularstructureprediction, byintegratingahierarchicalmulti-agentarchitecturewithiterativeclosed-loopreasoning(Figure1e, ExtendedDataFigure2a).Specifically,wedevelopedaniterativeclosed-loopreasoningprocesswithina hierarchicalmulti-agentframework,whichdynamicallycoordinatedmultipleagentstoexecutespecialized tasks,includingdrafting,rethinking,andrefiningmolecularstructuresThesetaskswereseamlessly interconnectedthroughafeedbackloop:draftinggeneratedinitialSMILES-basedmolecularstructures, rethinkingevaluatedtheirvaliditybysimulatingspectraandcomparingthemtoexperimentaldata,and refiningoptimizedthesestructuresusingsimilarityscoresandmolecularconfidencemetrics.Thisiterative cycledrovetest-timescaling,wheredecipheringperformanceimprovedwithincreasingiterations.This multi-agentsystemwasstructuredbottom-up,progressingfrommodality-specificagentstofunction-specific subsystems(Figure2a). Self-supervisedpretrainingendowedmodality-specificagentswithdeepchemicalandspectralknowledge. Fourmodality-specificdecoder-onlytransformerswerepretrainedonover100millionmulti-sourcespectra vianext-tokenprediction,acquiringuniversalspectralgrammarwithoutlabels(Figure2b,ExtendedData Figure2b).Vibrationalspectraweretokenizedaspatches,NMRasdiscretepeaks,andmoleculesas SMILEScharacters(Figure2b).Pretrainedsolelyonspectralsequences,MACROSagentsspontaneously discoveredvibrationalmodes,carbonhybridizationstates,andprotonenvironments,revealinganAI-native chemicalintuitionemergentfromself-supervisedlearning.UnsupervisedclusteringofIRagentembeddings revealsstrongseparationbyfunctionalgroupchemistry,includingpatternsconsistentwithO–H,N–H,C–X, N=O,C=O,andC≡Cvibrations(Figure2c).Unsupervisedclusteringof¹³CNMRagentembeddings revealedseparationbycarbonhybridization(sp²,sp³;Figure2d-e).Similarly,the¹HNMRagentstratifies protonenvironmentswithhighchemicalhomogeneity(Figure2f-g). Supervisedfine-tuningassembledpretrainedagentsintotask-specificsubsystems(Figure2e-f).Inforward structureprediction,asummaryagentintegratedspectralencodingsandfeedsthemtoamolecularagentthat autoregressivelygeneratesinitialSMILESdrafts(draftingworkflow).Therefiningsubsystemextendedthe draftingworkflowbyincorporatingmolecularcontext,e.g.draftmolecules,fragments,orreaction precursors(Figure2e).Ablationofmodalitymaskingratiosrevealed60%asoptimal(ExtendedData Figure3a).Atthislevel,MACROSachievedpeakperformanceonbothcompleteandincompletespectral inputs.Token-levelconfidence(maximumprobabilitypergeneratedtoken)increasedprogressivelyoverthe firsteighttokensandstabilizedthereafter,consistentacrosslow-,medium-,andhigh-complexitymolecules (ExtendedDataFigure3b-c).ThispatternwashighlyconsistentwithSMILESsyntaxambiguity,as organicmoleculeshadmultiplevalidstartingpointsandnon-canonicalforms.MonteCarlohybridsearch wastherebydesignedtoexplorethediverseSMILESrepresentationspaceandoutperformsbeamandgreedy decoding(ExtendedDataFigure3d–g).Clipconfidence(productofpost-8th-tokenprobabilities) correlatedwithTanimotosimilarity(Pearsonr>0.4;ExtendedDataFigure3h,i),enabling high-confidencemolecular-levelfilteringofpredictedstructures. IntheMACROSrethinkingsubsystem,asummaryagentaggregatedencodedmoleculartokensandtransfers themtospectralagentsforspectrasimulation,followedbysimilarityscoringagainstexperimentaldata (Figure2d).Infraredsimilaritywascomputeddirectlyviacosinesimilarity.ForNMRmodalities,peaks werefirstalignedusingtheHungarianalgorithmonabsolutechemical-shiftdifferences,thennormalizedto yieldaboundedsimilarityscore(Materialsandmethods).Incontrasttoconventionalspectralsimulation methods,whichrelyoncomputationallyintensive3DconformergenerationandMerckmolecularforcefield (MMFF)optimization(typicallyrequiringsecondstominutespermolecule),Mol2Specpredictedspectra directlyfromSMILESstringsinasingleforwardpass.Furthermore,Mol2Speccanbetrainedonraw spectrum–structurepairswithoutrequiringpeakassignments,therebysubstantiallyreducingtheannotation burdenandfacilitatingdirectfine-tuningonexperimentallaboratorydata(SupplementaryNote5).The spectralsimulationagentsachievedaccuratereversepredictionwithlowerrors(n=1000):¹HNMR(MAE 0.16ppm,95%CI0.15–0.17ppm),¹³CNMR(1.34ppm,95%CI1.20–1.48ppm),HSQC(¹H:0.32ppm, 95%CI0.30–0.34ppm;¹³C:2.29ppm,95%CI2.08–2.50ppm),andIR(cosinesimilarity0.761,95%CI 0.754–0.768).OnIRprediction,thesevaluesoutperformedDetaNet 29 (0.719,95%CI0.712–0.726)and AttentiveFP 33 (0.720,95%CI0.713–0.727)(ExtendedDataFigures4,5a–c;representativepredictionsin ExtendedDataFigures4e–gand5d–f).Cliptoken-levelprobabilitieswereaggregatedintomolecular confidencescores,enablinghigh-confidencestructureselection(Figure2f).IRsimilarityscoreandclip confidencebothcorrelatedstronglywithtruemolecularsimilarity(ExtendedDataFigure1b).This closed-loopintegrationofdrafting,rethinking,andrefiningensuredrobust,experimentallyalignedstructure prediction. MACROSlearnschemicalintuitionandinterpretablereasoning MACROSexhibitedinterpretabledecision-makingacrosssystemandsubmodulelevels.Althoughno explicitmodalitybiaseswereintroducedduringtraining,themodelpreferentiallyusedNMRspectrafor forwarddraftingandIRspectraforrethinkingvalidation.NMRnoiseimpairedbothdraftgenerationand finalpredictionaccuracy,whereasIRnoiseselectivelydegradesperformanceintherethinkingvalidation phase(ExtendedFigure1a).IRsimilarityscorescorrelatedstronglywithTanimotosimilarity,whileNMR scoresremainedconsistentlyhigh(ExtendedFigure1b).WehypothesizedthatMACROSleveragesNMR chemicalshiftinformationtoproducerobustdraftpredictionswithhighNMR-basedstructuralsimilarity, whileIRspectra,alongsideclip-confidencemetrics,facilitatevalidationintherethinkingphase.Modality ablationsconfirmedthisasymmetry:excludingIRcausesminimalperformancedrop,whereasexcluding NMRseverelyimpairsaccuracy(Figure1b). Tounderstandtheinternalreasoningprocessofthemodel,weanalyzedattentionmapsatthefinalSMILES generationstep,aggregatingscoresfortokenscorrespondingto13functionalgroups(Figure3a,Extended DataFigure6).MACROSrecoveredspectroscopicconsensus,e.g.,ketones(¹H:2–3ppm;¹³C:195–225 ppm),aldehydes(¹H:9–10ppm;¹³C:190–205ppm),withoutpeak-levelsupervision(Figures3b–c).The samemechanisminrethinkingmapspredictedspectrabacktomolecularsubstructures,yieldingconsistent, interpretablealignments(ExtendedFigures4a,c).Itwasworthnotingthatthe chemical-shift/functional-grouprelationshipscompiledinthisstudyoriginatedfromdecadesofempirical practice,andtheirshiftaccuracyremainedcontingentonfactorssuchasspectrometerconfiguration,choice ofdeuteratedsolvent,andmeasurementconditions.Consequently,theinterpretationprecisionofMACROS inassigningfunctional-group-specificchemicalshiftswasfundamentallyrelatedtotheaccuracyand consistencyoftheunderlyingdatabases.Thisclosecorrespondencefurtherconfirmedthereliabilityofthe datasource.ThesefindingsdemonstratedthatMACROSacquiresexpert-likechemicalintuitionthrough self-supervisedspectralmodeling,enablingreliable,automatedstructureelucidation.Crucially,noexplicit correspondencebetweenindividualspectralpeaksandmolecularsubstructureswaseverprovidedduring training.ThesolesupervisorysignalconsistedofrawmultimodalspectrapairedwithcorrectSMILES. Nevertheless,followinglarge-scaleself-supervisedpretrainingonunpairedspectraldataandsubsequent end-to-endsupervisedfine-tuningonannotatedpairs,MACROSautonomouslyrecoveredtheentire canonicalchemical-shift–functional-groupcorrelationtablewithtextbookfidelity. Predictivepatternsexhibitedaconsistentandchemicallyintuitivestrategyofstructureelucidation.Across diversedatasets(QM9,NMRMIND,USPTO,andSimPubchem),MACROSconsistentlyinitiatedSMILES sequenceswithringsystems(>40%,Figure3d).Thisring-firstpreferencedirectlymirroredtheexpert heuristicofanchoringinterpretationonrigidcyclicscaffoldsbeforeappendingsubstituents,despitethe autoregressivedecoderreceivingnoexplicitstructuralorsyntacticpriorityduringtraining 16,17 .During trainingphase,usageofnon-canonicalSMILEScanenhancethepriority(Figure3d).UMAPvisualizations ofthelearned¹Hand¹³CNMRchemical-shiftembeddingsfromthepretrainedNMRagentsprovideddirect mechanisticsupportforthisintuition(Figure3d).Trajectoriesexhibitedpronounceddirectionalchanges nearthediagnosticallycriticalregionsof 13 C ≈ 120ppmand 1 H ≈ 7ppm,regionsclassicallyassociatedwith aromaticsystems(ExtendedDataFigure7).Thispatternemergedalreadyduringself-supervised pretrainingonunpairedspectraandwassubsequentlyrefinedduringsupervisedfine-tuning,indicatingthat themodelautonomouslyidentifiesaromaticmotifsashighlysalientfeatures.Theobservedring-first generationstrategythusreflectednotonlytheacquisitionofcorespectroscopic–structuralcorrelationsbut alsotheemergenceofanexpert-likehierarchicalreasoningprocessthatsystematicallyprivilegesrigidcyclic fragmentsasprimaryinterpretiveanchors. MACROSgeneralizestoreal-worldspectraandaugmentshumanexperts Toevaluatezero-shotperformanceonauthenticexperimentalspectra,weassembledacuratedreal-world evaluationsetof1,701moleculesfrommultipleindependentsources,includingrandomlysampledentries fromNMR-BANK(n=540),literature-reportedspectrafromrecentJ.Am.Chem.Soc.publications (2019–2024;n=450),challengingnaturalproductswithHSQCdata(n=398),andasubsetfromthe NMR-Expdatabase²⁴(n=313).Allstructureswereverifiedindependentlyandconfirmedtobeabsentfrom themodel’strainingandfinetuningcorpus. Validationsonnaturalproducts.Arandomlyselectedsubsetofthenatural-productcohortfromNP-MRD wasusedforevaluation(n=380uniquecompounds,eachwithexperimental¹Hand¹³CNMRdata;mean heavyatomcount31.2,95%CI:29.8–32.7,range8–112;meanmolecularweight438Da,95%CI:417–458, range112–1570).Onthissubset,thezero-shotpredictionachievedaninitialdraftfingerprintsimilarityof 0.559(95%CI:0.535–0.585).Thisincreasedmarkedlyto0.625(95%CI:0.600–0.651)afterclosed-loop reasoning(Figure4a).Targetedfine-tuningonnaturalproductsfurtherelevatedperformanceto0.799(95% CI:0.778–0.820),withanadditionalimprovementto0.843(95%CI:0.822–0.864)followingreasoning (Figure4b).Intotal,54.2%ofthemoleculeswerecorrectlyresolvedfromthespectra.Zero-shotpredictions successfullycapturedbroadstructuraldiversity,includingterpenoids,prenylatedflavones,flavonoid glycosides,andlignanglycosides(Figure4i,ExtendedDataFigure9).Moreover,MACROSdemonstrated promiseforstereoisomerdiscrimination,encompassingenantiomersandcis–transisomers(ExtendedData Figure10).Notably,MACROSinitiallytrainedongenericorganiccompoundsachievedhigh-fidelity structuralassignmentofdiversenaturalproductsafteronlyminimalfine-tuningwithpartialdomain-specific knowledge,highlightingefficienttransferlearningthatavoidstheneedforextensivespecializedtraining. Validationsonliterature-reportedspectra.Arandomlyselectedsubsetofthenatural-productcohortfrom NMRBANKcomprised544uniquesampleswithexperimental¹Hand¹³CNMRdata(meanheavyatom count:25.6,95%CI24.9–26.4,range3–62;meanmolecularweight:369Da,95%CI358–379,range 69–970Da).Onthissubset,zero-shotpredictionsyieldedaTanimotosimilarityof0.478(95%CI 0.460-0.497)(Figure4c,ExtendedDataFigure11).Fine-tuningondataofchemicalcompoundsraised performanceto0.742(95%CI0.705-0.779),demonstratingrapidadaptationtoliterature-derivedspectral distributions.Arandomlyselectedsubsetofthenatural-productcohortfromNMRExpcomprised313 uniquesampleswithexperimental¹Hand¹³CNMRdata(meanheavyatomcount:24.9,95%CI23.8–25.9, range9–101;meanmolecularweight:355Da,95%CI340–369,range119–1414Da).Onthissubset, zero-shotperformanceyieldedaninitialdraftsimilarityof0.431(95%CI0.405,0.457),improvingto0.479 (95%CI0.452-0.507)afterclosed-looprefinement.Afterfine-tuning,draftsimilarityincreasedto0.683 (95%CI0.649-0.711),reaching0.743(95%CI0.714-0.772)afterreasoning.Onin-houseacquiredcrude reactionproductsusing¹³CNMRalone,MACROSattainedamediansimilarityof0.456(95%CI 0.440–0.472)zero-shot(SupplementaryNote8),withrobustpredictiondespiteseveresignaloverlapand impurities.TheseresultsestablishthatMACROSdeliversrobustzero-shotcapabilityacrossdiverse real-worlddataandcanberapidlyspecializedtolaboratory-specificchemicalspaceswithminimal additionalfinetuning. Validationsonmetabolitemolecules.ThemetaboliccohortfromBMRBcomprised409uniquesamples (meanheavyatomcount:18.7,95%CI17.5–19.9,range2–76;meanmolecularweight:268Da,95%CI 251–284,range32–1062Da).Onthiscohort,usingonly¹Hand¹³CNMRdata,thezero-shotprediction achievedaninitialdraftsimilarityof0.519(95%CI:0.491-0.546).Itimprovedmarkedlyto0.589(95%CI: 0.560-0.618)followingclosed-loopreasoning(Figure4e).Afterfine-tuning,performancerosesubstantially to0.742(95%CI:0.713,0.770),andfurtherincreasiedto0.816(95%CI:0.790,0.843)afterreasoning (Figure4f).Intotal,68.9%ofthemoleculeswerecorrectlyresolvedfromthemetabolicspectra.The zero-shotpredictionssuccessfullyencompassedabroaddiversityofmetabolicclasses,includingamino acids,alkaloids,nucleotides,andglycosides(Figure4j).Despitedecadesofmetabolomicsresearch,known metabolitesrepresentonlyatinyfractionofpossiblestructuresinthemammalianmetabolome,leaving extensiveunidentifiedsignalsasmetabolic“darkmatter”.MACROSpotentiallyenablesproactivediscovery ofunknownmetabolitesfromNMRdata,providingacomplementaryalternativetoDeepMet'spassive MS-basedvalidationstrategyandsupportingsystematicexplorationofmetabolicdarkmatterin NMR-centricdatasets 34,35 . Validationsonautonomousreactionannotations.Spectralanalysisofreactionproductsremainsa cornerstoneofmechanisticelucidationandreal-timereactionmonitoringinsyntheticchemistry.MACROS introducesreactantstructuresasmolecularpromptstothePromptSpec2Molsubsystem,enablingsubstantial accuracyimprovementswithoutanyexplicittrainingonreactionmechanismsoroutcomeprediction.Onthe USPTObenchmark(n=360),performancewithreactantpromptingyieldedanaverageTFPS0.844(95% CI0.825–0.863)fortheinitialdraft,risingto0.889(95%CI0.874–0.904)afterrethinking.MACROSwas furtherevaluatedon450real-worldcrudereactionproductscuratedfromrecentJ.Am.Chem.Soc. publications(2019–2024).Intotal,MACROSsuccessfullyannotated44%ofchemicalreactionsfrom real-worldcasesusingzero-shotprediction.Usingonly¹³CNMRspectracombinedwithreactantmolecular prompts,thezero-shotpredictionachievedadraftTFPSof0.556(95%CI0.533–0.580),increasingto 0.5886(95%CI0.566–0.612)afterrethinking(Figure4k).Using 1 Hand¹³CNMRspectracombinedwith reactantmolecularprompts,thezero-shotpredictionachievedadraftTFPSof0.642(95%CI0.620–0.664), increasingto0.677(95%CI0.656–0.698)afterrethinking.SelectedcasesweredemonstratedinExtended DataFigure12andSupplementaryNote9.Thesedemonstratethatcontextualpriorinformationfrom reactantsalonecaneffectivelyconstrainthecombinatorialsearchspaceandsteerpredictionstoward chemicallyplausibleproducts. Human–AIcollaborationinreal-worldspectralanalysis.MACROSenablesefficientcharacterizationof complexchemicalsystemsinautomatedlaboratoryworkflowsandsupportschemistsinpreciseproduct elucidation.InacuratedPubChemdataset(chemicalengineering,environmental,pharmaceuticalmolecules), MACROSaloneachieves>90%structuralaccuracyin<1minutespermoleculeonasingle24GBGPU.A blindedhuman-expertstudywasconductedwiththreeorganicspectroscopists(MaterialsandMethods). Eachexpertfirstelucidatedstructuresfromrawmultimodalspectrawithoutassistance(about1hper molecule,0.102TFPS).Afterafour-weekwash-outperiod,thesamespectrawerere-analyzedwith MACROSassistance(fullclosed-loopmodewithconfidencescoresdisplayed).Meantimewasreducedto about10minpermoleculeandTFPSincreasedto0.807(Figure4l).Closed-loopreasoningtrajectories furtherrevealsystematicrefinement(Figure4l).Initialdraftsexhibithighstructuraldiversity;eachiteration pruneslow-confidencecandidates,monotonicallyincreasingmolecularconfidence,reversespectral similarity,andground-truthTFPS.Thisconvergenceismostpronouncedforcomplexnaturalproductsand macrocycles,whereaccuracygainsexceed60%withinfivecycles.Theseresultsdemonstratethat MACROSnotonlyenablesfullyautonomousspectralanalysisatexpertlevel,butalsodramatically acceleratesandaugmentshumanexpertise,establishingapracticalsymbiosisbetweenchemistsandAIin modernautomatedlaboratories. Discussion Spectral-to-structureelucidationremainsoneofthemostcognitivelydemandingtasksinorganicchemistry, requiringiterativehypothesistestingacrosscomplementaryyetheterogeneousmodalities.Chemicalexperts requireconsiderablebreadthofthoughtwheninterpretingspectra.Typically,thisinvolvesdissectingspectral peaksandthecomplexrelationallogicofmolecularstructuresfrommultipleperspectives,ultimately identifyingamolecularstructurethatcorrespondstoallobservedpeaks.Inthisprocess,thepresenceof shiftedorinterferingpeaksintroducedbycomplexsystemsfurtheramplifiesthedifficultyofthis“logical puzzle”.Elucidatingmolecularstructuresfromspectroscopicdataisacomplex,systematictaskthat necessitatestheintegrationofmultidisciplinaryknowledge.Whilethestructuralinformationaffordedby routinespectroscopicmethodsisinsufficientforunambiguous,fullmolecularelucidation,especiallyfor highlycomplexspecies,theseapproachesremainindispensableforexploringthevastexpanseofuncharted chemicalspace.Forautomatedmolecularstructureelucidation,thekeyistoestablishascalablefoundation withcorrectspectroscopicprinciple,whichcanbereadilyextendedtonovelmodalitiesandspecialized fields. Here,weproposeMACROS,ahierarchicalmulti-agentframeworkthatemulatestheiterativereasoningof expertspectroscopists.MACROSachievesrobustperformanceonarbitrarycombinationsofroutinespectra andexhibitsrobustzero-shotgeneralizationtodiversereal-worldexperimentalspectra,includingchemical reactions,naturalproductsandmetabolites.Onasingleconsumer-gradeGPU(24GB),MACROS processes>20moleculesperminuteinquickmodeand>1moleculesperminuteinfullclosed-loopmode. MACROSstreamlinesexpert-drivenstructureelucidation,deliveringsixfoldfasteranalysisand substantiallyenhancedaccuracy.Throughlarge-scalepretrainingandclosed-loopreasoning,MACROS establishesascalablefoundationforgeneralizablemolecularstructureelucidationfrommultimodalspectra, withthepotentialtosupportself-drivinglaboratoriesformoleculardiscoveryandvalidation. Thisadvantagederivesinpartfromdeliberatemodality-specificpretrainingonnativedatarepresentations. Ratherthanforcingheterogeneousspectroscopicdataintogenerictextorimagerepresentations,wepretrain lightweightagentsdirectlyontheirphysicallynativeformatsthatpeaklistsforNMR,continuous waveformsforIR,andchemistry-specifictokenizationforSMILES.Thispreservestheintrinsicinductive biasesthatareinevitablylostinmonolithicLLM/VLMbackbones.Theresultingagentsareintegrated throughlightweightprojectionlayersintotask-specificsubsystems,achievingmarkedlyhigheraccuracy thancontemporaryLLM-basedapproacheswhilerequiringorders-of-magnitudefewerparametersfor adaptation.Leveragingmulti-sourcelarge-scaleself-supervisedpretraining,individualagentswithin MACROSacquirebroadlygeneralizablespectroscopicandmolecularknowledge.Thispositionsthemas promisingcandidatesforservingasfoundationalmodelstailoredtoNMR,IR,andmolecularrepresentation learningmorebroadly,therebyenablingrobustmultimodalmolecularnetworking 42,43 .Thisphysically grounded,reusableagentparadigmoffersapractical,resource-efficientalternativetotheprevailingstrategy forscientificdataanalysis. Equallyimportantistheclosed-loopreasoningframework,whichenablestest-timescalingwithoutgradient updates.Unlikeapproachesthatrelyonreinforcementlearningtodiscoverelucidationstrategies autonomously,MACROSemploysanexplicitmulti-agentworkflowwhoseinteractionpatternsare deliberatelyderivedfromexpertorganicchemists’systematicreasoning.Therethinkingstageprovidesa chemistry-nativeconsistencycheckfromanorthogonalphysicalperspective:candidatestructuresare re-encodedandpassedthroughreversespectralagents,yieldingpredictedNMRchemicalshiftsandIRband patternsthatarescoredagainstrawexperimentaldata.Thisstepeffectivelymeasuresagreementbetween hypothesisandobservationusingphysicallymeaningfuldistancesratherthanlearnedproxies.Therefining stagefurtherextendsthebasedraftingmodelthroughtargetedfine-tuningonprefix-conditionedinputs.By feedingspectroscopicallyvalidatedfragments(reactantsortopdrafts)asprefixes,thegenerationisrestricted tothelow-dimensionaldetailsubspacearoundchemicallycorrectanchors,transformingexponentialde novosearchintoefficientlocaloptimizationwhilepreservingestablishedstructuralintegrity.Althoughthe currentagenttopologyisdeliberatelyminimalandfullyhand-craftedforchemicaltransparency,italready deliversexpert-levelspectralparsingacrosshighlydiversereal-worlddatasets.Morerefinedworkflow designinfutureiterationscouldallowMACROStotackleanevenbroaderrangeofmoleculardiscovery taskswiththesametransparent,expert-inspiredclosed-loopprinciple. InterpretabilityanalysesconfirmthatMACROSfaithfullyrecapitulatesthecanonicalcorrespondence betweenspectralfeaturesandmolecularsubstructures.Modalityablationandnoise-perturbationstudies revealalearneddiagnostichierarchythatmirrorsexpertpractice.Attentionscoresfurtherdemonstratethat MACROSautonomouslyreconstructsthecanonicalmappingbetweenNMRchemicalshiftsandmolecular substructures.Remarkably,MACROShasnevershownanyexplicitcorrespondencebetweenchemicalshifts andmolecularstructures,yetitrediscoveredtheseempirical,textbook-levelcorrelationspurelyfromraw paireddata,illustratingascalableparadigmforextractinglatentnaturallawsdirectlyfromunannotated experimentalrecords.Thisemergentrediscoveryoftextbookspectroscopic–structuralcorrelations, traditionallydistilledoverdecadesofchemicalexperimentation,illustratesabroaderprinciple.This capabilityestablishesatransferableframeworktominelatentprinciplesfromunannotatedexperimentaldata, withbroadpotentialfordenovomulti-modaldatainterpretation.MACROSdisplaysemergentchemical intuitionthrougharing-firststructuralassemblystrategy,preferentiallyinitiatingSMILESgenerationwith rigidcyclicscaffoldsacrossalltesteddatasets.Thisdata-driven,AI-nativeunderstandingofstructure elucidationalignswiththeheuristicpreferencesrefinedbyexpertchemiststhroughextensivepractical experience.Collectively,thesefindingsconfirmthatMACROShasinitiallydevelopedanexpert-like reasoningframeworkforstructureelucidation. MACROSrepresentsasignificantstepforwardinautonomouschemicalresearch,offeringanefficientand stablemethodformolecularstructureelucidationthroughitsmulti-agent,closed-looparchitecture.By decouplingmodalitiesandleveragingdomain-specificknowledge,MACROSoutperformsgeneral-purpose LLMswhilemaintainingcomputationalefficiency.Itsexplicitmulti-agentdesignandtest-timescaling capabilitiesprovideaflexibleframeworkforaddressingcomplexscientificchallenges.Futureevolution, includingbroaderdatasets,improvedtrainingandinferencestrategies,andtighterintegrationwith automatedexperimentation,shouldfurtherexpanditsutilityinchemicalsynthesis,naturalproductdiscovery andautomatedlaboratories.ThisworkunderscoresthepotentialofspecializedAIsystemstotransform scientificdiscovery,offeringablueprintforfutureadvancementsincomputationalchemistryandbeyond. Reference 1.Jia,Y.etal.Robot-assistedmappingofchemicalreactionhyperspacesandnetworks.Nature2025645:8082 645,922–931(2025). 2.Zhang,Z.etal.Amultimodalroboticplatformformulti-elementelectrocatalystdiscovery.Nature20251–3 (2025)doi:10.1038/S41586-025-09640-5. 3.Boiko,D.A.,MacKnight,R.,Kline,B.&Gomes,G.Autonomouschemicalresearchwithlargelanguage models.Nature624,570–578(2023). 4.Qin,Y.etal.Amulti-scalemapofcellstructurefusingproteinimagesandinteractions.Nature2021 600:7889600,536–542(2021). 5.Beck,M.,Covino,R.,Hänelt,I.&Müller-McNicoll,M.Understandingthecell:Futureviewsofstructural biology.Cell187,545–562(2024). 6.Burns,D.C.,Mazzola,E.P.&Reynolds,W.F.Theroleofcomputer-assistedstructureelucidation(CASE) programsinthestructureelucidationofcomplexnaturalproducts.Nat.Prod.Rep.36,919–933(2019). 7.Liu,Y.etal.UnequivocaldeterminationofcomplexmolecularstructuresusinganisofropicNMR measurements.Science(1979).356,(2017). 8.Sindelar,M.&Patti,G.J.ChemicalDiscoveryintheEraofMetabolomics.J.Am.Chem.Soc.142, 9097–9105(2020). 9.Jeppesen,M.J.&Powers,R.Multiplatformuntargetedmetabolomics.MagneticResonanceinChemistry61, 628–653(2023). 10.Segler,M.H.S.&Waller,M.P.ModellingChemicalReasoningtoPredictandInventReactions.Chemistry –AEuropeanJournal23,6118–6128(2017). 11.Guo,D.etal.DeepSeek-R1incentivizesreasoninginLLMsthroughreinforcementlearning.Nature2025 645:8081645,633–638(2025). 12.Barone,V.etal.Computationalmolecularspectroscopy.NatureReviewsMethodsPrimers20211:11,38- (2021). 13.Vignoli,A.,Cacciatore,S.&Tenori,L.DerivingthreeonedimensionalNMRspectrafromasingle experimentthroughmachinelearning.NatureCommunications202516:116,10159-(2025). 14.Buchanan,C.etal.UnidecNMR:automaticpeakdetectionforNMRspectrain1-4dimensions.Nature Communications202516:116,449-(2025). 15.Elyashberg,M.IdentificationandstructureelucidationbyNMRspectroscopy.TrACTrendsinAnalytical Chemistry69,88–97(2015). 16.Wishart,D.S.etal.NP-MRD:theNaturalProductsMagneticResonanceDatabase.NucleicAcidsRes.50, D665–D677(2022). 17.Williams,A.J..,Martin,G.E..&Rovnyak,David.ModernNMRapproachestothestructureelucidationof naturalproducts.Volume2,Dataacquisitionandapplicationstocompoundclasses. https://books.google.com/books/about/Modern_NMR_Approaches_to_the_Structure_E.html?hl=zh-CN&id =DHMoDwAAQBAJ(2017). 18.Yao,L.etal.ConditionalMolecularGenerationNetEnablesAutomatedStructureElucidationBasedon13C NMRSpectraandPriorKnowledge.Anal.Chem.95,5393–5401(2023). 19.Yu,N.etal.Pseudodata-basedmolecularstructuregeneratortorevealunknownchemicals.NatureMachine Intelligence20257:117,1879–1887(2025). 20.Swanson,K.,Wu,W.,Bulaong,N.L.,Pak,J.E.&Zou,J.TheVirtualLabofAIagentsdesignsnew SARS-CoV-2nanobodies.Nature2025646:8085646,716–723(2025). 21.MacKnight,R.etal.Rethinkingchemicalresearchintheageoflargelanguagemodels.Nat.Comput.Sci.5, 715–726(2025). 22.M.Bran,A.etal.Augmentinglargelanguagemodelswithchemistrytools.Nat.Mach.Intell.6,525–535 (2024). 23.Bisson,J.etal.DisseminationofOriginalNMRDataEnhancesReproducibilityandIntegrityinChemical Research.Nat.Prod.Rep.33,1028(2016). 24.Karunanithy,G.,Shukla,V.K.&Hansen,D.F.Solution-statemethylNMRspectroscopyoflarge non-deuteratedproteinsenabledbydeepneuralnetworks.NatureCommunications202415:115,5073- (2024). 25.Elyashberg,M.etal.Computer-assistedmethodsformolecularstructureelucidation:Realizinga spectroscopist’sdream.J.Cheminform.1,1–26(2009). 26.Dührkop,K.etal.SIRIUS4:arapidtoolforturningtandemmassspectraintometabolitestructure information.NatureMethods201916:416,299–302(2019). 27.Xing,S.,Shen,S.,Xu,B.,Li,X.&Huan,T.BUDDY:molecularformuladiscoveryviabottom-upMS/MS interrogation.Nat.Methods20,881–890(2023). 28.Alberts,M.,Schilter,O.,Zipoli,F.,Hartrampf,N.&Laino,T.UnravelingMolecularStructure:A MultimodalSpectroscopicDatasetforChemistry.https://arxiv.org/pdf/2407.17492v2(2024). 29.Zou,Z.etal.Adeeplearningmodelforpredictingselectedorganicmolecularspectra.Nature ComputationalScience20233:113,957–964(2023). 30.Jin,Y.etal.NMR-Solver:AutomatedStructureElucidationviaLarge-ScaleSpectralMatchingand Physics-GuidedFragmentOptimization.https://arxiv.org/pdf/2509.00640v1(2025). 31.Xue,X.etal.NMRMind:ATransformer-BasedModelEnablingtheElucidationfromMultidimensional NMRtoStructures.Anal.Chem.97,22603–22614(2025). 32.Fang,Z.etal.NMRGym:AComprehensiveBenchmarkforNuclearMagneticResonanceBasedMolecular StructureElucidation.ProceedingsofProceedingsofthe32ndACMSIGKDDConferenceonKnowledge DiscoveryandDataMining(KDD’26)1,(2026). 33.Xiong,Z.etal.PushingtheBoundariesofMolecularRepresentationforDrugDiscoverywiththeGraph AttentionMechanism.J.Med.Chem.63,8749–8760(2019). 34.Qiang,H.etal.Languagemodel-guidedanticipationanddiscoveryofmammalianmetabolites.Nature2026 12,1–10(2026). 35.DaSilva,R.R.,Dorrestein,P.C.&Quinn,R.A.Illuminatingthedarkmatterinmetabolomics.Proc.Natl. Acad.Sci.U.S.A.112,12549–12550(2015). 36.Jonas,E.,Kuhn,S.&Schlörer,N.PredictionofchemicalshiftinNMR:Areview.MagneticResonancein Chemistry60,1021–1031(2022). 37.Han,H.&Choi,S.TransferLearningfromSimulationtoExperimentalData:NMRChemicalShift Predictions.J.Phys.Chem.Lett.12,3662–3668(2021). 38.Bross-Walch,N.,Kühn,T.,Moskau,D.&Zerbe,O.Strategiesandtoolsforstructuredeterminationof naturalproductsusingmodernmethodsofNMRspectroscopy.Chem.Biodivers.2,147–177(2005). 39.Jiang,Z.etal.Mechanismofboricacidextractionbytrioctylamineandtartaricacid.Sep.Purif.Technol. 331,125597(2024). 40.Yu,N.etal.Pseudodata-basedmolecularstructuregeneratortorevealunknownchemicals.NatureMachine Intelligence20257:117,1879–1887(2025). 41.MarcosAnghinoni,J.,Irum,UrRashid,H.,JoãoLenardão,E.&SantosSilva,M.31PNuclearMagnetic ResonanceSpectroscopyforMonitoringOrganicReactionsandOrganicCompounds.ChemicalRecord24, e202400132(2024). 42.Bushuiev,R.etal.Self-supervisedlearningofmolecularrepresentationsfrommillionsoftandemmass spectrausingDreaMS.NatureBiotechnology20251–11(2025)doi:10.1038/s41587-025-02663-3. 43.Wang,M.etal.SharingandcommunitycurationofmassspectrometrydatawithGlobalNaturalProducts SocialMolecularNetworking.NatureBiotechnology201634:834,828–837(2016). 44.Wang,Q.etal.NMRExtractor:leveraginglargelanguagemodelstoconstructanexperimentalNMR databasefromopen-sourcescientificpublications.Chem.Sci.16,11548–11558(2025). 45.Steinbeck,C.&Kuhn,S.NMRShiftDB–compoundidentificationandstructureelucidationsupportthrough afreecommunity-builtwebdatabase.Phytochemistry65,2711–2717(2004). 46.Hoch,J.C.etal.BiologicalMagneticResonanceDataBank.NucleicAcidsRes.51,D368–D376(2023). 47.Hu,F.,Chen,M.S.,Rotskoff,G.M.,Kanan,M.W.&Markland,T.E.AccurateandEfficientStructure ElucidationfromRoutineOne-DimensionalNMRSpectraUsingMultitaskMachineLearning.ACSCent. Sci.10,2162–2170(2024). 48.Hu,Y.etal.Revvingup13CNMRshieldingpredictionsacrosschemicalspace:benchmarksfor atoms-in-moleculeskernelmachinelearningwithnewdatafor134kilomolecules.Mach.Learn.Sci. Technol.2,035010(2021). 49.Su,J.etal.RoFormer:EnhancedtransformerwithRotaryPositionEmbedding.Neurocomputing568, (2024). 50.Silverstein,R.W.&Bassler,G.C.Spectrometricidentificationoforganiccompounds.J.Chem.Educ.39, 546–553(1962). 51.Pretsch,E.,Bühlmann,P.&Badertscher,M.StructureDeterminationofOrganicCompoundsTablesof SpectralDataFifthEdition. Figure1.Multi-agentsystemsforclosed-loopmulti-modalspectralreasoning. a.PerformancebenchmarkontheUSPTO-derivedtestset(n=200compounds).Fingerprintsimilarity(± 95%confidenceinterval)ofpredictedversusground-truthstructuresisshownforMACROSandcurrent state-of-the-artmethods,includingdedicatedspectrum-to-structuremodelsandprompt-engineeredlarge languagemodels(LLMs).b.Accuracyscalingwithreasoningfordifferentmultimodalinputcombinations (¹HNMR,¹³CNMR,HSQC,IR).Performanceoftheinitialdraft(iteration0)andsubsequentrefinement cyclesisreported.c.Concordancebetweencandidaterankingbyrethinkingscores(IRspectralsimilarity, clippedmolecularconfidence,andtheirweightedcombination)andrankingbyground-truthfingerprint similarityacrossthefulltestset.d.Functional-grouprecallofinitialdraftsversusfinalclosed-loop predictionsfor13commonorganicfunctionalgroups.eClosed-loopreasoningworkflowfortest-time scalinginmultimodalspectralstructureelucidation.Theiterativepipelineconsistsof(i)draftcandidate generationusinghybriddecodinganduncertainty-awareconfidenceestimation,(i)spectralsimulationof candidatesfollowedbymulti-criteriaranking,and(i)guidedrefinementovermultiplecyclesusing top-rankedstructuresasprefixes.Optionalextendedinformationmaybeincorporatedintotheworkflow, suchasreactants,molecularformulaederivedfromMS,ordatafromelementalanalysis. Figure2.MACROSdevelopmentviaself-supervisedpre-trainingandfine-tuning. a.Hierarchicalmulti-agentassembly.Modality-specificagents(¹HNMR,¹³CNMR,HSQC,IR,and molecule)arefirstpretrainedindependently,thenjointlyfine-tunedintotask-orientedsubsystems(forward Spec2MolandreverseMol2Spec),andfinallyintegratedintoclosed-loopworkflowsthatenableiterative drafting,rethinking,andrefinement.b.Self-supervisedautoregressivepretrainingparadigm.Each modality-specificagentispretrainedon>10⁷unpairedexamplesofitsnativedatatype(peaklistsforNMR, fullwaveformsforIR,SMILESstringsformolecules)usinganext-tokenpredictionobjective.c-d. Emergentchemicalknowledgefrompretraining.Withoutexplicitsupervision,theIRagentlearnscanonical functional-groupwavenumberranges(c),the¹³CNMRagentdistinguishessp,sp²,sp³,andquaternary carbons,andthe¹HNMRagentresolvesprotonenvironments(representativeattention-derivedcorrelations shown)(d).e.Forwarddecipheringsubsystem(Spec2Mol).Spectralagentsencodemultimodalinputs;a summaryagentintegratesfeaturesandfeedsthebasemolecularagentfordenovoSMILESgeneration. Optionaldraft-moleculeprefixescanbesuppliedforguidedrefinement.f.Reversespectralsimulation subsystem(Mol2Spec).Thelargemolecularagentencodescandidatestructures;spectralagents independentlypredictcorresponding¹HNMR,¹³CNMR,HSQC,andIRdata,enablingsimilarity-based candidatescoringintherethinkingphase. Figure3.Interpretationofspectralparsingprocessatmulti-scale. a.Attention-basedinterpretability.Causalattentionweightsfromthefinalthreetransformerlayersofthe molecularagentareprunedtothetop10%perquerytoken,averagedacrossheads,andaggregatedby predefinedfunctionalgroups.HeatmapsshowattentionfromSMILEStokens(x-axis)tospectralpeaks (y-axis). b.Learned¹H(i)and¹³C(i)NMRchemical-shift–functional-groupcorrelations.Heatmapshows aggregatedattentionfromfunctional-groupatomsto¹HNMRpeaktokens.Greyshadedregionsindicate empiricalrangescompiledfromdecadesofexperimentalNMRliteratures 50,51 .Learnedcorrelationshighly overlapwithempiricalranges c.UMAPprojectionoflearned¹H(i)and¹³C(i)NMRchemical-shiftembeddingsfromthepretrainedNMR agents.Trajectoriesshowpronounceddirectionalchangesnear¹³C≈120ppmand¹H≈7ppm,regions classicallyassociatedwitharomaticsystems.Barcharts(i)confirmthatMACROSconsistentlyinitiates correcttop-1SMILESpredictionswithringsystemsacrossdatasets. Figure4.Real-worldvalidationsandapplications. a-h.Real-worldzero-shotandfine-tunedperformanceofthemoleculargenerationmodelevaluatedon diverseNMR-basedvalidationsources(a-b.naturalproducts,c-fchemicalreactionsandg-hmetabolites) usingmultiplespectraldatacombinations(¹HNMRonly,¹³CNMRonly,andcombined¹H/¹³CNMR). i-j.Representativeexamplesofzero-shotparsingfornaturalproductsfrom(i)NP-MRDand(j)BMRB. k.Zero-shotandfine-tunedannotationperformanceofchemicalreactions. l.Schematicdiagramofreaderstudies:chemistsvs.chemists+MACROSparsingdistinctspectracollected fromPubChem.ThereasoningtrajectoryinMACROSsignificantlyimprovesbothparsingaccuracyand efficiency.