Source-linked AI summary
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
TL;DR
Scalable, fully automated structure elucidation from routine spectra remains an outstanding challenge despite its foundational role in discovery. MACROS uses a multi-agent closed loop to emulate expert hypothesis testing. It generalizes across routine spectra and diverse real-world samples, with sixfold faster analysis and substantially enhanced accuracy.
Problem
Scalable, fully automated de novo structure elucidation from spectroscopic data remains an outstanding challenge despite its foundational role in chemical, biological, and biomedical discovery.
Source: Introduction · p. 2Method
MACROS is a purpose-built hierarchical multi-agent system that uses closed-loop expert-like hypothesis testing to elucidate structures from routine multimodal spectra.
Source: Introduction · p. 3Results
Sixfold faster analysis with substantially enhanced accuracy accompanies robust zero-shot generalization across arbitrary combinations of routine spectra and diverse real-world experimental samples.
Source: Discussion · pp. 8–9Takeaways & Limitations
The work establishes a scalable foundation for generalizable molecular structure elucidation from multimodal spectra.
Source: Discussion · p. 9The reported results were obtained under strict zero-shot conditions, without test-time fine-tuning or modality-specific adaptation.
Source: MACROS generalizes across diverse… · p. 4
Abstract
from arXiv · showhide
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.
Introduction
The introduction frames de novo elucidation of complex molecules from routine spectra as a persistent challenge requiring interpretable, hypothesis-driven integration of complementary modalities. It presents MACROS as a purpose-built multi-agent system that emulates expert iterative reasoning and adapts to variable spectral inputs.
Motivation
- Reliable de novo elucidation remains difficult for large, complex, or ambiguous molecules because it requires sequential integration of complementary spectroscopic signals.The introduction distinguishes this task from matching expected compounds to reference standards or retrieving similar spectra from libraries, which cannot identify molecules absent from those libraries.Source: Introduction · p. 2
Limitations of prior approaches
- Existing spectra-guided generators lack flexibility beyond fixed training distributions, while repurposed language and vision-language models perform poorly at symbolic, iterative spectroscopic reasoning.Expert elucidation instead generates structural hypotheses, cross-validates them across modalities, and refines assignments by eliminating inconsistent isomers.Source: Introduction · p. 2
Design requirements
- The introduction identifies three requirements for effective systems: self-supervised learning from unpaired spectra, tolerance of arbitrary or missing modalities, and explicit closed-loop hypothesis testing.These requirements are motivated by the need to handle real-world spectral variability and emulate expert workflows.Source: Introduction · pp. 2–3
MACROS
- MACROS combines modality-specific agents in a hierarchical system trained on over 100 million spectra-structure pairs and uses repeated candidate generation, spectral simulation, ranking, and refinement.Spec2Mol proposes structures, Mol2Spec simulates and ranks spectra, and PromptSpec2Mol refines candidates over cycles; the workflow accepts additional prompts such as reactant structures or MS-derived information and supports test-time performance scaling.Source: Introduction · p. 3
Capabilities and interpretability
- MACROS is presented as generalizing zero-shot across real-world molecular classes, while attention-based interpretations and atom-level confidence support interpretable human–AI collaboration.The introduction also reports recovered textbook chemical-shift relationships and an emergent ring-first preference for rigid cyclic scaffolds exceeding 40%.Source: Introduction · p. 3
Results
MACROS combines multimodal structure prediction with iterative reasoning, generalizes across diverse spectra, and recovers interpretable chemical patterns. In real-world studies, it improved structure elucidation and accelerated expert analysis.
MACROS generalizes across diverse sources for de novo structure elucidation
- MACROS achieved 0.88 fingerprint similarity and >0.95 recall of common functional groups on the USPTO-derived test set, with accuracy improving after closed-loop refinement.It also improved fingerprint similarity across independent datasets: QM9S, 0.55 to 0.71; SimPubChem, 0.57 to 0.67; NMRMIND, 0.47 to 0.56; and NMRGym, 0.47 to 0.64, from draft to refined predictions. All results were obtained under strict zero-shot conditions.Source: MACROS generalizes across diverse… · pp. 3–4, Figure 1e · p. 4
Closed-loop multi-agent reasoning enhances structure elucidation at inference
- MACROS coordinates drafting, spectral rethinking, and refinement in an iterative multi-agent loop that improves predictions with additional iterations.Its Mol2Spec agents predict spectra directly from SMILES in one forward pass; for IR prediction, cosine similarity was 0.761, versus 0.719 for DetaNet and 0.720 for AttentiveFP.Source: Closed-loop multi-agent reasoning… · pp. 4–5
MACROS learns chemical intuition and interpretable reasoning
- Without peak-level supervision, MACROS recovered textbook chemical-shift–functional-group correlations and produced interpretable alignments between spectra and molecular substructures.The reported correlations included ketones at ¹H 2–3 ppm and ¹³C 195–225 ppm, and aldehydes at ¹H 9–10 ppm and ¹³C 190–205 ppm; their precision depends on database and measurement conditions.Source: MACROS learns chemical intuition and… · pp. 5–6
- Across four datasets, MACROS initiated >40% of SMILES sequences with ring systems, while learned NMR embeddings showed directional changes near aromatic ¹³C≈120 ppm and ¹H≈7 ppm.The ring-first preference mirrored an expert heuristic, emerged during self-supervised pretraining, and was refined during supervised fine-tuning.Source: MACROS learns chemical intuition and… · p. 6
MACROS generalizes to real-world spectra and augments human experts
- On natural-product and metabolite spectra, closed-loop reasoning improved zero-shot fingerprint similarity, while targeted fine-tuning raised performance further.For natural products, similarity rose from 0.559 draft to 0.625 after reasoning, and from 0.799 after fine-tuning to 0.843 after reasoning; 54.2% were correctly resolved. For metabolites, it rose from 0.519 draft to 0.589 after reasoning, and from 0.742 after fine-tuning to 0.816 after reasoning; 68.9% were correctly resolved.Source: MACROS generalizes to real-world… · pp. 6–7
- Zero-shot predictions reached 0.478 Tanimoto similarity on NMRBANK spectra and median similarity of 0.456 on in-house crude reaction products using ¹³C NMR alone.On NMRExp, draft similarity rose from 0.431 to 0.479 after reasoning, and from 0.683 after fine-tuning to 0.743 after reasoning.Source: MACROS generalizes to real-world… · p. 7
- With reactant prompts, MACROS annotated 44% of real-world reactions zero-shot and achieved a 0.677 TFPS draft score with ¹H and ¹³C NMR after rethinking.On the USPTO benchmark, reactant prompting yielded an average TFPS of 0.844 for the initial draft, rising to 0.889 after rethinking; on real-world cases, ¹³C-only TFPS rose from 0.556 to 0.5886 after rethinking.Source: MACROS generalizes to real-world… · pp. 7–8
- About 10 minutes per molecule and 0.807 TFPS with MACROS assistance, versus about 1 hour and 0.102 TFPS unaided, in a blinded expert study.MACROS alone achieved >90% structural accuracy in <1 minute per molecule on a single 24 GB GPU.Source: MACROS generalizes to real-world… · p. 8
Discussion
MACROS combines modality-specific pretrained agents with transparent closed-loop reasoning to elucidate structures from diverse routine spectra, with reported real-world generalization and emergent expert-like behavior. The discussion presents it as a scalable foundation for broader autonomous chemical research that will require further development.
Discussion
- Routine spectra remain insufficient for unambiguous full elucidation of especially complex molecules, though they are indispensable for exploring chemical space; broader data and tighter automated-experiment integration are future priorities.The discussion also proposes improved training and inference strategies and extension to chemical synthesis and natural-product discovery.Source: Discussion · pp. 8, 10
- Sixfold faster analysis with substantially enhanced accuracy is reported for MACROS, which processes >20 molecules per minute in quick mode and >1 per minute in full closed-loop mode on a 24-GB consumer GPU.The system is described as robust to arbitrary combinations of routine spectra and diverse real-world experimental samples, including reactions, natural products and metabolites.Source: Discussion · pp. 8–9
- Modality-specific pretraining on native spectral and molecular representations preserves physical inductive biases, while lightweight agents achieve higher accuracy than contemporary LLM-based approaches with orders-of-magnitude fewer adaptation parameters.NMR agents use peak lists, IR agents use continuous waveforms, and molecular representation agents use chemistry-specific SMILES tokenization.Source: Discussion · p. 9
- MACROS enables test-time scaling without gradient updates through expert-inspired drafting, reverse-spectrum consistency checks, and prefix-conditioned refinement around validated structural fragments.Reverse spectral agents predict NMR shifts and IR bands for comparison with raw experimental data; refinement constrains generation around chemically correct anchors.Source: Discussion · p. 9
- Without explicit shift-to-structure correspondence training, MACROS rediscovers textbook NMR chemical-shift and substructure mappings, while ablation and noise studies reveal an expert-like diagnostic hierarchy.Attention analyses support the recovered correspondence between spectral features and molecular substructures.Source: Discussion · p. 9
- MACROS exhibits a ring-first structural assembly strategy, preferentially starting SMILES generation with rigid cyclic scaffolds across tested datasets.The discussion interprets this behavior as aligning with expert chemists’ heuristic preferences.Source: Discussion · p. 10