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
Automated de novo structure elucidation from routine spectra remains difficult because existing systems lack reliable reasoning over unseen, multimodal data. MACROS addresses this with a purpose-built hierarchical multi-agent architecture and expert-modeled closed-loop refinement, achieving broad zero-shot generalization and faster, more accurate human-assisted elucidation.
Problem
Scalable automated de novo structure elucidation remains challenging for complex spectra and requires reasoning across heterogeneous modalities.
Method
MACROS uses modality-specific agents, large-scale spectral pretraining, hierarchical coordination, and explicit closed-loop hypothesis testing modeled on expert spectroscopists.
Results
MACROS generalizes zero-shot across diverse real-world spectra, while human collaboration delivers sixfold faster analysis and substantially enhanced accuracy.
Takeaways & Limitations
MACROS provides a scalable framework for multimodal molecular structure elucidation and supports automated laboratory workflows.
Takeaways & Limitations
Reported evaluations were conducted under strict zero-shot conditions without test-time fine-tuning or modality-specific adaptation.
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.
Summary:
The paper concerns structure elucidation, multi-agent systems, closed-loop reasoning, and spectroscopy.
- The study combines structure elucidation with multi-agent closed-loop reasoning for spectroscopy.
Introduction
Automated de novo structure elucidation from routine spectra remains difficult, especially for complex molecules and variable real-world data. MACROS addresses this gap with spectroscopic foundation modeling, flexible multimodal handling, and expert-like closed-loop reasoning.
- Introduction: Scalable, fully automated de novo structure elucidation remains difficult for high-molecular-weight, architecturally complex, or stereochemically dense molecules.Structure determination underpins chemical, biological, and biomedical discovery.
- Introduction: Routine ¹H NMR, ¹³C NMR, HSQC NMR, and IR spectra form the primary basis for structural characterization in laboratory workflows.MS and heteronuclear NMR can provide supplementary formula or heteronuclear validation.
- Introduction: Existing spectra-guided generators lack flexibility across real-world spectral variability, while repurposed language and vision-language models degrade on structure elucidation.The task requires symbolic pattern matching, multistep deduction, and iterative validation.
- Introduction: Effective systems require self-supervised spectral pretraining, flexible handling of arbitrary or missing modalities, and explicit closed-loop reasoning.
- Introduction: MACROS implements these principles through modality-specific agents, hierarchical integration, large-scale training, and iterative reasoning modeled on expert spectroscopists.It is designed specifically for multimodal routine spectra rather than repurposed general-purpose foundation modeling.
Results
MACROS combines hierarchical multi-agent processing with closed-loop refinement, achieving accurate and interpretable structure prediction across simulated, experimental, natural-product, metabolite, reaction, and human-assisted settings.
- Core performance: 0.88 fingerprint similarity and over 0.95 recall of common functional groups were achieved on 560 USPTO test molecules.Closed-loop rethinking improved accuracy, while the combined rethinking score achieved a 0.71 concordance index.
- Core performance: 0.71 concordance index ranked the true structure within the top 25% in 52% of cases and within the top 10% in 31% of cases.
- Robustness and interpretability: The system remains robust to noisy and incomplete spectra, scales to molecules with more heavy atoms than training examples, and provides atom-level and molecular-level confidence scores.
- Cross-source generalization: 0.55 initial and 0.71 refined fingerprint similarity were obtained on QM9S, with comparable performance on SimPubChem.The reported evaluations were conducted under strict zero-shot conditions without test-time fine-tuning or modality-specific adaptation.
- Cross-source generalization: 0.47 to 0.56 fingerprint similarity on NMRMIND and 0.47 to 0.64 on NMRGym followed closed-loop refinement.
- Robustness and interpretability: NMR primarily supports forward drafting, whereas IR primarily supports rethinking validation; corresponding noise selectively impaired these stages.
- Interpretable reasoning: Self-supervised pretraining and supervised fine-tuning recovered textbook chemical-shift–functional-group correlations despite no explicit correspondence being shown.The recovered relationships remained contingent on spectrometer configuration, deuterated solvent, and measurement conditions.
- Interpretable reasoning: MACROS initiated SMILES sequences with ring systems over 40% of the time across QM9, NMRMIND, USPTO, and SimPubChem.This ring-first strategy aligned with the expert heuristic of anchoring interpretation on rigid cyclic scaffolds.
Discussion
MACROS combines modality-specific learning with closed-loop, expert-inspired reasoning to address the cognitive and multimodal demands of structure elucidation. It generalizes across routine spectra and real-world samples while providing interpretable chemical reasoning and a foundation for autonomous chemical research.
- Discussion: MACROS addresses structure elucidation as iterative hypothesis testing across complementary, heterogeneous spectra, a cognitively demanding task for organic chemists.Routine spectroscopic methods remain indispensable despite insufficient structural information for unambiguous elucidation of highly complex species.
- Discussion: MACROS supports arbitrary combinations of routine spectra and shows robust zero-shot generalization to chemical reactions, natural products, and metabolites.The framework is designed to tolerate real-world spectral variability while extending across diverse experimental samples.
- Discussion: MACROS processes >20 molecules per minute in quick mode and >1 molecules per minute in full closed-loop mode on a single consumer-grade GPU (24 GB).The system is reported to deliver sixfold faster analysis with substantially enhanced accuracy in expert-driven structure elucidation.
- Discussion: MACROS uses modality-specific native-format pretraining and an explicit multi-agent workflow for drafting, rethinking, and refining candidate structures.The rethinking stage re-encodes candidates through reverse spectral agents, while refinement uses validated fragments or drafts as prefixes for local optimization.
- Discussion: Interpretability analyses show that MACROS reconstructs canonical spectral–substructure correspondences and develops a ring-first strategy across tested datasets.The model recovers textbook NMR correlations without explicit correspondence supervision and preferentially begins SMILES generation with rigid cyclic scaffolds.
- Discussion: MACROS establishes a scalable foundation for generalizable multimodal structure elucidation and a blueprint for integration with automated experimentation.The discussion identifies broader datasets, improved training and inference, and tighter laboratory integration as future directions.