MACROS: Multi-Agent Closed-Loop Reasoning for Organic Structure Elucidation
Autonomous structure determination from routine multimodal spectra: ¹H NMR, ¹³C NMR, HSQC, and IR
MACROS (Multi-Agent Closed-Loop Reasoning for Organic Structure) is a purpose-built multi-agent system designed for automated structure elucidation of organic molecules from routinely acquired multimodal spectroscopic data: ¹H NMR, ¹³C NMR, HSQC, and IR.
Unlike approaches that adapt general foundation models, MACROS is engineered from the ground up with modality-specific components and an explicit multi-agent closed-loop reasoning mechanism that emulates the iterative, expert human process for resolving complex spectra.
- Modality-Specific Pretrained Agents — Each agent is self-supervised on large-scale simulated data tailored to its spectral type.
- Hierarchical Integration & Joint Training — Agents are unified in a hierarchical framework and jointly trained on ~105 million simulated spectra–structure pairs.
- Fine-Tuning on Real Data — Final adaptation to experimental variability using unassigned real-world spectra.
The following checkpoints are (or will be) available, each optimized for different application domains:
-
General Pretrain checkpoint
Broad simulated pretraining — foundation model trained on the full ~105M simulated spectra–structure pairs -
Human Metabolism checkpoint
Fine-tuned for human metabolites — enhanced performance on endogenous small molecules and metabolic pathways -
Natural Product checkpoint
Adapted for complex natural products (>800 Da) — optimized for large, structurally intricate molecules from natural sources -
Organic Chemistry checkpoint
General-purpose for synthetic / reaction products — suited for small-to-medium synthetic organic compounds and reaction mixtures