Answers are grounded in this paper and cite their sources.
Identifying molecular structures from multiple analytical measurements is a longstanding challenge across natural sciences because different analytical results provide complementary information that is not readily coupled through consistent physical principles. Here, we propose a mixture of analytical reasoning agents (MARAs), an autonomous framework for molecular identification through comprehensive chemical analysis across infrared spectroscopy, Raman spectroscopy, and elemental analysis (EA). MARA organizes multiple reasoning large language models to collectively analyze, critique, and decide the spectroscopic and EA data, emulating chemical analysis processes across human experts. Without any human intervention, MARA achieved 95.2% accuracy in identifying 186,102 molecular structures from their spectroscopic and EA data. Furthermore, MARA showed 87.4% accuracy on an experimental dataset containing experimentally collected analytical observations of 5939 molecules, even though it contains large uncertainties from experimental procedures. For further evaluation under a fully realistic analytical task, we applied MARA to a PubChem-driven molecular library comprising 38.9 million molecules, presenting the majority of synthesizable compounds. MARA achieved 45.7% molecular identification accuracy despite being trained on only 0.015% of the library.
No takes yet. Share an insight, caveat, or question.
Na et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: