AIMe Maps Chemical Space of Small Molecules

Researchers at the Boyce Thompson Institute and Cornell University have unveiled AIMe, an artificial intelligence platform designed to overcome a persistent bottleneck in biomedical research: the identification of small molecules in biological samples. Developed by Frank Schroeder of the Boyce Thompson Institute and Cornell chemistry department, alongside Carla Gomes, director of Cornell’s AI for Science Institute, AIMe leverages neuro-symbolic AI to map and search the mass spectra of over 100 million known small organic compounds. The tool directly addresses a critical limitation in mass spectrometry, where traditional reference libraries capture fewer than one percent of known compounds, leaving the vast majority of detected metabolites structurally unannotated. At the core of the platform is DeepMS2Reasoner, a model that computationally simulates molecular fragmentation within mass spectrometers. By combining symbolic chemical rules to generate physically plausible breakdown pathways with neural networks that assign probabilistic outcomes, the system produces predicted tandem mass spectra alongside chemically interpretable fragmentation maps. These predictions have been organized into MS2KOSMOS, a searchable database encompassing more than 800 million spectra derived from the PubChem repository. This expansion effectively increases the searchable chemical space by a factor of one thousand compared to existing experimental libraries. The researchers validated AIMe through comparative metabolomics analysis of germ-free versus conventional mice, focusing on microbiota-dependent compounds. Of the 111 most abundant unidentified features, approximately one-third yielded direct spectral matches or structurally related candidates. The platform also successfully navigated spectral ambiguities by mapping unknown compounds to molecular neighborhoods, enabling hypothesis-driven structure elucidation. In a notable application, the team identified two polyamine derivatives that defied conventional matching. Through iterative candidate synthesis and spectral comparison guided by AIMe, researchers confirmed a linear putrescine derivative and discovered a previously unreported macrocyclic polyamine. Subsequent public database screening located the novel ring-shaped compound in human fecal samples, revealing a substantial gap in the understood catalog of microbiome-derived metabolites at the intersection of diet, immunity, and microbial ecology. Beyond targeted studies, AIMe demonstrated repository-scale utility by processing 7.4 million spectral clusters from the Global Natural Products Social Molecular Networking database. The platform generated putative annotations for approximately 2.69 million clusters using established similarity thresholds, drastically outpacing prior manual annotation efforts. AIMe is currently accessible via web interface, with peer-reviewed publication and open-source code release pending. By transforming fragmented mass spectral data into actionable structural hypotheses, the tool establishes a new scalable paradigm for metabolomics, toxicology, and small-molecule discovery.