Researchers at the Boyce Thompson Institute and Cornell University are developing AIMe, short for AI Molecule Explorer. It uses neuro-symbolic artificial intelligence (AI) to predict, organize, and search mass spectra for more than 100 million known small organic molecules.
AIMe predicts spectra by computer and stores them in a resource called MS2KOSMOS. It produced more than 800 million predicted spectra for essentially all small organic molecules in PubChem, a large public catalog of chemicals. That is about a thousand times more searchable space than experimental libraries offer. At the center is DeepMS2Reasoner, which simulates how a molecule would fall apart. Symbolic rules list physically possible breaks; a neural network scores how likely each break is. The output is both a prediction and a map of the breakup that a chemist can read.
From mouse gut chemistry to human samples
The researchers then used the tool on gut microbes for metabolomics data, which means a survey of many small molecules in a sample. They compared germ-free mice, raised with no gut microbes, with mice that had a normal gut community. Thousands of chemical signals differed. For the 111 most abundant unnamed compounds that depended on microbes, AIMe found close predicted matches or related structures for about one third. For the rest it pointed to neighborhoods of similar molecules.
Two signals looked like polyamines, a known family of small nitrogen-rich compounds, but did not match any described structure. Guided by AIMe, the researchers built candidate structures, predicted their spectra, and compared. One compound was a linear relative of putrescine and was confirmed by making the authentic chemical. The other only made sense as a macrocyclic polyamine, a ring-shaped form not previously reported from mouse or human biology. The same signal later appeared in 57 of 99 human fecal samples in a public database.
Applied to more than 7 million spectral clusters in the GNPS public repository, where earlier work had annotated about 416,000 clusters, AIMe offered tentative annotations for about 2.69 million. The tool is online; code is to be released with the paper. The work is a described in a preprint published in bioRxiv.