DEEPNET-LLM: Highly accurate structure prediction of unknown metabolites from mass spectra via their molecular neighborhoods

Abstract

Most molecular features detected by untargeted mass spectrometry-based metabolomics remain unannotated. This so-called 'dark matter' currently comprises more than 90% and likely holds uncharacterized bioactive chemistry. Structure prediction from fragmentation spectra remains an outstanding challenge, because a spectrum does not uniquely determine a structure and the space of candidate structures is mostly unbounded; correspondingly, existing methods rarely return the correct de novo structure as their top prediction. Here we describe De novo Ego network-Enabled Propagation and Neighborhood Extrapolation of Topology using Large Language Models (DEEPNET-LLM) reasoning approach, which uses Molecular Community Networking (MCN) to restrict the search of a query spectrum to its local neighborhood within the global metabolome, and applies AI reasoning to propagate structural information from neighbors to the query. For a collection of 900 authentic standards (2,249 spectra) that have never been publicly released to exclude possibility of data leakage, DEEPNET-LLM returned the correct structure as its top prediction for 84.7% of spectra. Accuracy was independent of precursor mass across a range over which the number of candidate structures grows by 30 orders of magnitude, and was comparable across ion polarity, adducts and in-source fragmentation. On a random subset of the public MassSpecGym benchmark, DEEPNET-LLM reached 85.9% top-1 accuracy excluding stereochemistry. We showcase the use of DEEPNET-LLM with an example of a discovery of tryptamine amidate of a bile acid in mouse gut, confirmed against a synthetic standard. The DEEPNET-LLM approach is orthogonal to existing in silico methods and can be combined with them.