AIMe

Untargeted metabolomics detects far more molecules than it can name. Reference libraries cover a few hundred thousand compounds out of the tens of millions of known small molecules, so the structures behind most tandem mass (MS²) spectra in public repositories stay unassigned. AIMe (AI Molecule Explorer) approaches that gap from the opposite direction: instead of matching a measured spectrum against other measured spectra, it predicts MS² spectra for essentially every known small organic molecule and searches that predicted space.

AIMe is a multi-agent neuro-symbolic framework from the Gomes and Schroeder laboratories at Cornell University, the Boyce Thompson Institute, and Weill Cornell Medicine. Only one of its three agents is a trained neural system: DeepMS²Reasoner, which predicts a spectrum from a structure. The other two — MS²KOSMOSGenerator and MS²KOSMOSMapper — are orchestration and symbolic-retrieval layers built on top of it, responsible respectively for running DeepMS²Reasoner across PubChem and for querying the resulting index.

That division places AIMe apart from related work. LSM-MS2 embeds measured spectra to identify compounds; AIMe runs the forward direction — structure to spectrum — and treats its own predictions as the searchable reference. Because each predicted peak comes from an explicit chain of bond-breaking events, a match carries fragment formulas and fragment structures rather than a bare similarity score.

#Key Features

• Fragmentation as a sequence of actions: rather than regressing a peak vector, the model builds a directed acyclic graph of fragments, expanding chemically feasible operators — linear-bond cleavage, ring-bond cleavage, excision, and hydrogen transfer — with a learned halting probability at each node.
• Interpretable peak assignments: atoms are tracked through every bond-breaking action, so a predicted peak resolves to a molecular formula, a fragment structure, and the pathway that produced it.
• One neural component inside a symbolic scaffold: action enumeration, feasibility checks, fragment-graph edits, and canonical state merging are deterministic; learning enters only through the action and halting probabilities.
• Repository-scale retrieval: predicted spectra are indexed in a discrete Formula Coordinate System spanning roughly 30 million dimensions in positive mode, so a query scores only molecules sharing at least one fragment formula with it.
• Conditioned on acquisition settings: collision energy, instrument, and adduct enter as global features, matching predictions to the conditions a spectrum was acquired under.

#Technical Details

The learned component, the Action Likelihood Estimator Advisor (ALEA), is a Graphormer-style graph transformer over a feature-augmented molecular graph of the current fragment state. Node and edge features encode atom identity, bond type, ring context, and hydrogen bookkeeping; learned structural bias terms enter the self-attention, and action-specific heads emit logits for linear-bond cleavage, ring opening, excision, and halting. Training uses the NIST 2020 MS² library on an 80/10/10 split, with a KL-divergence loss between normalized predicted and observed spectra, regularizers that softly favor simple linear-bond explanations, and a curriculum that introduces higher-mass molecules progressively.

On the held-out NIST20 split, predicted spectra reach an average cosine similarity of 0.83 to experiment with a median of 0.89, against 0.61 to 0.75 for the other models evaluated on identical splits. Zero-shot, DeepMS²Reasoner records the highest average cosine on both CASMI 2022 (0.67) and MassSpecGym (0.59), and is the closest of six models in over 40% of cases versus under 20% for the runner-up, ICEBERG. Against decoy sets containing every PubChem isomer of the target, retrieval reaches 35.6%, 64.4%, and 79.8% top-1, top-10, and top-50 accuracy.

#Applications

Run at a fixed checkpoint across every PubChem organic compound below m/z 1000 — over 105 million structures, in two ionization modes at four collision energies — DeepMS²Reasoner produced more than 800 million spectra, organized into the MS²KOSMOS resource. In comparative metabolomics of germ-free and specific-pathogen-free mouse feces, mapping unknowns into that space guided elucidation of previously undescribed microbiota-dependent polyamines, including a macrocyclic spermine derivative confirmed by chemical synthesis and also detected in samples of human origin.

#Impact

Applied to 7,144,480 consensus spectra from the GNPS repository, AIMe returned annotations above cosine 0.7 for roughly 2.69 million of them, against 416,277 previously, adding about 1.27 million more at the stricter 0.8 threshold. That reframes MS² annotation as interpretable search across the known small-molecule universe rather than curated-library matching. The work is a preprint awaiting peer review, and neither the weights, the source code, nor MS²KOSMOS itself has been released; the authors state that code will be published to a named GitHub repository and MS²KOSMOS made available as a community resource on acceptance. No model card or data card accompanies the release.