Mass Spectrometry Molecule Retrieval

Tandem mass spectrometry breaks a molecule into fragments and measures their masses relative to charge. The resulting peaks are clues to the original structure. Strong peaks identify common fragments; a weak peak can distinguish two otherwise similar molecules.

This task gives the model a spectrum and a list of candidate structures. It must rank them, placing the true molecule first. The difficulty is ambiguity: related molecules can fragment similarly, and changing the instrument or collision energy changes the pattern.

A real spectrum, and a real retrieval mistake

Public validation query MassSpecGymID0079471 · 45 measured peaks · QTOF · [M+H]+ · recorded collision energy 20. Three structures are shown from its 256-candidate pool.

MassSpecGym developers (2024), MIT license. Example data · Source license. This real-data panel is separate from the synthetic interactive experiment below.

Two routes to one complete ranking

Candidate molecular graphAtoms + bonds
Acquisition metadata

→

Graph networkConnect atoms and bonds
Learned graph features

→

Fragment + intensity headsWhich atoms stay together?
How strong is each peak?

→

Predicted spectrumEntropy comparison with observed peaks
Rank supplied candidates

Fallback path: observed peaks → Fourier-DeepSets → predicted structural features → compare candidate fingerprints → complete ranking. Public acquisition/structure checks choose the route; the true identity is never a routing input.

Train: MAGMa fragment targets + experimental intensities supervise the simulator. Inference: compare every supplied candidate with the observed spectrum. Native model: 15.1M parameters, random initialization. Method and results.

MassSpecGym provides about 194,000 training spectra, 19,000 validation spectra and 17,556 test queries. A query contains peak positions and intensities plus acquisition information, with up to 256 candidate structures. Intensities describe relative peak strength; they do not directly measure the abundance of molecules across experiments.

The correct formula is not an extra clue. Models must distinguish mass-compatible alternatives using the spectrum and candidates. Training starts without pretrained weights. Optional simulated spectra can broaden chemical coverage, but their patterns may differ from real measurements.

The test set is publicly available, making careful separation from model development especially important. Results measure this known benchmark rather than a newly collected, secret test.

Hit@1 is the fraction of queries with the correct first choice. Hit@5 and Hit@20 ask whether it appears in a shortlist. A structural-distance metric also describes how chemically different a wrong first answer is.

For an illustrative four-query set with correct ranks 1, 2, 6 and 22, Hit@1 is 25%, Hit@5 is 50% and Hit@20 is 75%. Every query counts equally. The normalized reward rescales first-choice accuracy between two reference systems; it is not the probability that a proposed structure is correct.

An inverse model starts with the observed peaks and predicts a molecular fingerprint: a compact description of structural features. It then compares that fingerprint with each candidate. This is the approach used by the simpler DeepSets reference.

GLACIER works in the other direction. It reads a candidate as a graph of atoms and bonds, predicts how it might fragment, and estimates the resulting spectrum. Candidates whose simulated spectra best match the observation rank higher.

The simulator does not support every instrument and molecular input. The reference system therefore uses it where supported and falls back to the inverse model elsewhere. Its overall score measures this complete system, including the difficult cases.

Training annotations associate observed peaks with plausible subsets of the molecule’s atoms. These teach the simulator which fragments can form; measured intensities teach which fragments are prominent under particular acquisition conditions.

But reconstructing a spectrum is not identical to identifying a molecule. A model can reproduce the dominant peaks while missing the small diagnostic peak that separates two candidates. Validation must therefore measure candidate ranking as well as spectral similarity.

The from-scratch GLACIER system identifies about 40% of test molecules as its first choice, versus about 4% for the simple reference. It places the answer in the first five about 57% of the time and in the first twenty about 71% of the time.

An official-weight hybrid reaches about 49%, but uses different pretrained components and training history. This is a reference target, not a controlled test of architecture alone. A separate eight-seed reviewer summary reports about 40.5% first-choice accuracy.

The fallback handles about 43% of queries but identifies only about 7% correctly, compared with about 66% on the simulator-supported group. Most of the gap to the official-weight hybrid occurs in that fallback group.

This suggests improving coverage or inverse matching could be valuable. It does not prove that replacing the fallback will recover the entire gap. These test diagnostics describe a fixed system; new design choices should be developed on validation data.

Train on confusing candidate pairs, improve acquisition-aware simulation, or combine forward and inverse evidence more effectively. The key test is whether better fragment predictions help distinguish similar structures across the full range of instruments.