Fragment Onset Thresholds as an Isomer Fingerprint

Mass spectrometers weigh molecules and the pieces they break into. Two molecules with the same atoms wired differently (structural isomers) weigh the same and often break into similar pieces. Current machine-learning tools for identifying unknown molecules from their fragment spectra therefore have a characteristic failure: they rank a "decoy" isomer, one whose spectrum looks almost identical, above the true structure. Particle physicists face the same kind of problem with exotic hadrons. A resonance's mass and width are often equally well explained by a compact multiquark state, a loosely bound two-hadron molecule, or a non-resonant threshold effect. The tool that has helped them is to stop treating the signal as a static peak. They study how it behaves as energy sweeps across the point where a decay channel opens, and they use that behavior to tell the hypotheses apart.

We propose to carry this idea into tandem mass spectrometry. Instead of scoring a candidate isomer by one fragment spectrum at a fixed collision energy, we model each fragment's breakdown curve: its intensity as collision energy is ramped. The curve is constrained to a threshold shape whose onset energy is anchored to a density functional theory (DFT) estimate of the energy needed to make that fragment. DFT is the standard quantum-mechanical method for computing molecular energies from the electron density. Isomers that share fragments rarely share onset energies, because the bonds that must break differ. The claim is that onsets separate decoy pairs that static spectra cannot, and that a model which knows where thresholds should sit can also report honestly when the data cannot break a tie.

Core Claim

Energy-resolved fragment onsets, predicted by a forward model whose thresholds are tied to cheap DFT-derived dissociation energies, carry isomer-discriminating information that collision-energy-conditioned black-box simulators do not exploit. Plugging such a scorer into an existing isomer-search method should raise top-1 identification specifically on the hard subset of spectra where the leading candidates are near-degenerate at any single energy. The same model should produce a calibrated "undecidable" verdict when even the breakdown curves fail to separate candidates.

This is different from the closest existing work in three ways. ICEBERG simulates fragmentation with a graph model that predicts bond-cleavage paths and fragment intensities, and it conditions on collision energy. But it learns the energy dependence as an unconstrained input. It has no notion of a threshold and no electronic-structure grounding. GEMS searches the space of same-formula graphs by local bond edits and documents that cosine similarity is a poor judge between decoy isomers. It needed a learned stopping probability to cope, but its scorer still compares static spectra. FRIGID uses ICEBERG inside a self-correction loop and so inherits the same static scoring. None of these uses DFT. None treats the shape of intensity versus energy as the primary evidence.

What carries over from hadron spectroscopy

The hadron literature motivates the design and also constrains it. On the motivating side, the BESIII study of the $\eta'\pi^+\pi^-$ spectrum, Observation of an anomalous line shape near the $p\bar{p}$ threshold (Observation of an anomalous line shape of the $\\eta^{\\prime}\\pi^{+}\\pi^{-}$ mass spectrum near the $p\\bar{p}$ mass threshold in $J/\\psi\\rightarrow\\gamma\\eta^{\\prime}\\pi^{+}\\pi^{-}$), found that the structural information sat in a slope change exactly at a channel threshold. A smooth resonance shape could not describe it, and a threshold-aware (Flatté) parameterization was required. The BESIII Dalitz-decay measurement, Observation of $\psi(3686)\rightarrow e^{+}e^{-}\chi_{c1,c2}$, makes the general point that how a rate changes with probe energy acts as a structural ruler. Deep learning exotic hadrons trained a classifier on $10^5$ synthetic line shapes drawn from a physical amplitude model. Its attribution analysis found that the decisive evidence lay in the energy window right next to the threshold. We adopt both lessons: generate training signal from a physically parameterized model, and focus scoring on the near-onset region.

On the constraining side, the same BESIII paper reports that a Flatté threshold molecule and a coherent sum of two interfering states fit the data equally well. A single observable can therefore be genuinely degenerate between mutually exclusive structures. We take this as a design requirement rather than an embarrassment. The scorer must compute a likelihood ratio between the top candidates and abstain when that ratio is not decisive, instead of always emitting a ranked answer.

We do not claim any physical equivalence between S-matrix poles and ion dissociation. The transfer is methodological. Thresholds carry structure, a parameterized threshold model is a better scorer than a free-form one near onsets, and degeneracy should be reported.

Research Program

The forward model predicts, for each candidate structure and each fragment $f$ it proposes, a breakdown curve

$$ I_f(E) = A_f \,\sigma\!\left(\frac{E - E_f^{0}}{w_f}\right) \exp\!\left(-\frac{\max(0,\,E - E_f^{0})}{\tau_f}\right). $$

The terms are defined as follows. $E$ is collision energy and $I_f(E)$ is the predicted relative intensity. $\sigma$ is the logistic function giving a smooth onset. $E_f^{0}$ is the onset energy and $w_f$ the onset width. $\tau_f$ is the scale at which the fragment is depleted by further breakdown. $A_f$ is the amplitude. The functional form follows the qualitative rise-then-decay shape of breakdown curves and is deliberately low-dimensional, so that the onset parameter cannot be absorbed by the amplitude. Fragment candidates come from ICEBERG's cleavage generator, reused unchanged.

The key ingredient is the onset energy:

$$ E_f^{0} = \alpha + \beta\,\Delta E_f^{\mathrm{DFT}} + r_\theta(G, f). $$

Here $\Delta E_f^{\mathrm{DFT}}$ is the DFT reaction energy for forming fragment $f$ from the protonated precursor, $\alpha$ and $\beta$ are global calibration constants, and $r_\theta$ is a small graph-network residual on molecular graph $G$. The residual is regularized toward zero so that the physics term carries the ordering.

DFT is far too slow to run inside a search loop. Li et al. report 2 to 10 CPU hours per molecule for related excited-state calculations in ECDFormer. We therefore run DFT once, at a modest level, on precursor and fragment pairs for a training subset of a few thousand NIST'20 compounds. We then train a surrogate network to predict $\Delta E_f^{\mathrm{DFT}}$ from graph and fragment. Only the surrogate is called during search. This follows the pattern ECDFormer used to replace TD-DFT with a 1.5-second network, applied here to ground-state fragmentation energies rather than spectra. DFT's standing as the method for barriers and reaction energies, illustrated in mechanistic work such as the Breslow intermediate review, is what justifies the anchor. Its known weaknesses for barrier heights are why we keep a learned residual and report how much the residual moves the onsets.

The scorer compares observed multi-energy spectra of an unknown with predicted curves. It uses a likelihood that upweights energies within one onset width of predicted thresholds, mirroring the near-threshold focus of Ng et al. This scorer replaces the spectral-discrepancy signal inside GEMS's edit-based search, with GEMS's policy and seeds otherwise untouched. For the final top two candidates, it reports a log-likelihood ratio. Spectra whose ratio falls below a threshold tuned on validation data are returned as "undecidable between A and B."

Baselines

The naive reference point is library-style cosine similarity between the unknown's spectra and each candidate's ICEBERG-predicted spectrum, summed over available energies. GEMS shows this is weak on decoys. It tells us how hard the hard subset is.

The strongest existing scorer is ICEBERG itself, used with its native collision-energy conditioning at every energy for which the unknown was recorded. This is the real competitor: it already sees energy and could in principle learn thresholds implicitly. ICEBERG reports 40% top-1 retrieval on NIST'20 [M+H]+, against 25% for earlier models (ICEBERG).

The strongest search method is GEMS with its own scorer. It reports 30.48% top-1 on the NIST unique subset of over 17,000 spectra and 72.19% search success within 16 trajectories, against DiffMS at 6.53% and MolSpecFlow at 8.22% (GEMS). As a generative reference we include FRIGID, which reports 25.03% top-1 on NPLIB1 and 18.29% on MassSpecGym (FRIGID). These are different datasets, so those numbers are context, not a direct comparison. We will rerun every method on our splits.

Fairness rules apply to all comparisons. Every scorer sees the same multi-energy spectra, the same candidate sets and seeds, and the same number of GEMS trajectories and model calls. We give ICEBERG the same retraining budget on the multi-energy training split that our model receives.

Two ablations isolate the key ingredient. The first keeps the threshold-shaped curve but sets $\beta = 0$, so onsets are purely learned. This tests whether DFT matters. The second keeps the DFT surrogate energies but feeds them as extra node features into ICEBERG without the threshold form. This tests whether the structured curve matters.

Beating the field means a statistically significant top-1 gain over GEMS-with-ICEBERG-scoring on the degenerate subset, at matched model calls. It also means no loss on the full set. Calibrated abstention must achieve higher accuracy on the answered spectra than the baseline achieves at the same coverage.

Why not just…?

Why not just give ICEBERG every collision energy and let it learn the energy dependence? This is the most serious threat. If the black-box energy conditioning already places onsets correctly, our gain will be zero. The first ablation and the ICEBERG multi-energy baseline settle it directly. We will also measure onset-ordering accuracy, meaning whether the predicted order of fragment appearance matches observation, for both models on held-out scaffolds. That is where an unconstrained model should fail if it is merely interpolating energies seen in training. If ICEBERG matches us, we will report that the hadron-style threshold structure adds nothing for this chemistry.

Why not just add DFT bond energies as features to an existing simulator? This is cheaper and plausible, and the second ablation is exactly that. Our hypothesis is that an energy feature without a threshold form lets the network trade onset against amplitude, which blurs the decoy-separating signal. If the feature-only variant matches the full model, we adopt the simpler design.

Why not just run DFT on every candidate? The file's cost figures rule this out for search. At 2 to 10 CPU hours per molecule (ECDFormer), a thousand-candidate set costs thousands of CPU hours per unknown. We do run full DFT on the top two candidates for a small audit sample, to check that the surrogate's onset ordering agrees with direct DFT.

Why not just use a more powerful generator such as FRIGID? Generators improve candidate proposal. The decoy problem, however, is a scoring problem, and FRIGID's refinement loop reuses static ICEBERG spectra. Our scorer can drop into FRIGID's loop as well, and we will test that as a secondary experiment.

Proposed Experiments

The primary benchmark is the subset of NIST'20 [M+H]+ compounds recorded at several collision energies. Confirming the size of this subset is the first step. If it is too small, the project is re-scoped to spectra with at least three energies. We split by scaffold. Within the test set we define the degenerate subset: spectra whose top two GEMS candidates have predicted single-energy spectral cosine above 0.9. This is the regime the method targets.

Metrics are top-1 and top-10 accuracy, pairwise isomer discrimination AUC on true-versus-decoy pairs, onset-ordering accuracy, and accuracy-versus-coverage curves for abstention. We also report CASMI zero-shot reachability against GEMS's 32.02% as an out-of-distribution check, and NPLIB1 and MassSpecGym to confirm that single-energy data, where the method reduces to a constrained ICEBERG, is not harmed.

Three outcomes decide the project. The first is whether the full model beats the ICEBERG multi-energy baseline on the degenerate subset at matched compute; if not, the central claim fails. The second is whether the $\beta = 0$ ablation loses most of that gain; if not, DFT is not contributing. The third is whether abstained spectra are enriched for true-versus-decoy pairs whose DFT-audited onsets differ by less than the onset width. That would show the "undecidable" verdict reflects real physical degeneracy, the analogue of the Flatté-versus-interference tie, rather than model noise.

References

  1. Neural Spectral Prediction for Structure Elucidation with Tandem Mass Spectrometry
  2. GEMS: Molecular Structure Identification via Geodesic Navigation of the Isomer Manifold
  3. FRIGID: Scaling Diffusion-Based Molecular Generation from Mass Spectra at Training and Inference Time
  4. Observation of an anomalous line shape of the $\eta^{\prime}\pi^{+}\pi^{-}$ mass spectrum near the $p\bar{p}$ mass threshold in $J/\psi\rightarrow\gamma\eta^{\prime}\pi^{+}\pi^{-}$
  5. Observation of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ψ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>3686</mml:mn><mml:mo stretchy="false">)</mml:mo><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mi>χ</mml:mi><mml:mrow><mml:mi>c</mml:mi><mml:mi>J</mml:mi></mml:mrow></mml:msub></mml:math> and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>χ</mml:mi><mml:mrow><mml:mi>c</mml:mi><mml:mi>J</mml:mi></mml:mrow></mml:msub><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mi>J</mml:mi><mml:mo stretchy="false">/</mml:mo><mml:mi>ψ</mml:mi></mml:math>
  6. Deep learning exotic hadrons
  7. Decoupled peak property learning for efficient and interpretable ECD spectra prediction
  8. Tale of the Breslow intermediate, a central player in N-heterocyclic carbene organocatalysis: then and now