acceptodds
Withdrawn from ICLR 2027

CLEF: Cleavage Events Guide Electron Ionization Mass Spectrum Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mass spectrometry is widely used to identify compounds and analyze molecular structures. Predicting electron ionization (EI) mass spectra from structures can extend reference libraries to compounds without measured spectra. Accurate EI spectrum prediction and reliable retrieval among constitutional isomers are complementary goals. Both depend on modeling how molecular connectivity affects fragment intensities. We introduce CLEF, a structure-guided EI spectrum predictor. It uses atom connectivity, bond types, and candidate cleavage events to score fragment formulas. The events are enumerated from the input structure and matched to formulas by elemental composition. Composition-matched attention updates formula representations with event information. These representations then guide atom pooling and formula scoring, allowing the same formula to receive different weights in constitutional isomers. The model is trained on paired structures and spectra without event annotations. CLEF outperforms six trained baselines in spectrum prediction and isomer retrieval on NIST23 and MoNA, and in MoNA-to-SWGDRUG transfer. With only 25% of NIST23 training data, CLEF achieves spectral accuracy comparable to that of a leading full-data predictor.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
EI mass spectrum prediction, graph neural networks, data-efficient learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qZGMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CLEF to predict EI-MS spectra from molecular graphs, enumerated bond-cleavage events, and candidate fragment formulas. Composition-matched cross-attention updates formula tokens with cleavage context to guide atom pooling and isotope rendering. The model trains end-to-end on structure-spectrum pairs without fragment-level annotations. Evaluation covers spectrum prediction, isomer retrieval, and cross-library transfer on NIST23, MoNA, and SWGDRUG. CLEF outperforms the baselines across these tasks by conditioning formula scoring on local cleavage context.

Critical strengths

Conditioning formula scores on candidate bond cleavages introduces a structural bias that separates constitutional isomers sharing fragment formulas. CLEF shows improvements across spectral and retrieval metrics. Strong zero-shot transfer from MoNA to SWGDRUG shows that predictions generalize across libraries. Scaffold and anonymous topology splits reduce structural leakage, while shared PubChem decoys standardize isomer retrieval.

Critical weaknesses

Methodological novelty is limited: formula scoring and fragmentation graphs are established, and the retrieval setup follows SCARF without initial attribution.

The ablations do not isolate composition-matched attention from raw event features, additional model capacity, and latent-set processing.

MoNA’s same-source retrieval covers only 6-8 formula groups per split, leaving uncertainty about small performance gains. The data-efficiency experiments also retain the full labeled validation set across all training fractions.

Questions for authors

Can the authors explain the retrieval protocol of SCARF and clarify the methodological advance over the existing fragmentation models?

Does composition-matched attention outperform simple aggregation of matching events or unrestricted event attention under comparable parameter and training budgets?

Can the authors report paired, formula-group-aware confidence intervals for retrieval and evaluate isomer retrieval on the external SWGDRUG cohort?

Do the limited-data gains hold with a proportionally reduced validation set, and what are the preprocessing times, inference throughput, and filter coverage losses?

4 comments

Thank you for the detailed review. Revisions specific to your comments are purple. Shared methodological, computational and filtering revisions are dark red.

Q1. SCARF retrieval protocol and methodological advance.

SCARF (Goldman et al., 2023) ranks a target with 49 same-formula PubChem isomers selected by Morgan fingerprint Tanimoto similarity, using spectral cosine similarity. Section 2, Tandem prediction, now explains and cites this protocol. Section 4.1 attributes the candidate-construction strategy, which we apply to EI with candidate lists shared across methods.

Section 2 also distinguishes the computation. ICEBERG generates and scores fragmentation graphs. FraGNNet predicts fragment-node/formula probabilities and marginalizes over nodes. CLEF combines cleavage descriptors and local atom states through composition-matched cross-attention to update each formula representation before atom pooling and formula scoring (Sections 3.2 and 3.3, Figure 2). The revised discussion in Section 2 is dark red. The detailed definitions and diagram revisions are blue.

Q2. Uniform aggregation, unrestricted attention and model capacity.

We compare learned composition-matched attention with uniform aggregation and unrestricted all-event attention in Section 4.3 and Table 4(c). Uniform aggregation assigns equal weight to the events matching each formula. Section 4.3 also reports parameter-count and attention-complexity comparisons. Table 4(a,b) and Appendix C.2 provide the branch and descriptor controls and event-erasure analysis.

Table 4. NIST23 ablations. Bold marks the best value in each block.

\hline & & \text{Spectrum prediction} & & & \substack{\text{Same-source}\\\K=5} & & \substack{\text{Cross-source}\\\K=50} & \\\\ \text{Variant} & \text{DP} & \text{SDP} & \text{COS} & \text{Top_1} & \text{R@1} & \text{MRR} & \text{R@1} & \text{MRR} \\\\ \hline \rlap{\text{(a) Conditioning branches}} & & & & & & & & \\\\ \hline \text{Graph} & 0.7691 & 0.5051 & 0.6303 & 0.3463 & 0.4814 & 0.6806 & 0.1832 & 0.3134 \\\\ \text{Graph + formula} & 0.8091 & 0.7437 & 0.7001 & 0.4401 & 0.5492 & 0.7240 & 0.2845 & 0.4277 \\\\ \text{CLEF} & \mathbf{0.8729} & \mathbf{0.8758} & \mathbf{0.8147} & \mathbf{0.5782} & \mathbf{0.7386} & \mathbf{0.8477} & \mathbf{0.5567} & \mathbf{0.6762} \\\\ \hline \rlap{\text{(b) Event-descriptor alignment}} & & & & & & & & \\\\ \hline \text{CLEF} & \mathbf{0.8729} & \mathbf{0.8758} & \mathbf{0.8147} & \mathbf{0.5782} & \mathbf{0.7386} & \mathbf{0.8477} & \mathbf{0.5567} & \mathbf{0.6762} \\\\ \text{Shuffled descriptors} & 0.8577 & 0.8525 & 0.7887 & 0.5464 & 0.6945 & 0.8202 & 0.4824 & 0.6166 \\\\ \text{Formula–graph null} & 0.8072 & 0.7471 & 0.6965 & 0.4313 & 0.5483 & 0.7228 & 0.2920 & 0.4319 \\\\ \hline \rlap{\text{(c) Event aggregation}} & & & & & & & & \\\\ \hline \text{CLEF} & \mathbf{0.8729} & \mathbf{0.8758} & \mathbf{0.8147} & \mathbf{0.5782} & \mathbf{0.7386} & \mathbf{0.8477} & \mathbf{0.5567} & \mathbf{0.6762} \\\\ \text{Composition uniform} & 0.8700 & 0.8709 & 0.8097 & 0.5738 & 0.7293 & 0.8418 & 0.5432 & 0.6642 \\\\ \text{Full attention} & 0.7720 & 0.6535 & 0.6268 & 0.3466 & 0.4584 & 0.6619 & 0.1848 & 0.3102 \\\\ \hline \end{array}$$ K=5 denotes same-source retrieval and K=50 cross-source retrieval.

Q3. Formula-group-aware uncertainty and external isomer retrieval.

For NIST23 cross-source retrieval, we provide paired formula-group bootstrap estimates here. After averaging query-level differences across three training runs, we resample formula groups with replacement, retain all queries in each selected group, and recompute the query-weighted difference using the same draw for both models. The cohort contains 10,156 queries in 3,527 groups. Over 10,000 resamples, the R@1 difference is 0.0413 with a 95 percent interval of [0.0338,0.0488][0.0338,0.0488], and the MRR difference is 0.0311 with interval [0.0256,0.0366][0.0256,0.0366]. These estimates concern NIST23. MoNA same-source retrieval retains 6 to 8 formula groups per split.

Table 2 adds all four retrieval metrics on SWGDRUG using MoNA-trained models. CLEF has same-source K=5K=5 R@1/MRR of 0.209/0.463 and cross-source K=50K=50 R@1/MRR of 0.270/0.427. RASSP:SN has 0.197/0.453 and 0.143/0.289. MassFormer has the highest same-source R@1, 0.337. CLEF has the highest cross-source R@1 and MRR among the transferred trainable models.

Table 2. SWGDRUG spectrum and isomer-retrieval evaluation. Trainable models transfer from MoNA, with mean ±\pm sample SD. Bold and underline denote the best and second-best means among matched methods. CFM-ID (†\dagger) is a frozen external reference. Details are in Appendix A.2.

\hline & & \text{Spectrum prediction} & & & \text{Same-source (}K{=}5\text{)} & & \text{Cross-source (}K{=}50\text{)} & \\\\ \text{Method} & \text{DP }\uparrow & \text{SDP }\uparrow & \text{COS }\uparrow & \text{Top_1 }\uparrow & \text{R@1 }\uparrow & \text{MRR }\uparrow & \text{R@1 }\uparrow & \text{MRR }\uparrow \\\\ \hline \text{CLEF} & \mathbf{0.724}\pm0.007 & \mathbf{0.651}\pm0.013 & \mathbf{0.605}\pm0.018 & \mathbf{0.413}\pm0.027 & 0.209\pm0.028 & 0.463\pm0.014 & \mathbf{0.270}\pm0.021 & \mathbf{0.427}\pm0.017 \\\\ \text{RASSP:SN} & \underline{0.648}\pm0.005 & \underline{0.569}\pm0.009 & \underline{0.439}\pm0.015 & \underline{0.212}\pm0.018 & 0.197\pm0.040 & 0.453\pm0.022 & \underline{0.143}\pm0.009 & \underline{0.289}\pm0.009 \\\\ \text{MassFormer} & 0.561\pm0.006 & 0.498\pm0.006 & 0.331\pm0.007 & 0.150\pm0.011 & \mathbf{0.337}\pm0.022 & \underline{0.572}\pm0.015 & 0.079\pm0.010 & 0.194\pm0.008 \\\\ \text{RASSP:FN} & 0.546\pm0.018 & 0.432\pm0.034 & 0.297\pm0.013 & 0.128\pm0.009 & 0.185\pm0.032 & 0.448\pm0.019 & 0.054\pm0.006 & 0.149\pm0.008 \\\\ \text{MolFormer} & 0.478\pm0.010 & 0.506\pm0.004 & 0.308\pm0.008 & 0.127\pm0.010 & 0.316\pm0.034 & 0.552\pm0.025 & 0.072\pm0.006 & 0.182\pm0.006 \\\\ \text{NEIMS} & 0.568\pm0.006 & 0.517\pm0.006 & 0.354\pm0.006 & 0.160\pm0.005 & \underline{0.332}\pm0.025 & \mathbf{0.577}\pm0.021 & 0.072\pm0.007 & 0.181\pm0.007 \\\\ \text{Hybrid} & 0.385\pm0.018 & 0.493\pm0.009 & 0.155\pm0.021 & 0.040\pm0.022 & 0.205\pm0.026 & 0.457\pm0.014 & 0.063\pm0.007 & 0.172\pm0.008 \\\\ \text{CFM-ID}^{\dagger} & 0.552\pm0.000 & 0.500\pm0.001 & 0.534\pm0.000 & 0.338\pm0.001 & \text{–} & \text{–} & \text{–} & \text{–} \\\\ \hline \end{array}$$

Q4. Reduced validation, preprocessing, throughput and filtering coverage.

We reselect checkpoints on nested validation subsets scaled with training fraction and evaluate on the corresponding nested test subsets (Appendix C.1, Figure 5). CLEF has higher means than RASSP:SN on all eight metrics at each fraction.

Figure 5 data. Each cell gives mean ± sample SD in CLEF / RASSP:SN order. Bold marks the higher mean. The 100% row uses the full validation and test sets.

Spectrum prediction

\hline \text{Train} & \text{DP} & \text{SDP} & \text{COS} & \text{Top_1} \\\\ \hline 1\\% & \mathbf{0.7389}\pm0.0135\\,/\\,0.6769\pm0.0159 & \mathbf{0.7420}\pm0.0330\\,/\\,0.6880\pm0.0350 & \mathbf{0.5930}\pm0.0232\\,/\\,0.4785\pm0.0304 & \mathbf{0.3147}\pm0.0196\\,/\\,0.1825\pm0.0465 \\\\ 5\\% & \mathbf{0.7952}\pm0.0018\\,/\\,0.7475\pm0.0028 & \mathbf{0.7919}\pm0.0065\\,/\\,0.7505\pm0.0117 & \mathbf{0.6875}\pm0.0008\\,/\\,0.6106\pm0.0040 & \mathbf{0.4210}\pm0.0087\\,/\\,0.3214\pm0.0062 \\\\ 10\\% & \mathbf{0.8209}\pm0.0015\\,/\\,0.7687\pm0.0029 & \mathbf{0.8197}\pm0.0039\\,/\\,0.7777\pm0.0090 & \mathbf{0.7243}\pm0.0044\\,/\\,0.6487\pm0.0082 & \mathbf{0.4582}\pm0.0162\\,/\\,0.3622\pm0.0123 \\\\ 25\\% & \mathbf{0.8488}\pm0.0010\\,/\\,0.8028\pm0.0005 & \mathbf{0.8479}\pm0.0037\\,/\\,0.8111\pm0.0032 & \mathbf{0.7723}\pm0.0028\\,/\\,0.7129\pm0.0017 & \mathbf{0.5181}\pm0.0046\\,/\\,0.4372\pm0.0020 \\\\ 100\\% & \mathbf{0.8743}\pm0.0015\\,/\\,0.8388\pm0.0006 & \mathbf{0.8766}\pm0.0011\\,/\\,0.8514\pm0.0015 & \mathbf{0.8174}\pm0.0026\\,/\\,0.7783\pm0.0015 & \mathbf{0.5832}\pm0.0048\\,/\\,0.5287\pm0.0033 \\\\ \hline \end{array}$$ **Isomer retrieval** $$\begin{array}{crrrr} \hline \text{Train} & \text{K=5 R@1} & \text{K=5 MRR} & \text{K=50 R@1} & \text{K=50 MRR} \\\\ \hline 1\\% & \mathbf{0.4815}\pm0.0602\\,/\\,0.3998\pm0.0351 & \mathbf{0.6764}\pm0.0379\\,/\\,0.6233\pm0.0144 & \mathbf{0.2594}\pm0.0281\\,/\\,0.1440\pm0.0556 & \mathbf{0.3932}\pm0.0266\\,/\\,0.2711\pm0.0303 \\\\ 5\\% & \mathbf{0.5596}\pm0.0090\\,/\\,0.5420\pm0.0072 & \mathbf{0.7313}\pm0.0078\\,/\\,0.7163\pm0.0019 & \mathbf{0.3306}\pm0.0083\\,/\\,0.3051\pm0.0115 & \mathbf{0.4763}\pm0.0097\\,/\\,0.4425\pm0.0093 \\\\ 10\\% & \mathbf{0.6089}\pm0.0065\\,/\\,0.5692\pm0.0150 & \mathbf{0.7622}\pm0.0060\\,/\\,0.7383\pm0.0092 & \mathbf{0.3948}\pm0.0050\\,/\\,0.3236\pm0.0105 & \mathbf{0.5358}\pm0.0045\\,/\\,0.4719\pm0.0080 \\\\ 25\\% & \mathbf{0.6545}\pm0.0112\\,/\\,0.6274\pm0.0041 & \mathbf{0.7935}\pm0.0057\\,/\\,0.7754\pm0.0026 & \mathbf{0.4673}\pm0.0047\\,/\\,0.4145\pm0.0154 & \mathbf{0.5976}\pm0.0064\\,/\\,0.5548\pm0.0100 \\\\ 100\\% & \mathbf{0.7192}\pm0.0050\\,/\\,0.6978\pm0.0022 & \mathbf{0.8353}\pm0.0030\\,/\\,0.8223\pm0.0014 & \mathbf{0.5596}\pm0.0036\\,/\\,0.5183\pm0.0044 & \mathbf{0.6787}\pm0.0025\\,/\\,0.6477\pm0.0032 \\\\ \hline \end{array}$$ **Table 5.** Inference costs on an NVIDIA RTX 4090 for 1,000 NIST23 test structures. $$\begin{array}{lrrrrr}\hline\text{Method}&\text{Params (M)}&\text{Latency (ms)}&\text{Batch}&\text{Throughput (mol/s)}&\text{GPU memory (MiB)}\\\\\hline\text{CLEF}&26.4&25.4&8&132.2&8245.6\\\\\text{RASSP:SN}&26.3&13.2&4&150.5&1147.3\\\\\text{MassFormer}&54.6&18.3&120&171.8&566.9\\\\\text{RASSP:FN}&19.7&13.8&8&181.6&8129.1\\\\\text{MolFormer}&58.8&38.7&120&1564.5&353.1\\\\\text{NEIMS}&39.3&2.9&512&25179.2&250.3\\\\\text{Hybrid}&2.9&4.8&128&16661.6&424.8\\\\\hline\end{array}$$ Latency is measured at batch size 1; throughput uses the listed batch size. CLEF uses cached event and formula features. **Table 7.** Stepwise MoNA selection. The first nine rows count spectrum records and the final row counts consensus targets. The last column reports records removed or collapsed at that step. $$\begin{array}{lrr}\hline\text{Stage}&\text{Retained}&\text{Removed or collapsed}\\\\\hline\text{GC–MS source export}&19,136&\text{–}\\\\\text{EI, MS1 and positive-mode metadata}&16,672&2,464\\\\\text{Exclude explicit non-70 eV energy}&13,799&2,873\\\\\text{Exclude derivatization or mixture annotations}&11,205&2,594\\\\\text{Accepted license metadata}&7,844&3,361\\\\\text{Structure, component, isotope and element checks}&5,729&2,115\\\\\text{Capacity limits: N48/M400/F8192}&5,308&421\\\\\text{Spectrum quality control}&5,280&28\\\\\text{Same-identity exact-spectrum deduplication}&5,265&15\\\\\text{Consensus target construction}&3,977&1,288\\\\\hline\end{array}$$

Q4 (continued). MoNA filtering coverage, relaxed selection and stratification.

Appendix A.3, Tables 7 and 8, reports stepwise exclusions from the 19,136-record GC-MS export to 5,265 retained spectra and 3,977 consensus targets. Tables 9 and 10 give the relaxed-cohort splits and exclusions. Shared computation and stepwise-coverage revisions are dark red.

Table 8. Record-level exclusion accounting. Capacity exclusions are assigned to the first failed rule.

\hline \text{Stage} & \text{Exclusion rule} & \text{Records} \\\\ \hline \text{Acquisition metadata} & \text{CI-B / generic GC–MS / missing instrument / MS2 / negative mode / other} & \text{992 / 918 / 390 / 143 / 8 / 13} \\\\ \text{Structure checks} & \text{SMILES–InChI conflict} & 1,298 \\\\ & \text{Header identity mismatch} & 1 \\\\ & \text{Multiple components} & 5 \\\\ & \text{Isotope-labelled structures} & 133 \\\\ & \text{Unsupported elements} & 678 \\\\ \text{Capacity checks} & \text{Atom count / mass / formula count, sequential first failure} & \text{374 / 11 / 36} \\\\ \text{Spectrum checks} & \text{Nonpositive intensity} & 2 \\\\ & \text{Outside-grid total ion current} & 2 \\\\ & \text{Signal outside the precursor isotope envelope} & 24 \\\\ \text{Deduplication} & \text{Exact spectra within the same identity} & 15 \\\\ \hline \end{array}$$ Under both relaxed filters (N48/M400 and N96/M800), CLEF leads all four spectrum metrics and both $K=50$ retrieval metrics. RASSP:SN leads both $K=5$ metrics in N96/M800 (Table 11, dark orange). **Table 11.** Spectrum prediction and isomer retrieval under relaxed MoNA selection. $$\begin{array}{llrrrrrrrr}\hline&&&\text{Spectrum prediction}&&&K=5&&K=50&\\\\\text{Setting}&\text{Method}&\text{DP }\uparrow&\text{SDP }\uparrow&\text{COS }\uparrow&\text{Top_1 }\uparrow&\text{R@1 }\uparrow&\text{MRR }\uparrow&\text{R@1 }\uparrow&\text{MRR }\uparrow\\\\\hline\text{N48/M400}&\text{CLEF}&{\bf0.7940}&{\bf0.7814}&{\bf0.7315}&{\bf0.5066}&{\bf0.7477}&{\bf0.8480}&{\bf0.4848}&{\bf0.6054}\\\\&\text{RASSP:SN}&0.7619&0.7412&0.6810&0.4468&0.7451&0.8436&0.4560&0.5621\\\\&\text{MassFormer}&0.7464&0.7050&0.6219&0.3484&0.6432&0.7718&0.2704&0.4076\\\\\hline\text{N96/M800}&\text{CLEF}&{\bf0.7982}&{\bf0.7619}&{\bf0.7332}&{\bf0.4743}&0.7503&0.8620&{\bf0.4712}&{\bf0.6061}\\\\&\text{RASSP:SN}&0.7635&0.7367&0.6752&0.3653&{\bf0.7678}&{\bf0.8642}&0.4241&0.5509\\\\&\text{MassFormer}&0.7392&0.6938&0.5966&0.2731&0.6885&0.8082&0.3166&0.4427\\\\\hline\end{array}$$ On the original filtered cohort, CLEF has higher DP than RASSP:SN in all 17 strata (Table 13, dark orange). **Table 13.** MoNA molecular-property and energy-metadata strata. Record counts pool the three test splits, and scores are equal-cell means over nine split–training combinations. SN denotes RASSP:SN and MF denotes MassFormer. Bold marks the highest value among the three listed methods within each stratum. $$\begin{array}{llrrrrrrr}\hline&&&&\text{DP }\uparrow&&&\text{SDP }\uparrow&\\\\\text{Axis}&\text{Stratum}&\text{Records}&\text{CLEF}&\text{SN}&\text{MF}&\text{CLEF}&\text{SN}&\text{MF}\\\\\hline\text{Mass (Da)}&\lt 200&783&{\bf0.7927}&0.7749&0.7716&{\bf0.7601}&0.7394&0.7232\\\\&200\text{–}\lt 400&404&{\bf0.7342}&0.6854&0.6522&{\bf0.7244}&0.6767&0.6204\\\\\hline\text{Heavy atoms}&\leq10&392&{\bf0.7775}&0.7629&0.7654&{\bf0.7585}&0.7487&0.7462\\\\&\text{11–15}&528&{\bf0.8005}&0.7735&0.7589&{\bf0.7545}&0.7190&0.6830\\\\&\geq16&267&{\bf0.7133}&0.6618&0.6265&{\bf0.7185}&0.6707&0.6123\\\\\hline\text{Rings}&0&460&{\bf0.8558}&0.8261&0.8042&{\bf0.7778}&0.7185&0.6872\\\\&1&292&{\bf0.7051}&0.6921&0.6950&0.6629&0.6607&{\bf0.6644}\\\\&\geq2&435&{\bf0.7280}&0.6914&0.6738&{\bf0.7728}&0.7549&0.7058\\\\\hline\text{Rotatable bonds}&0&162&0.7478&0.7268&{\bf0.7674}&0.7802&{\bf0.7919}&0.7797\\\\&\text{1–3}&369&{\bf0.7364}&0.7155&0.7040&{\bf0.7599}&0.7524&0.7192\\\\&\geq4&656&{\bf0.7995}&0.7652&0.7372&{\bf0.7333}&0.6809&0.6485\\\\\hline\text{Formula count}&\leq255&226&0.7616&0.7452&{\bf0.7724}&0.7767&0.7816&{\bf0.7856}\\\\&\text{256–1,023}&676&{\bf0.8000}&0.7735&0.7619&{\bf0.7606}&0.7279&0.6988\\\\&\geq1{,}024&285&{\bf0.7177}&0.6754&0.6244&{\bf0.6964}&0.6457&0.5866\\\\\hline\text{Energy metadata}&\text{All report 70 eV}&1,071&{\bf0.7829}&0.7551&0.7419&{\bf0.7599}&0.7276&0.6984\\\\&\text{All unreported}&80&{\bf0.6438}&0.5924&0.5761&{\bf0.5811}&0.5543&0.5128\\\\&\text{Mixed member metadata}&36&{\bf0.7656}&0.7652&0.7543&0.7471&{\bf0.7771}&0.7472\\\\\hline\end{array}$$
Reviewer_vXC6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CLEF, an EI-MS spectrum predictor built on RASSP-style formula scoring. It enumerates cleavage events by cutting one to three heavy-atom bonds with hydrogen shifts, encodes these events, and lets each candidate formula attend only to events whose elemental composition matches it exactly. Training uses spectrum-level supervision only. On NIST23 and MoNA datasets, CLEF achieves the best mean scores on spectral similarity metrics and on a newly proposed same-formula isomer retrieval task.

Critical strengths
  • Clear and well-motivated design. Composition-matched sparse attention links candidate formulas to structural fragmentation evidence without requiring fragment annotations, and the method is easy to follow
  • Consistent gains on large-scale data. On NIST23, CLEF improves DP from 0.839 for RASSP:SN to 0.874 and cross-source retrieval R@1 from 0.518 to 0.560.The structural-group splits and shared candidate pools make these comparisons more informative than a random-split evaluation.
  • Rigorous evaluation protocol. The grouped splits also handle acyclic molecules, overlapping compounds are removed for SWGDRUG transfer, and candidate lists are frozen and shared across methods. The isomer retrieval task is itself a useful contribution to evaluating EI predictors.
Critical weaknesses
  • Limited scope and dataset coverage The proposed method is evaluated exclusively on 70 eV electron-ionization spectra. Its applicability to tandem mass spectrometry (MS/MS) remains unexplored, and existing large-scale public benchmarks such as MassSpecGym and NPLIB1 are not used to examine its potential for extension. Meanwhile, the authors adapt MassFormer, originally developed for MS/MS, as an EI baseline, but do not explain whether CLEF’s core cleavage-event modeling relies on assumptions specific to EI or identify the concrete obstacles to extending it to MS/MS. The current experiments therefore establish effectiveness under a particular spectrometric setting while leaving the broader utility of the modeling approach unresolved.

  • Insufficient justification of data selection The MoNA experiments ultimately use only 3,977 molecular structures. Although the appendix lists the filtering criteria, the authors do not sufficiently justify the necessity of these restrictions or analyze how these choices affect the results. This uncertainty about dataset coverage weakens the empirical evidence for the model’s generalization.

  • Missing sensitivity analysis of the bond-cut limit Candidate events are generated by cutting one to three bonds, but the choice of this upper limit is neither justified nor evaluated. Moreover, the method enumerates many candidate events and formulas but provides no comparison of training and inference costs against the baselines, leaving the trade-off between performance gains and computational expense unclear.

Questions for authors
  • Which components of CLEF rely on assumptions specific to EI? Could the authors evaluate CLEF's applicability on MassSpecGym or NPLIB1?
  • Could the authors report the stepwise sample counts from the original MoNA data to the final 3,977 structures and justify the filtering thresholds?
  • Have the authors assessed how data filtering affects the results? Do CLEF’s gains over the baselines persist when key restrictions are relaxed or performance is stratified by molecular size and complexity?
  • How does performance change when the maximum number of bond cuts is set to one, two, or four? With the candidate-event budget controlled, how do fewer or more cuts affect spectrum prediction and isomer retrieval?
  • How do CLEF’s training time, per-molecule inference time, and memory requirements compare with the main baselines?
5 comments

Thank you for the detailed questions. Revisions specific to this review are highlighted in dark orange. Shared revisions are highlighted in dark red.

Q1. EI assumptions and applicability to MassSpecGym and NPLIB1.

Graph encoding and composition matching operate on molecular structure and elemental counts. The original configuration targets 70 eV EI with a neutral-molecule candidate inventory and nominal-mass rendering (Sections 3.1 to 3.3). We extended the ion inventory to [M+H]+[\mathrm{M}+\mathrm{H}]^+ and trained CLEF-MS/MS on both requested datasets. Section 4.2 and Table 3 report mean COS/DP of 0.5985/0.6697 on NPLIB1, exceeding MassFormer, ICEBERG2.1 and SCARF on both metrics. On MassSpecGym, CLEF-MS/MS reaches 0.4952/0.5518, compared with 0.3177/0.4244 for MassFormer. Appendix A.4 and Tables 14 to 16 give the adaptations, filtering and training/validation/test counts of 2,676/312/330 for NPLIB1 and 59,939/1,001/1,055 for MassSpecGym. The results support applicability under the evaluated protonated MS/MS settings.

Table 3. MS/MS spectrum prediction on a common 1 Da grid with 512 bins. Values report means and sample SD over three training runs. COS uses raw intensities and DP uses square-root intensities.

\hline \text{Dataset} & \text{Method} & \text{Cosine (COS) }\uparrow & \text{Square-root cosine (DP) }\uparrow \\\\ \hline \text{NPLIB1} & \text{CLEF-MS/MS} & \mathbf{0.5985}\pm0.0037 & \mathbf{0.6697}\pm0.0033 \\\\ & \text{MassFormer} & 0.4734\pm0.0067 & 0.6194\pm0.0022 \\\\ & \text{ICEBERG2.1} & 0.5301\pm0.0245 & 0.5536\pm0.0133 \\\\ & \text{SCARF} & 0.5028\pm0.0065 & 0.5846\pm0.0033 \\\\ \hline \text{MassSpecGym} & \text{CLEF-MS/MS} & \mathbf{0.4952}\pm0.0200 & \mathbf{0.5518}\pm0.0144 \\\\ & \text{MassFormer} & 0.3177\pm0.0074 & 0.4244\pm0.0091 \\\\ \hline \end{array}$$ **Table 14.** Frozen MS/MS data after task-specific filtering. Independent structures are source connectivity keys for MassSpecGym and full InChIKeys for NPLIB1. $$\begin{array}{llrrr} \hline \text{Dataset} & \text{Count} & \text{Train} & \text{Validation} & \text{Test} \\\\ \hline \text{MassSpecGym} & \text{Spectra} & 59,939 & 1,001 & 1,055 \\\\ & \text{Independent structures} & 5,746 & 306 & 295 \\\\ \text{NPLIB1} & \text{Merged targets} & 2,676 & 312 & 330 \\\\ & \text{Independent structures} & 2,304 & 268 & 283 \\\\ \hline \end{array}$$ **Table 15.** Successive MassSpecGym selection counts within the official simulation folds. The final ion check removes four training records. $$\begin{array}{lrrr} \hline \text{Selection stage} & \text{Train} & \text{Validation} & \text{Test} \\\\ \hline \text{Official simulation fold} & 99,341 & 9,734 & 9,954 \\\\ \text{Supported elements} & 97,323 & 9,435 & 9,652 \\\\ \text{Neutral formal charge} & 97,302 & 9,432 & 9,652 \\\\ \text{Single component, no isotope label} & 97,302 & 9,432 & 9,652 \\\\ \text{Neutral exact mass }\leq400\text{ Da} & 79,465 & 3,875 & 3,280 \\\\ \text{Explicit-H graph atoms }\leq48 & 68,317 & 2,973 & 2,408 \\\\ \text{Neutral formula inventory }\leq8,192 & 60,508 & 1,133 & 1,110 \\\\ \text{H+1 formula inventory }\leq8,192 & 59,954 & 1,001 & 1,055 \\\\ \text{Source-formula/adduct compatibility} & 59,943 & 1,001 & 1,055 \\\\ \text{Precursor/formula agreement within 0.01 Da} & 59,939 & 1,001 & 1,055 \\\\ \hline \end{array}$$ **Table 16.** NPLIB1 selection within the published `Fold_0` assignment. $$\begin{array}{lrrr} \hline \text{Selection stage} & \text{Train} & \text{Validation} & \text{Test} \\\\ \hline \text{Source identifiers with fold assignment} & 8,680 & 989 & 1,040 \\\\ \text{Parsed targets and sanitized structures} & 6,492 & 738 & 800 \\\\ \text{Common structure and ion rules} & 2,677 & 313 & 330 \\\\ \text{Precursor/formula agreement within 0.01 Da} & 2,676 & 312 & 330 \\\\ \hline \end{array}$$

Q2. Stepwise MoNA filtering counts and the rationale for the thresholds.

Appendix A.3 now provides the complete selection ledger and exclusion accounting (Tables 7 and 8, dark red). Table 7 gives the count after each step, from 19,136 to 5,265 spectrum records. The retained spectra form 3,977 full-InChIKey consensus targets, representing 3,973 connectivities. Acquisition and structure checks align the targets with EI conditions and molecular inputs. Limits of 48 explicit-H atoms, 400 Da and 8,192 candidate formulas bound the model's input and enumeration sizes. Replicate measurements are combined into consensus targets. The effects of the selection choices are evaluated separately below.

Table 7. Stepwise MoNA selection. The first nine rows count spectrum records and the final row counts consensus targets. The last column reports records removed or collapsed at that step.

\hline \text{Stage} & \text{Retained} & \text{Removed or collapsed} \\\\ \hline \text{GC–MS source export} & 19,136 & \text{–} \\\\ \text{EI, MS1 and positive-mode metadata} & 16,672 & 2,464 \\\\ \text{Exclude explicit non-70 eV energy} & 13,799 & 2,873 \\\\ \text{Exclude derivatization or mixture annotations} & 11,205 & 2,594 \\\\ \text{Accepted license metadata} & 7,844 & 3,361 \\\\ \text{Structure, component, isotope and element checks} & 5,729 & 2,115 \\\\ \text{Capacity limits: N48/M400/F8192} & 5,308 & 421 \\\\ \text{Spectrum quality control} & 5,280 & 28 \\\\ \text{Same-identity exact-spectrum deduplication} & 5,265 & 15 \\\\ \text{Consensus target construction} & 3,977 & 1,288 \\\\ \hline \end{array}$$ **Table 8.** Record-level exclusion accounting. Capacity exclusions are assigned to the first failed rule. $$\begin{array}{llr} \hline \text{Stage} & \text{Exclusion rule} & \text{Records} \\\\ \hline \text{Acquisition metadata} & \text{CI-B / generic GC–MS / missing instrument / MS2 / negative mode / other} & \text{992 / 918 / 390 / 143 / 8 / 13} \\\\ \text{Structure checks} & \text{SMILES–InChI conflict} & 1,298 \\\\ & \text{Header identity mismatch} & 1 \\\\ & \text{Multiple components} & 5 \\\\ & \text{Isotope-labelled structures} & 133 \\\\ & \text{Unsupported elements} & 678 \\\\ \text{Capacity checks} & \text{Atom count / mass / formula count, sequential first failure} & \text{374 / 11 / 36} \\\\ \text{Spectrum checks} & \text{Nonpositive intensity} & 2 \\\\ & \text{Outside-grid total ion current} & 2 \\\\ & \text{Signal outside the precursor isotope envelope} & 24 \\\\ \text{Deduplication} & \text{Exact spectra within the same identity} & 15 \\\\ \hline \end{array}$$

Q3. Effects of relaxed filtering and molecular-size/complexity stratification.

We removed the license and derivatization whitelists and the replicate-consensus requirement. N48/M400 and N96/M800 retain 5,056 and 5,615 structures (Appendix A.3, Tables 9–10). CLEF leads all spectrum and K=50K=50 metrics in both cohorts. At K=5K=5, CLEF leads in N48/M400 and RASSP:SN leads in N96/M800 (Table 11). Table 12 stratifies N96/M800 by mass. CLEF exceeds RASSP:SN in DP in all 17 original-cohort strata. MassFormer leads on DP in the zero-rotatable-bond and smallest-formula-count strata (Table 13).

Table 9. Split sizes of the relaxed MoNA cohorts. Structures are distinct full InChIKeys.

SettingUnitPoolRetainedTrainVal.TestN48/M400Records16,9098,9467,445749752Structures–5,0564,046504506N96/M800Records18,9899,7428,083824835Structures–5,6154,496560559\begin{array}{llrrrrr}\hline\text{Setting}&\text{Unit}&\text{Pool}&\text{Retained}&\text{Train}&\text{Val.}&\text{Test}\\\\\hline\text{N48/M400}&\text{Records}&16,909&8,946&7,445&749&752\\\\&\text{Structures}&\text{–}&5,056&4,046&504&506\\\\\hline\text{N96/M800}&\text{Records}&18,989&9,742&8,083&824&835\\\\&\text{Structures}&\text{–}&5,615&4,496&560&559\\\\\hline\end{array}

Table 11. Spectrum prediction and isomer retrieval under relaxed MoNA selection.

\begin{array}{llrrrrrrrr}\hline&&&\text{Spectrum prediction}&&&K=5&&K=50&\\\\\text{Setting}&\text{Method}&\text{DP }\uparrow&\text{SDP }\uparrow&\text{COS }\uparrow&\text{Top_1 }\uparrow&\text{R@1 }\uparrow&\text{MRR }\uparrow&\text{R@1 }\uparrow&\text{MRR }\uparrow\\\\\hline\text{N48/M400}&\text{CLEF}&{\bf0.7940}&{\bf0.7814}&{\bf0.7315}&{\bf0.5066}&{\bf0.7477}&{\bf0.8480}&{\bf0.4848}&{\bf0.6054}\\\\&\text{RASSP:SN}&0.7619&0.7412&0.6810&0.4468&0.7451&0.8436&0.4560&0.5621\\\\&\text{MassFormer}&0.7464&0.7050&0.6219&0.3484&0.6432&0.7718&0.2704&0.4076\\\\\hline\text{N96/M800}&\text{CLEF}&{\bf0.7982}&{\bf0.7619}&{\bf0.7332}&{\bf0.4743}&0.7503&0.8620&{\bf0.4712}&{\bf0.6061}\\\\&\text{RASSP:SN}&0.7635&0.7367&0.6752&0.3653&{\bf0.7678}&{\bf0.8642}&0.4241&0.5509\\\\&\text{MassFormer}&0.7392&0.6938&0.5966&0.2731&0.6885&0.8082&0.3166&0.4427\\\\\hline\end{array}

Table 12. Mass strata of the N96/M800 test cohort. DP averages spectrum records within each band. K=50K=50 R@1 retains equal-connectivity aggregation and square-root-intensity cosine scoring. SN and MF denote RASSP:SN and MassFormer. Dashes indicate no spectra or eligible retrieval queries in that stratum.

[200,400)&302&230&{\bf0.7649}&0.7090&0.7019&{\bf0.4707}&0.3880&0.3560\\\\ [400,600)&12&8&{\bf0.7419}&0.7142&0.6264&\text{–}&\text{–}&\text{–}\\\\ [600,800]&0&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\\hline\end{array}$$ **Table 13.** MoNA molecular-property and energy-metadata strata. Record counts pool the three test splits, and scores are equal-cell means over nine split–training combinations. SN denotes RASSP:SN and MF denotes MassFormer. Bold marks the highest value among the three listed methods within each stratum. $$\begin{array}{llrrrrrrr}\hline&&&&\text{DP }\uparrow&&&\text{SDP }\uparrow&\\\\\text{Axis}&\text{Stratum}&\text{Records}&\text{CLEF}&\text{SN}&\text{MF}&\text{CLEF}&\text{SN}&\text{MF}\\\\\hline\text{Mass (Da)}&\lt 200&783&{\bf0.7927}&0.7749&0.7716&{\bf0.7601}&0.7394&0.7232\\\\&200\text{–}\lt 400&404&{\bf0.7342}&0.6854&0.6522&{\bf0.7244}&0.6767&0.6204\\\\\hline\text{Heavy atoms}&\leq10&392&{\bf0.7775}&0.7629&0.7654&{\bf0.7585}&0.7487&0.7462\\\\&\text{11–15}&528&{\bf0.8005}&0.7735&0.7589&{\bf0.7545}&0.7190&0.6830\\\\&\geq16&267&{\bf0.7133}&0.6618&0.6265&{\bf0.7185}&0.6707&0.6123\\\\\hline\text{Rings}&0&460&{\bf0.8558}&0.8261&0.8042&{\bf0.7778}&0.7185&0.6872\\\\&1&292&{\bf0.7051}&0.6921&0.6950&0.6629&0.6607&{\bf0.6644}\\\\&\geq2&435&{\bf0.7280}&0.6914&0.6738&{\bf0.7728}&0.7549&0.7058\\\\\hline\text{Rotatable bonds}&0&162&0.7478&0.7268&{\bf0.7674}&0.7802&{\bf0.7919}&0.7797\\\\&\text{1–3}&369&{\bf0.7364}&0.7155&0.7040&{\bf0.7599}&0.7524&0.7192\\\\&\geq4&656&{\bf0.7995}&0.7652&0.7372&{\bf0.7333}&0.6809&0.6485\\\\\hline\text{Formula count}&\leq255&226&0.7616&0.7452&{\bf0.7724}&0.7767&0.7816&{\bf0.7856}\\\\&\text{256–1,023}&676&{\bf0.8000}&0.7735&0.7619&{\bf0.7606}&0.7279&0.6988\\\\&\geq1{,}024&285&{\bf0.7177}&0.6754&0.6244&{\bf0.6964}&0.6457&0.5866\\\\\hline\text{Energy metadata}&\text{All report 70 eV}&1,071&{\bf0.7829}&0.7551&0.7419&{\bf0.7599}&0.7276&0.6984\\\\&\text{All unreported}&80&{\bf0.6438}&0.5924&0.5761&{\bf0.5811}&0.5543&0.5128\\\\&\text{Mixed member metadata}&36&{\bf0.7656}&0.7652&0.7543&0.7471&{\bf0.7771}&0.7472\\\\\hline\end{array}$$
Reviewer_6LrJOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Summary of the paper: The paper proposes CLEF, a model that predicts mass spectra from a molecular graph. It mainly uses a relational GCN to produce atom states. Candidate "cleavage events" are enumerated deterministically from the structure by cutting one to three heavy-atom bonds. These events are also sent into the model. The proposed method achieves the best performance across all datasets.

Critical strengths

CLEF has the highest mean scores among the six trainable baselines on all four spectrum metrics.

Critical weaknesses

The "Graph-cut enumeration" and "Encoding each event" paragraphs depend too heavily on Figure 2 and are very hard to understand.

The paper says the input is the molecular graph, but where do the events come from? I hope to see a formal description.

Where do the candidate fragments come from? Where do the events come from? In "Its descriptor" (L211), whose descriptor is meant? What is the formula token f_j, and what does the index j range over?

Tokens are a concept for sequential data, such as the inputs to LLMs. What does "token" mean here?

The contribution in methodology is incremental. Graph neural networks have been applied to this task. Enumerating fragments by breaking bonds is also well established in prior work.

Questions for authors

I hope the authors can address the issues raised in the section of critical weaknesses.

3 comments

Thank you for the detailed comments on the presentation and methodological positioning. We have revised the object definitions, event construction, terminology, and Figure 2. Please see the blue revisions in Sections 3.2 and 3.3 and Figure 2. The methodological positioning shared with another review is highlighted in dark red in Section 2.

Q1. Where do the cleavage events and candidate fragments come from?

The events are constructed deterministically from the input molecular graph. Section 3.2, "Graph-cut enumeration," now defines the heavy-atom bond set and enumerated cut sets SS with 1≤∣S∣≤31\leq \lvert S\rvert\leq 3. Removing these bonds yields connected atom components. Each retained component, its cut context, and a hydrogen-count adjustment define a candidate event. The unchanged molecular ion is added separately. This construction is described in the text and summarized in Figure 2b. Appendix B.1 specifies filtering, deduplication, and enumeration budgets.

Q2. What is the event descriptor, and how does it relate to candidate formulas?

Section 3.2 now distinguishes the two objects. A candidate formula is an element-count vector within the precursor composition. A cleavage event associates such a composition with a particular atom component and cut context. Several events can therefore match one formula. In "Encoding each event," the descriptor of event ee is the 45-dimensional vector de\mathbf{d}_e, which records its event type, cut count, hydrogen-count change, compositions, mass, and cut-bond properties. Equation (2) shows how the event descriptor is combined with global atom context and mean/max atom states at the cut boundary to produce its learned embedding.

Q3. What do fj\mathbf{f}_j and jj mean, and why was the word "token" used?

Section 3.2 defines the candidate-formula set F(c)\mathcal{F}(\mathbf{c}) containing J=∣F(c)∣J=\lvert\mathcal{F}(\mathbf{c})\rvert element-count vectors cj\mathbf{c}_j. Section 3.3 defines fj\mathbf{f}_j as the learned embedding of candidate formula jj, obtained by applying a linear layer to its element-count encoding. The index thus ranges over j=1,…,Jj=1,\ldots,J. We have replaced "token" with "formula embedding" or "event embedding," as appropriate, in the text and main diagram. Equations (4) to (6) specify how matching events update the formula embedding and how that embedding guides atom pooling.

Q4. What is the methodological distinction from existing graph and fragmentation models?

Graph neural encoding and bond-cut enumeration are established components. CLEF couples the resulting cleavage evidence to formula scoring through composition-matched cross-attention. Cleavage descriptors and local atom states encode alternative contexts for the same formula. Their attention-weighted combination updates its representation before formula-specific atom pooling and scoring, with joint spectrum-level supervision. Section 2 now explains this computation alongside the fragment-generation/scoring approach of ICEBERG, the fragment-node/formula probability model of FraGNNet, and the formula-generation approach of SCARF. Section 4.3 and Table 4(a,b) further document the effects of event conditioning and descriptor alignment. The methodological discussion in Section 2 is highlighted in dark red.

Table 4. NIST23 ablations. Bold marks the best value in each block.

\hline & & \text{Spectrum prediction} & & & \substack{\text{Same-source}\\\K=5} & & \substack{\text{Cross-source}\\\K=50} & \\\\ \text{Variant} & \text{DP} & \text{SDP} & \text{COS} & \text{Top_1} & \text{R@1} & \text{MRR} & \text{R@1} & \text{MRR} \\\\ \hline \rlap{\text{(a) Conditioning branches}} & & & & & & & & \\\\ \hline \text{Graph} & 0.7691 & 0.5051 & 0.6303 & 0.3463 & 0.4814 & 0.6806 & 0.1832 & 0.3134 \\\\ \text{Graph + formula} & 0.8091 & 0.7437 & 0.7001 & 0.4401 & 0.5492 & 0.7240 & 0.2845 & 0.4277 \\\\ \text{CLEF} & \mathbf{0.8729} & \mathbf{0.8758} & \mathbf{0.8147} & \mathbf{0.5782} & \mathbf{0.7386} & \mathbf{0.8477} & \mathbf{0.5567} & \mathbf{0.6762} \\\\ \hline \rlap{\text{(b) Event-descriptor alignment}} & & & & & & & & \\\\ \hline \text{CLEF} & \mathbf{0.8729} & \mathbf{0.8758} & \mathbf{0.8147} & \mathbf{0.5782} & \mathbf{0.7386} & \mathbf{0.8477} & \mathbf{0.5567} & \mathbf{0.6762} \\\\ \text{Shuffled descriptors} & 0.8577 & 0.8525 & 0.7887 & 0.5464 & 0.6945 & 0.8202 & 0.4824 & 0.6166 \\\\ \text{Formula–graph null} & 0.8072 & 0.7471 & 0.6965 & 0.4313 & 0.5483 & 0.7228 & 0.2920 & 0.4319 \\\\ \hline \rlap{\text{(c) Event aggregation}} & & & & & & & & \\\\ \hline \text{CLEF} & \mathbf{0.8729} & \mathbf{0.8758} & \mathbf{0.8147} & \mathbf{0.5782} & \mathbf{0.7386} & \mathbf{0.8477} & \mathbf{0.5567} & \mathbf{0.6762} \\\\ \text{Composition uniform} & 0.8700 & 0.8709 & 0.8097 & 0.5738 & 0.7293 & 0.8418 & 0.5432 & 0.6642 \\\\ \text{Full attention} & 0.7720 & 0.6535 & 0.6268 & 0.3466 & 0.4584 & 0.6619 & 0.1848 & 0.3102 \\\\ \hline \end{array}$$ K=5 denotes same-source retrieval and K=50 cross-source retrieval. Bold marks the highest values.
Reviewer_6LrJOpenReview ↗

Thank you for your clarification

Thank you for your responses. The new version certainly shows improvement. However, I would still like to see stronger methodological innovations, which are currently limited in this submission.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for your careful reviews and constructive questions. The additional experiments took longer than expected with the computing resources available to us, so we posted our responses toward the end of the discussion period. Your feedback helped us clarify how CLEF uses cleavage events and evaluate the method under a wider range of settings. We appreciated the opportunity to do that work.

As you finalize your assessments, we would be grateful if you could take our responses and the additional results in the individual review threads into account. Thank you again for your time and consideration.

The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.