acceptodds
Under review as a conference paper at ICLR 2027

Anchora: Retrieval-Guided AEMS Completion for Molecular Structure Elucidation from Tandem Mass Spectra

Abstract

Inferring molecular structure from tandem mass spectra is difficult because isomers can share same formula and similar fragments. Structure-database retrieval is applicable when the target or a close analogue is covered, but depends strongly on library coverage and cannot directly recover an out-of-library target. De novo generation can propose unseen molecules, yet constraining an entire molecular graph from sparse spectral evidence creates a large search space and can cause structural drift. To combine retrieval reliability with generative reach, we introduce Anchora, a retrieval-augmented generation framework based on an anchor-extended molecular scaffold (AEMS), which explicitly encodes side-chain anchoring points. Given a spectrum and formula, Anchora ranks formula-compatible complete molecules, extracts AEMS from the top reference, and represents it as a canonical AEMS–SAFE prefix for side-chain completion by a pretrained discrete diffusion model. Large-scale completion pretraining and spectrum–decoder adaptation support generation, while three ICEBERG-guided refinement rounds edit complete molecules without preserving the prefix, providing a spectrum-guided correction route beyond an erroneous prior. Reference-guided top-1 exact accuracy reaches 38.98% on CANOPUS and 23.50% on MassSpecGym, exceeding the best baseline by 13.95 and 5.21 points, respectively. Zero-shot evaluation further shows that Anchora-Base improves top-1 accuracy over FRIGID-base by 5.10–27.14 percentage points, supporting robustness to distribution shifts. These results prove that Anchora bridges chemical-structure retrieval and de novo molecular generation for MS/MS structure elucidation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
tandem mass spectrometry, molecular structure elucidation, retrieval-augmented generation, molecular generation, discrete diffusion models, scaffold-conditioned generation, molecular fingerprints

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tUMNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents ANCHORA, a retrieval-guided framework for molecular structure elucidation from MS/MS spectra with a known molecular formula. It retrieves a formula-compatible reference molecule, converts its anchor-extended molecular scaffold into an AEMS-SAFE prefix, completes the side chains with a conditional discrete diffusion model. It then subsequently applies ICEBERG-guided full-molecule refinement that can edit beyond the retrieved prefix. The paper reports substantial improvements over reported baselines on CANOPUS and MassSpecGym, as well as zero-shot transfer results on five external GNPS libraries.

Critical strengths
  • The core design is sensible and targets a real limitation of both retrieval-only and unconstrained de novo methods. In particular, the distinction between a bare Murcko scaffold and an AEMS that encodes side-chain attachment sites is meaningful for molecular completion.
  • The representation contribution is more than a superficial serialization choice. It explicitly recognizes that canonicalization can permute symmetry-equivalent anchors, and proposes selecting only scaffold-to-canonical isomorphism mappings whose reconstructed SAFE string decodes to the input canonical molecular graph.
  • The reported empirical improvements in Table 1 are large. Under reference priors, Anchora reaches 38.98% Acc@1 on CANOPUS and 23.50% on MassSpecGym, compared with 25.03% and 18.29% for FRIGID, respectively. The corresponding improvements are not limited to exact matching: Sim and MCES also improve at both Top-1 and Top-10. The oracle-prior results additionally establish that the completion model can make strong use of a correct AEMS, rather than merely benefiting from ranking a complete retrieved molecule.
  • The paper does make a useful effort to separate pipeline components. Table 3 shows complementary benefits from large-scale AEMS completion pretraining and spectrum-decoder joint adaptation. Table 6 also provides a focused diagnostic showing that predicting completion length rather than full sequence length improves generation-length RMSE from 5.33 to 3.18 and increases AEMS pass rate from 62.92% to 74.99%.
  • The external-library evaluation is a positive addition. Table 2 consistently favors Anchora-Base over FRIGID-base across all five GNPS libraries, including difficult MIADB and IQAMDB subsets where FRIGID-base has nearly no exact recovery. This is useful evidence that the pipeline is not confined to the two primary benchmarks, although the interpretation needs stronger controls as discussed below.
Critical weaknesses
  • The central empirical comparison does not isolate the effect of retrieval augmentation, and therefore cannot yet support the broad claim that the proposed generation method itself is responsible for the headline gains. ANCHORA has access to a 672,681 structure external catalogue, retrieves a formula-compatible reference using a MIST-derived fingerprint, and then uses a highly informative AEMS extracted from the top molecule. In contrast, the listed generative baselines in Table 1 are reported as de novo methods without this retrieval channel.
  • The paper reports AEMS accuracy of 59.78% on CANOPUS and 43.27% on MassSpecGym, which already indicates that retrieval provides a highly informative structural prior. However, it does not report: (i) the Top-1/Top-K exact retrieval accuracy of the complete molecule, (ii) the fraction of test targets contained in the retrieval database, (iii) the performance of simply returning d*, (iv) the performance of ranking retrieved complete candidates directly by Equation 2, or (v) the accuracy conditional on whether the reference AEMS is correct. These are essential controls. Without them, the reader cannot determine whether the main improvement comes from AEMS completion, the MIST-based retrieval score, direct target coverage in the external database, or their combination.
  • The paper’s de novo and out-of-library framing is too broad relative to the evaluation protocol. The paper argues that the framework can recover an out-of-library target because it generates a completion rather than returning a database entry. That is plausible in principle, but the presented experiments do not distinguish targets absent from DD from targets already present in DD. This distinction matters because a retrieved target molecule has the exact target AEMS by construction, making the subsequent completion problem much easier.
  • The full-molecule refinement method in Section 3.6 is underspecified, yet it is a central claimed mechanism. The text states that discrepancies between ICEBERG-predicted and observed spectra guide masking, but does not define the discrepancy function, fragment or peak matching procedure, masking unit, mask size, thresholding, handling of unmatched peaks, or how the selected molecular regions are mapped back to tokens. It is similarly unclear which general full-molecule diffusion model is used, whether it shares parameters with the completion model, what its training distribution is, and how the K=16, M=4 edit variants are chosen.
  • The refinement ablation is confounded with a substantially larger inference budget. Section 4.3 states that Anchora-Base uses 128 initial attempts, while each of three refinement rounds adds 64 completions and up to 64 full-molecule edits, for up to 512 attempts total. Table 4 shows improvements over rounds, but it does not establish whether these gains derive from ICEBERG-guided editing rather than simply generating more candidates. This is especially important because the CANOPUS reference setting reaches 38.98% after Round 2 and does not improve in Round 3.
Questions for authors

See the weaknesses above. And could the authors please mention if they perform the data contamination check (from the dataset training perspective). Or equivalently if the authors could directly test their method on MassSpecGym v1.5 (follow the instructions). If you have done such, please mention that in the paper and this would no longer be the question from my side.

5 comments

Response to Reviewer tUMN

We sincerely thank the reviewer for carefully analyzing attachment-aware priors, symmetry-equivalent anchor mapping, and the training and length ablations. Your questions on system-level attribution, retrieval controls, and refinement implementation are very specific. Below, we distinguish complete-system performance, completion under partial structural conditions, and direct retrieval accuracy.

W1——The contributions of retrieval and generation have not been fully isolated

The external structure database supplies additional information, so the improvements in Table 1 cannot be directly attributed to the generative framework itself.

Response. Thank you for raising the attribution question. Introducing AEMS priors through an external structure database and completing unknown structures from spectra is Anchora's core design. The main experiments, zero-shot evaluation, and ablations center on this design, testing whether introducing AEMS priors improves structure elucidation. Retrieval is intended to contribute to the gains. Table 1 evaluates the complete system rather than attributing all improvements to the decoder.

For the shared-resource comparison, please see the "MADGEN (CSU-MS2) versus Anchora-Base (Reference)" table in our response to Reviewer FBze W2. It reports prior quality and six generation metrics using the same CSU-MS2 source database.

W2——Five controls to separate retrieval, database coverage, and completion

Please report: (i) full-molecule Top-1/Top-K retrieval accuracy; (ii) target database membership; (iii) directly returning d*; (iv) ranking by Equation 2; and (v) accuracy stratified by whether the reference AEMS is correct.

Response. Thank you for these five specific questions. The results are as follows:

  1. Full-molecule Top-1/Top-K retrieval accuracy: We recomputed the complete ranking using the same retrieval rules. Results are below; matches require molecular-graph identity after stereochemistry removal, without merging tautomers.
DatasetTop-1 (%)Top-5 (%)Top-10 (%)
CANOPUS52.5067.2872.16
MSG40.8151.5255.72

From Top-1 to Top-10, accuracy increases by 19.66/14.91 percentage points on the two datasets. Retrieval therefore concentrates more correct structures near the top of the ranking, supporting the use of multiple candidates as structural priors.

  1. Database membership: 81.44% for CANOPUS and 57.35% for MSG.
  2. Directly returning d:* Returning the first molecule in the ranking, without completion or refinement, gives target-molecule recovery accuracy of 52.50% on CANOPUS and 40.81% on MSG.
  3. Ranking by Equation 2: Ranking candidates using this equation gives Top-1 accuracy of 52.50% on CANOPUS and 40.81% on MSG. Its first candidate is d*, so this is identical to directly returning d*.
  4. Prior-conditional accuracy: Target-molecule recovery is grouped by whether the reference AEMS is correct.
DatasetReference AEMSBase exact Top-1 (%)Refined exact Top-1 (%)
CANOPUSCorrect64.7968.85
CANOPUSIncorrect3.023.36
MSGCorrect46.8251.01
MSGIncorrect2.943.03

Correct AEMS priors correspond to higher target-molecule recovery. Refinement further improves accuracy by 4.06 and 4.19 percentage points on CANOPUS and MSG, respectively; both datasets also show small improvements with incorrect priors.

W3——Is the claim of out-of-library generation supported?

The ability to generate a completion does not establish recovery of out-of-library targets; targets inside and outside the library should be distinguished.

Response. Thank you for suggesting this distinction. Anchora extracts AEMS from the retrieved molecule as a local structural prior rather than directly returning that molecule, then completes the remaining structure from the spectrum. Its final output therefore need not belong to the database. We further stratify all 803/17,082 CANOPUS/MSG test spectra. Database membership and molecular recovery Acc use the same stereochemistry removal and tautomer normalization; both models below use the Reference setting.

DatasetTarget in retrieval database?Anchora-Base Acc@1/10 (%)Anchora Acc@1/10 (%)
CANOPUSIn database42.59 / 46.0445.04 / 48.20
CANOPUSOutside database1.03 / 1.091.06 / 1.10
MSGIn database32.98 / 35.0035.93 / 38.16
MSGOutside database0.80 / 0.840.65 / 0.89

Both out-of-library subsets achieve nonzero exact target recovery, showing that Anchora's final outputs are not restricted to retrieval database entries. Overall out-of-library accuracy remains lower than in-library accuracy, indicating that database coverage remains an important factor.

W4——Implementation details of full-molecule refinement

Please specify the discrepancy function, peak/fragment matching, mask units and sizes, thresholds, unmatched peaks, token mapping, the editor's training source and relationship to completion, and the meanings of K=16 and M=4.

Response. Thank you for considering the implementation details. Full-molecule editing follows FRIGID[1]'s ICEBERG-guided refinement workflow, with the following settings (see Section 3.6 of the revised manuscript, lines 282–305):

  1. Discrepancy and peak matching: Intensities on each side are independently normalized by their maximum, with a threshold of 0.05. We select up to 3 peaks with the largest distance from a predicted peak to its nearest experimental peak, d(m)=min⁡x∣m−x∣d(m)=\min_x|m-x|, without an additional distance threshold. 20 ppm is used only to exclude the predicted precursor peak; experimental peaks without associated predicted fragments do not directly trigger editing.
  2. Masking and mapping: The fragment IDs of selected peaks define a union of atoms, which is mapped to SMILES character spans and then to tokens through tokenizer offsets. Masking is token-wise and independent, with probability 0.5, so mask size varies. Mapping failure triggers full-sequence masking; missing predicted spectra skip editing for that candidate without removing the query.
  3. Model and training data: The editor is the frozen FRIGID-base full-molecule masked diffusion model. According to [1], it learns complete SAFE sequences from the spectrum-free SAFE-GPT molecular-structure corpus, conditioned on formulas and fingerprints. The completion decoder is adapted separately from the same pretrained initialization; the two models have independent parameters, and the editor does not fix the AEMS prefix.
  4. Meaning of K/M: Each round selects up to K=16 unique molecules using formula/AEMS rules and joint fingerprint ranking. It independently samples M=4 random masked variants per molecule, yielding up to 64 editing inputs, rather than selecting the best four from a larger pool. Editing outputs are merged with fresh completion candidates for further filtering and ranking.

[1] Bohde, M., et al. FRIGID: Scaling Diffusion-Based Molecular Generation from Mass Spectra at Training and Inference Time. arXiv:2604.16648, 2026. https://arxiv.org/abs/2604.16648

W5——Additional budget and saturation in the third round

Round-wise improvements may simply come from adding candidates; CANOPUS Reference reaches 38.98% in the second round and does not improve further in the third.

Response. Thank you for considering the contribution of additional sampling. The relevant control is provided in 9cDw W2: with matched generation-request counts, full refinement has higher Acc@1/10 point estimates than Sampling-only on both datasets. Additional sampling therefore explains part of the improvement, while full refinement provides further numerical gains. The plateau in CANOPUS third-round Acc@1 indicates limited marginal gains from further refinement under this configuration. Repeated exploration of similar structures under a single prior is one possible explanation; see 9cDw Q3 for the diversity discussion and the fixed-total-budget multi-prior proposal.

Q1——Training contamination checks and MassSpecGym v1.5

Were training-side contamination checks performed? Alternatively, was the method evaluated under the MassSpecGym v1.5 protocol?

Response. Thank you for considering training-data contamination. When constructing our own AEMS completion pretraining set, we excluded complete molecules from CANOPUS, MSG, and the five external GNPS test sets using the connectivity block of non-stereochemical InChIKeys. Rechecking the actual training manifests confirmed zero structural overlap with all seven test sets. Further structural-similarity analysis is provided in FBze W3. The five GNPS evaluation sets also exclude structures overlapping with CANOPUS supervised training and validation, and use frozen CANOPUS models without external fine-tuning. We have not evaluated the method on MassSpecGym v1.5.

Dear Reviewer tUMN,

Thank you once again for your detailed and constructive review. Your questions about the respective roles of retrieval and generation, recovery of molecules outside the library, and refinement implementation have helped us make the evidence and technical details clearer.

We recognize the demands on your time. With the Paper Discussion deadline (November 18) approaching, we would be grateful for any further thoughts on our responses and revisions. If there are aspects that would benefit from further clarification, we would welcome the opportunity to discuss them before the discussion closes. We hope the added diagnostics and implementation details are helpful as you revisit your assessment.

Thank you again for your time and careful consideration.

Reviewer_9cDwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Anchora, a retrieval-guided framework for molecular structure elucidation from MS/MS spectra. It retrieves molecular references, extracts anchor-extended scaffolds, completes missing side chains using discrete diffusion, and applies full-molecule refinement to potentially correct incorrect structural priors.

Critical strengths
  • The overall motivation is compelling. Combining retrieval with constrained generation is a practical way to reduce the otherwise enormous molecular search space.

  • The canonical AEMS–SAFE representation is a technically thoughtful contribution, particularly in preserving attachment correspondence during molecular completion.

  • The empirical improvements over FRIGID are substantial, and the oracle-prior and zero-shot experiments provide useful insights into the benefits and limitations of structural priors.

Critical weaknesses
  • The framework uses a large external structure database, but the paper does not clearly establish whether target structures are excluded from retrieval. A retrieval-only baseline and target-excluded evaluation are important for understanding the actual generative contribution.

  • Refinement yields relatively modest improvements, and additional sampling may account for part of these gains. I would particularly like to see performance on examples where the retrieved AEMS is incorrect, along with a compute-matched ablation without ICEBERG guidance.

  • Much of the framework builds on existing AEMS conditioning and FRIGID-style refinement. The recently released GEMS work also demonstrates a retrieval-guided alternative, achieving stronger results on MassSpecGym. A discussion or controlled comparison would help clarify Anchora's distinctive advantages.

Questions for authors
  • Are ground-truth molecules excluded from retrieval? How does Anchora compare to a retrieval-only baseline?

  • How often does refinement successfully correct an incorrect AEMS prior? Are the gains beyond simply increasing sampling budget?

  • Why is the improvement from Acc@1 to Acc@10 so small? Is candidate diversity a bottleneck?

  • How does Anchora compare with GEMS under matched retrieval and inference budgets?

5 comments

Response to Reviewer 9cDw

We sincerely thank the reviewer for recognizing the motivation, the attachment correspondence in canonical AEMS–SAFE, and the value of the oracle and external-library experiments. Your questions help distinguish the information supplied by priors, how generation uses that information, and the incremental contribution of refinement. We provide the available evidence and direct explanations below.

W1——Controls for database coverage, retrieval-only prediction, and target exclusion

It is unclear whether the target is excluded from retrieval; retrieval-only and target-excluded evaluations are needed to understand the contribution of generation.

Response. Thank you for considering the role of structural resources. The external structure database supplies AEMS priors and is part of the method's design. The model must still use the spectrum to distinguish same-formula isomers and complete the unknown structure. Even when the target is present, database membership does not identify the correct entry. Therefore, database membership is not equivalent to successful retrieval, and coverage is not equivalent to structure elucidation accuracy.

For the shared-database baseline comparison, please see our response to Reviewer FBze W2. For the role of the structure database, training leakage, and pretraining similarity, please see our response to Reviewer FBze W3.

W2——Does refinement merely increase the number of samples?

Performance with incorrect AEMS priors and a compute-matched ablation without ICEBERG guidance are needed.

Response. Thank you for proposing this control. Recovery with incorrect priors is reported in FBze W4; molecular recovery is not equivalent to strict AEMS correction. To distinguish refinement from additional sampling, we fix the model and reference and compare against Sampling-only, without ICEBERG guidance or editing:

DatasetMethodAcc@1 (%)Acc@10 (%)Strict AEMS@1 (%)
CANOPUSAnchora-Base36.7439.8551.68
CANOPUSSampling-only38.3641.5953.42
CANOPUSAnchora refinement39.1042.5954.05
MSGAnchora-Base25.0626.5942.66
MSGSampling-only26.6227.0944.53
MSGAnchora refinement27.3828.8945.87

Relative to Sampling-only, refinement improves Acc@1/10 by 0.75/1.00 percentage points on CANOPUS and 0.76/1.80 percentage points on MSG. Additional sampling explains part of the improvement, while full refinement provides further numerical gains on both datasets.

W3——Differences from existing conditional generation methods and GEMS

The method relies on existing AEMS-conditioned generation and FRIGID-style refinement; GEMS also uses retrieval, so the differences or a controlled comparison should be provided.

Response. Thank you for recommending GEMS. Anchora's advantage is to directly reuse the scaffold and attachment sites from a retrieved structure, allowing the model to focus on generating the unknown portion. This narrows the structural search space relative to generating a complete molecule from scratch. Whereas GEMS performs graph editing from a complete candidate, Anchora retains only AEMS and can regenerate sidechains from spectral evidence without incrementally repairing the reference molecule's sidechains.

Canonical AEMS–SAFE further ensures correct correspondence between new sidechains and anchors, making structural priors usable as generative conditions. We therefore exploit existing structural knowledge while retaining flexibility to reconstruct the unknown portion. The numerical comparison is provided in Q4.

Q1——Is the target excluded, and does Anchora outperform retrieval-only methods?

Is the ground truth excluded from retrieval? How does Anchora compare with retrieval-only methods?

Response. Thank you for requesting clarification. We retain targets already present in the database. Frozen MIST predicts fingerprints from query spectra, but under the training/test splits used in this experiment, its supervised training did not use test spectra. The retrieval database contains molecular structures only, without corresponding experimental spectra. Thus, the presence of a target in the database does not imply that its test spectrum participated in supervised training.

The external structure database introduces AEMS priors by design. Unlike generating complete molecules from scratch, Anchora uses retrieved local structures to complete unknown portions. Even when the target is present, the model must still distinguish same-formula candidates using the spectrum and recover the structure: database membership is not equivalent to correct retrieval or recovery. Further explanations and the shared-database comparison are provided in FBze W2/W3.

Published retrieval-only results are compared below. ESP, MIST, and JESTR directly rank structural candidates without molecular generation and do not require candidates to have experimental spectra. Retrieval results on both datasets use known-formula conditions: CANOPUS/MSG values are from Table 2 of the JESTR paper; Anchora uses the main Reference results in our paper.

DatasetMethodTop-1 (%)
CANOPUSESP (retrieval-only)23.69
CANOPUSMIST (retrieval-only)27.96
CANOPUSJESTR (retrieval-only)45.76
CANOPUSAnchora (Reference)38.98
MSGESP (retrieval-only)11.05
MSGMIST (retrieval-only)9.57
MSGJESTR (retrieval-only)11.85
MSGAnchora (Reference)23.50

Under their respective evaluation settings, Anchora's MSG Top-1 exceeds all three retrieval-only methods; on CANOPUS, it exceeds ESP/MIST but is below JESTR. JESTR uses the larger PubChem source for CANOPUS candidates, whereas Anchora uses CSU-MS2.

More importantly, retrieval-only outputs are restricted to the database: an absent target cannot be returned exactly. Anchora instead uses retrieved AEMS as a local prior and regenerates the unknown portion from the spectrum, so the final structure need not exist in the retrieval database. Anchora therefore exploits existing structural knowledge while providing a route to discovering out-of-library molecules.

Q2——How often are incorrect priors corrected, and are the gains beyond additional sampling?

How often does refinement correct an incorrect AEMS? Do the gains exceed those from simply increasing the sampling budget?

Response. Thank you for distinguishing these questions. Recovery results and identity criteria are provided in FBze W4, and the sampling control is provided in W2 of this response. In the W2 comparison, refinement has higher point estimates than Sampling-only on both datasets, indicating that additional sampling is not the sole source of improvement.

Q3——Does the small Top-1/Top-10 gap indicate a diversity bottleneck?

Why is the improvement from Acc@1 to Acc@10 small? Is candidate diversity a bottleneck?

Response. Thank you for highlighting this diagnostic.

A single AEMS imposes a specific mechanistic limitation: additional sampling under the same prior mainly explores sidechain combinations and may not cover alternative scaffold and anchor hypotheses. Thus, "generating different molecules" does not imply "covering the correct structural hypothesis." A targeted future direction is to use three distinct AEMS priors, allocate sampling within a fixed total generation budget, and then merge, deduplicate, and rank the candidates together.

Q4——How does Anchora compare with GEMS under matched retrieval and inference budgets?

How does Anchora compare with GEMS under the same retrieval and inference budgets?

Response. Thank you for recommending GEMS. The table places Table 1 of the GEMS paper alongside Anchora's main Reference results to show their respective performance.

DatasetMethodTop-1 (%)MCES@1Sim@1Top-10 (%)MCES@10Sim@10
CANOPUS / NPLIB1GEMS29.647.730.5950.816.080.71
CANOPUSAnchora (Reference)38.985.420.6541.724.630.68
MSGGEMS30.229.680.5541.338.260.61
MSGAnchora (Reference)23.5010.570.5025.049.420.54

Dear Reviewer 9cDw,

Thank you again for your careful review and constructive questions about retrieval, incorrect AEMS priors, and the contribution of refinement beyond additional sampling. We particularly appreciate your suggestions, which motivated the additional analyses and Sampling-only comparison presented in our responses.

We understand that this is a busy period. As the Paper Discussion deadline (November 18) approaches, we would greatly appreciate any further feedback on whether these results and clarifications address your concerns. We would be happy to discuss any remaining questions before the discussion closes, and hope the additional evidence is useful as you revisit your assessment.

Thank you again for your time and thoughtful consideration.

Reviewer_FBzeOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces ANCHORA, a retrieval-guided framework for molecular structure elucidation from tandem mass spectra. The method first retrieves formula-compatible molecular structures using MIST-predicted fingerprints and extracts an anchor-extended molecular scaffold (AEMS) from the top-ranked reference. It then uses an AEMS–SAFE representation and a discrete diffusion model to complete the missing side chains, followed by three rounds of ICEBERG-guided full-molecule refinement. The proposed method achieves 38.98% and 23.50% top-1 accuracy on CANOPUS and MassSpecGym, respectively, outperforming the reported generative baselines.

Critical strengths
  • The paper is generally well organized, and the overall framework is easy to follow. Combining structure retrieval with molecular generation is a reasonable approach to reducing the search space for MS/MS structure elucidation.
  • The experimental results on CANOPUS and MassSpecGym are promising. The authors also evaluate the method on five external spectral libraries to examine its generalization ability.
Critical weaknesses
  • The methodological novelty is somewhat limited. The framework mainly combines existing components, including MIST for fingerprint prediction and retrieval, AEMS from MSAnchor, discrete diffusion from FRIGID, and ICEBERG for refinement. I have a negative assessment of this paper, but if the AC and other reviewers think it can be recognized as a solid contribution, I would raise my score if the authors can address the following issues:

  • The main concern is the fairness of the baseline comparisons in Table 1. ANCHORA uses an external molecular structure database containing 672,681 molecules (Lines 160–186), while the compared generative baselines do not have access to the same retrieval information. Therefore, the improvements cannot be directly attributed to the proposed generation framework. The authors should reproduce MADGEN and MSAnchor using the same database, as the main components are borrowed from these two methods.

  • The potential data leakage and retrieval database coverage need to be clarified. In Lines 351–355, the authors state that overlapping validation and test molecules are excluded from molecular completion pretraining. However, it is unclear whether test molecules are present in the retrieval database. Moreover, the authors should report the similarity distribution between the pretraining dataset and the test set.

  • The claim that refinement can correct an incorrect retrieved scaffold requires stronger experimental evidence. In Lines 282–290, the authors emphasize that full-molecule refinement can modify the initial AEMS. However, Appendix B states that candidates satisfying the retrieved AEMS constraints are prioritized during ranking. I am curious how often the refinement actually corrects an incorrect AEMS prior. The authors should report the exact recovery accuracy specifically on samples with incorrect retrieved AEMS.

  • The zero-shot comparisons in Table 2 are not entirely convincing. ANCHORA-Base benefits from an external retrieval database, while FRIGID-base relies on molecular generation. Thus, the reported improvements do not necessarily demonstrate better generalization of the generative model itself.

  • The computational cost of the framework is not sufficiently discussed. The method involves molecular retrieval, diffusion-based completion, ICEBERG spectrum simulation, and multiple refinement rounds. Given the relatively small improvement from refinement, reporting inference time and computational cost compared with FRIGID would help assess the practical value of the proposed framework.

  • The overall training pipeline is relatively complex, involving multiple pretrained models and training stages. I suggest adding a table summarizing the dataset sources, number of samples, and model initialization or checkpoint sources for each stage, including MIST and ICEBERG.

Questions for authors

The main concern is the fairness of the baseline comparisons in Table 1. The authors should reproduce these baselines to ensure a fair comparison.

8 comments

Response to Reviewer FBze

We sincerely thank the reviewer for carefully reading our paper and recognizing the motivation for combining retrieval and generation, as well as the experiments on two benchmarks and external spectral libraries. We respect your concerns but believe it is important to distinguish "retrieval is part of the system" from "the method merely returns retrieved results." We address each point below using the mechanisms and available controls.

W1——Is the methodological novelty merely a combination of existing components?

The framework mainly combines MIST, AEMS from MSAnchor, FRIGID discrete diffusion, and ICEBERG, with limited methodological novelty.

Response. Thank you for considering the methodological novelty. Our core innovation is to retrieve AEMS from an external structure database, explicitly introducing local structural priors into molecular generation. Unlike de novo baselines that generate complete molecules from scratch, Anchora uses retrieved priors to specify part of the structure and completes the unknown portion based on the spectrum. The main experiments, zero-shot evaluation, and ablations all center on this design, testing whether introducing AEMS priors improves structure elucidation. The primary contribution is therefore this prior-guided generation design, rather than replacing the backbone or simply assembling existing components.

A further technical challenge is how to turn retrieved attachment-aware structural evidence into a generative condition that supports consistent decoding and correct completion. Simply using a scaffold string as a prefix does not guarantee correct correspondence between symmetry-equivalent anchors and sidechains after canonicalization. Canonical AEMS–SAFE establishes explicit anchor-to-sidechain correspondence, making the retrieved prior usable for structural completion.

W2——Does the external structure database make baseline comparisons unfair?

Anchora uses an external structure database, whereas generative baselines cannot access the same information. The authors should reproduce MADGEN and MSAnchor using the same database.

Response. Thank you for raising the resource-matching issue. The external structure database is an explicit input through which Anchora obtains AEMS priors and is part of the method's design; Table 1 evaluates this complete system. The gains cannot all be attributed to the decoder, but retrieval-guided generation should not be discounted as a contribution simply because retrieval contributes to those gains.

MSAnchor predicts AEMS through a de novo pathway and has no retrieval database that can be directly replaced. Adding a retrieval module would extend the method rather than reproduce it by replacing its database. For MADGEN, we have completed experiments using the same CSU-MS2 source database and compare it with Anchora-Base (Reference) without refinement (see Section 4.6 and Table 3 of the revised manuscript, lines 479–499):

DatasetMethodPrior Scaf. (%)Prior AEMS (%)Acc@1 (%)MCES@1Sim@1Acc@10 (%)MCES@10Sim@10
CANOPUSMADGEN (CSU-MS2)37.46--5.5915.940.299.9714.620.37
CANOPUSAnchora-Base (Reference)66.5359.7836.865.540.6439.854.690.68
MSGMADGEN (CSU-MS2)36.67--2.3634.680.224.8933.570.28
MSGAnchora-Base (Reference)45.1443.2721.6310.700.4922.959.450.53

In the tabulated evaluations, Anchora-Base achieves higher scaffold accuracy and better performance on all six generation metrics on both datasets: higher exact recovery, lower MCES, and higher fingerprint similarity. This demonstrates the approach's competitiveness in both obtaining informative priors and reconstructing complete molecules.

W3——Training leakage, retrieval database coverage, and pretraining similarity

Are test molecules present in the retrieval database? Potential data leakage and database coverage require clarification, and the similarity distribution between pretraining and test sets should be reported.

Response. Thank you for considering data independence. We address these three issues separately.

1. Data leakage. Frozen MIST predicts fingerprints from query spectra; under the training/test splits used in this experiment, MIST's supervised training did not use test spectra, which are supplied only at inference. The retrieval database contains molecular structures only, without their corresponding experimental spectra. Thus, this procedure does not leak test information through supervised training on test spectra or through access to their experimental spectra in the retrieval database.

2. Retrieval database coverage. Unlike baselines that generate complete molecules from scratch, we deliberately retrieve AEMS from an external database to guide generation with local structural priors; the experiments test whether this design improves structure elucidation. Even when the target molecule is present in the database, the downstream generator must still use the spectrum to distinguish many same-formula isomers sharing the same AEMS; it does not directly know which entry is correct. For example, using a larger structure database such as PubChem could expand candidate coverage but would not automatically solve correct ranking. Database membership only supplies a candidate; successful retrieval and effective completion turn that access into predictive gains.

3. Pretraining similarity. We audited structural similarity between the pretraining set and seven test sets:

Test setNearest-neighbor median [Q1, Q3]All-pairs mean Tanimoto
CANOPUS0.789 [0.707, 0.857]0.1036
MSG0.657 [0.537, 0.788]0.1102
ECRFS0.750 [0.637, 0.825]0.0945
Pesticides0.819 [0.740, 0.833]0.1061
MIADB0.743 [0.626, 0.811]0.1081
IQAMDB0.804 [0.741, 0.879]0.0993
WineDB0.795 [0.699, 0.853]0.0981

The all-pairs mean similarity is approximately 0.10 for all seven test sets, while nearest-neighbor medians range from 0.657–0.819. This indicates that the pretraining corpus contains chemical patterns related to the test structures alongside substantial structural diversity, providing a rich reference for learning across structures.

W4——Exact target-molecule recovery accuracy when the retrieved AEMS is incorrect

Full-molecule refinement is claimed to modify the initial AEMS, but final ranking prioritizes candidates satisfying that AEMS. Exact recovery accuracy should be reported when the retrieved AEMS is incorrect.

Response. Thank you for proposing this mechanistic test. Full-molecule refinement is intended to prevent the initially retrieved AEMS from becoming an immutable generative boundary, rather than requiring recovery with an incorrect prior to approach recovery with a correct prior. Retrieval narrows the main search space, while refinement preserves a correction route for imperfect priors. Their roles are complementary; refinement is not expected to repeat the entire structural search.

Exact target recovery on samples with incorrect retrieved AEMS:

DatasetSamplesAnchora-Base Acc@1Anchora Acc@1Anchora-Base Acc@10Anchora Acc@10
CANOPUS2983.02% (9)3.36% (10)3.36% (10)4.36% (13)
MSG9,5442.94% (281)3.03% (289)3.46% (330)3.66% (349)

Even with an incorrect initial AEMS, target-molecule recovery remains possible. After refinement, Top-10 hits increase from 10 to 13 and 330 to 349, respectively, bringing more targets into the final candidate lists.

An incorrect retrieved AEMS is compatible with nonzero molecular recovery Acc: molecular matching in the paper removes stereochemistry and applies tautomer normalization, whereas strict AEMS Acc requires matching scaffold and anchor structures. The two metrics use different identity criteria.

W5——Does the zero-shot comparison establish generalization?

Anchora-Base benefits from an external retrieval database, so the improvements in Table 2 do not necessarily indicate better generalization of the generative model itself.

Response. Thank you for considering this setting. The external structure database supplies AEMS priors by design: retrieval introduces local structural evidence, and the model completes the unknown portion. The zero-shot experiment tests whether introducing AEMS priors continues to offer advantages over de novo generation without these priors on external libraries, with frozen models and no external fine-tuning. The structure database is thus part of the evaluated method, not an additional condition outside the experiment's purpose. This comparison evaluates the complete system's advantage without attributing all gains solely to the decoder.

Table 2 compares both methods on the same 1,066 spectra with known molecular formulas, without external fine-tuning or refinement, and excludes structural overlap with CANOPUS supervised training. Acc@1 improves by 5.10–27.14 percentage points across the five libraries, with improvements in Acc@10, MCES, and Sim, supporting cross-library effectiveness.

W6——Is the computational cost of refinement justified?

Refinement yields modest gains but involves retrieval, diffusion generation, ICEBERG, and multiple editing rounds. Inference time and cost relative to FRIGID should be reported.

Response. Thank you for considering computational cost. Stage timings are shown below; the two full-FRIGID timings are estimates extrapolated from previous logs:

DatasetModel and timed stageRuntime (h)Mean (s/spectrum)
CANOPUSFRIGID-base1.717.68
CANOPUSFRIGID15.971.4
CANOPUSAnchora: base AEMS generation1.245.57
CANOPUSAnchora: three additional refinement rounds4.2819.20
MSGFRIGID-base61.4412.95
MSGFRIGID571.2120.4
MSGAnchora: base AEMS generation44.789.44
MSGAnchora: three additional refinement rounds119.2525.13

The tabulated per-spectrum base-generation time falls from 7.68/12.95 s for FRIGID-base to 5.57/9.44 s, a reduction of about 27% on both datasets. Fixing the AEMS leaves only the unknown completion segment to generate, and shorter sequences can reduce decoding steps. Retrieval runs only once, taking 59.094/872.313 s in total on CANOPUS/MSG, or 0.0736/0.0511 s/spectrum, introducing structural priors at very low retrieval overhead.

W7——Training stages, data, and initialization sources

The training procedure is complex. A table should summarize the data sources, sizes, and checkpoints for each stage, including MIST and ICEBERG.

Response. Thank you for suggesting this format for verification. Counts below are train/validation records from the actual training manifests; the original CANOPUS/MSG splits are unchanged.

Stage or componentData and sizeInitialization and updates
AEMS pretrainingHMDB, DSSTox, COCONUT, MOSES; 2,659,168 / 54,271frigid_base.ckpt; all DLM parameters, 200,000 steps;
Dataset adaptationCANOPUS 6,119 / 354; MSG 161,939 / 18,476Pretrained DLM; frozen spectrum encoder, 10,000 steps each
Spectrum–decoder joint adaptationSame as above; dataset-specific, without cross-dataset mixingThis run's domain-adapted DLM + mist_canopus.pt / mist_msg.pt; joint updates, 5,000 / 10,000 steps
Completion-lengthgeneral 2,278,494 / 46,501; CANOPUS 2,284,613 / 354; MSG 2,440,433 / 18,476Separate NGBoost models; general/domain sample weights 1:4 for domain models, fixed hyperparameters
Retrieval MISTInherited full-molecule fingerprint predictorFrozen for retrieval; separate from the jointly adapted generation-conditioning encoder
Full-molecule editing / ICEBERGOriginal FRIGID / dataset-specific ICEBERGFrozen during inference; distinct from the completion checkpoint

This training curriculum uses shared pretraining for 200,000 steps of general completion learning, followed by 10,000 steps of domain adaptation per dataset and 5,000/10,000 steps of joint adaptation for CANOPUS/MSG. The stages first learn structural completion, then adapt to the domain, and finally jointly use spectral information; the length model specifically predicts the unknown completion length.

Q1——Have the reproductions needed for a fair comparison been completed?

Relevant baselines should be reproduced to ensure a fair comparison.

Response. Thank you for reiterating this priority. MADGEN has been reproduced using the same CSU-MS2 source database; see W2 for results. MSAnchor has no replaceable retrieval database, and adding a retrieval module would change the original method.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We hope this message finds you well. We sincerely thank you for the time and effort devoted to evaluating our submission. Our paper uses the AEMS extracted from retrieved molecules as prior knowledge to guide molecular generation. During rebuttal, we provided point-by-point responses to all three reviews, added experiments and analyses, and revised the manuscript.

Reviewer FBze returned with concerns about methodological novelty and pretraining similarity. We clarified how canonical AEMS–SAFE resolves anchor correspondence to support consistent molecular completion, beyond changing the source of the prior. We also distinguished complete-structure overlap from nearest-neighbor similarity and explained what the pretraining audit establishes.

Reviewer 9cDw carefully examined the additional results and raised further questions about generation’s incremental value beyond retrieval and controlled comparisons. We clarified the evidence from retrieval diagnostics, the Sampling-only comparison, and the MADGEN experiment using the same CSU-MS2 structure source. We sincerely appreciate this engagement, which helped us distinguish the demonstrated benefits of prior-guided generation from stronger claims that the current experiments do not establish.

Reviewer tUMN has not provided follow-up feedback after our detailed responses and reminders. We supplied the requested retrieval Top-1/Top-K results, database coverage, prior-conditional accuracy, and library-membership diagnostics, together with refinement implementation details, a Sampling-only comparison, and training-data overlap checks. Without further feedback, we remain uncertain whether these additions resolved the initial concerns or whether further clarification would have been helpful.

We respect the assessments of the two reviewers who continued the discussion and are grateful for their consideration. We respectfully ask the Area Chair to consider the complete responses and additional evidence alongside the original reviews, including the questions for which no follow-up feedback was available.

Thank you again for your thoughtful feedback, valuable time, and careful consideration. We remain available should any further clarification be helpful.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.