acceptodds
Under review as a conference paper at ICLR 2027

SOUR: Scaling Synthetic Data for Small Molecule Structure Elucidation via Train- and Test-Time Guidance

14 views

Abstract

Synthetic data can reduce reliance on scarce experimental annotations, but naively increasing its volume does not guarantee better inverse prediction. This is pronounced for structural elucidation from tandem mass spectra, where high-quality experimental annotation is limited while the forward simulators remain imperfect. We introduce SOUR, a framework for scaling synthetic supervision at training and/or test time by tailoring it to experimental evidence and accommodating imperfect simulations. We use forward simulation to select synthetic examples relevant to experimental spectra of interest and construct virtual refinement trajectories. In training, this supervision supports spectral encoder adaptation and teaches an autoregressive generator to interpret imperfect fingerprints and revise candidate structures. At test time, observed spectra guide query-specific adaptation, while comparisons with newly simulated spectra provide feedback for iterative correction. Our empirical study supports a general strategy for synthetic-data scaling whose benefits transfer across model architectures. On public benchmarks for mass spectrum-based structure elucidation, SOUR-full achieves state-of-the-art identification accuracy, substantially outperforming prior models. These results highlight the potential of imperfect forward simulation to turn an unlabeled observation of unknowns into synthetic supervision at training time and to guide adaptation and correction at test time.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Tandem Mass Spectrometry, Synthetic Data, Test-time Adaptation, Generative Modeling, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_bA57OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduce SOUR predicts molecular structures from mass spectra and known molecular formulae. It selects query-relevant synthetic spectra to train a spectrum-to-fingerprint encoder, then further adapts the encoder using query pseudo-labels. A molecular decoder is trained with perturbed fingerprints and learns to revise candidate structures using feedback from simulated spectra. On MassSpecGym and NPLIB1, SOUR achieves strong improvements in fingerprint prediction and retrieval.

Critical strengths
  • Comprehensive evaluations on the MassSpecGym and NPLIB1 benchmarks demonstrate the effectiveness of SOUR.
  • The approach is general and model-agnostic, with encoder adaptation improving performance across multiple backbones.
Critical weaknesses
  • The broad state-of-the-art claim is not fully supported by the reported comparisons. Published FRIGID results on MassSpecGym report 18.29% Top-1 accuracy, compared with 17.98% for SOUR. This raises concerns about the accuracy and fairness of the paper’s claims.
  • The models are trained and adapted separately on MassSpecGym and NPLIB1, with no cross-dataset evaluation. In real-world applications, a suitable target dataset may not be available for training or validation. It is therefore unclear whether the approach can generalize to new data sources in such settings, which limits its practical applicability.
Questions for authors
  1. Why does NPLIB1 use up to 2,048 candidate molecules per query, while MassSpecGym uses only 64? Could you evaluate both datasets with (K = 64, 256, 2048), using the same collision energies and keeping the encoder update budget and real-data mixing fixed within each dataset?
  2. Could you compare random and selected synthetic data, both with and without Phase 2, using different data sizes, such as 10K, 100K, and 1M synthetic pairs?
  3. MassSpecGym has more than 10,000 test spectra, yet the adaptation process uses 20 million synthetic pairs. This raises concerns about the computational and storage costs of generating and maintaining such a large synthetic dataset, which may limit practical use.
  4. Could you report accuracy by the maximum similarity to the decoder pretraining set and the synthetic-source molecules, using bins below 0.5, 0.5-0.7, 0.7-0.9, and at least 0.9? This would help assess how well the method generalizes to molecules that are less similar to the training data.
4 comments

We thank the reviewer for their thorough and thoughtful assessment of our work. Firstly, regarding the presentation score of the paper, we have carefully revised and elaborated upon the motivation, methodology and results of the paper in the updated version of the submission; we hope these revisions make it easier to understand the workflow and assess its merits. Below, we address your insightful observations and suggestions for revisions:

The broad state-of-the-art claim is not fully supported by the reported comparisons. Published FRIGID results on MassSpecGym report 18.29% Top-1 accuracy, compared with 17.98% for SOUR. This raises concerns about the accuracy and fairness of the paper’s claims.

We appreciate the reviewer’s attention to this discrepancy. We coordinated closely with FRIGID and MS-GPT authors on the benchmarking results, and based on the issues and suggestions we all agreed on, both papers have recently updated their preprints, and the updated results are consistent with the reported version in this paper. [1,2] The issue was that the shared MIST checkpoint both FRIGID and MS-GPT used was trained on structures that closely mimic test structures, and the only filtering rule was to exclude exact matches. They decided this is not a fair setting because testing information was used (though indirectly) during training, and the FRIGID authors rebuilt those baselines with a fairer version of MIST, the MIST 4M aug. version, which is reproducible and does not select synthetic data unfairly. We note that this differs fundamentally from SOUR's query-conditioned selection: SOUR only uses the test spectra, which are available at inference time, and never uses test structures, whereas the data selection behind the previous MIST checkpoint used information from test structures. On this fixed fair benchmark that all authors agreed on, SOUR demonstrates a step forward in performance on structural elucidation. We hope this detailed history of auditing and revision of results by community members reassures the reviewer about the legitimacy of the claims we present in the paper.

The models are trained and adapted separately on MassSpecGym and NPLIB1, with no cross-dataset evaluation. In real-world applications, a suitable target dataset may not be available for training or validation. It is therefore unclear whether the approach can generalize to new data sources in such settings, which limits its practical applicability.

We appreciate the reviewer’s consideration for the real-world utility of our approach. We did not perform naive cross-dataset evaluation in the main text because 645 of the 714 unique test molecules in NPLIB1 (90.3%, including 570 in MassSpecGym's training split) appear directly in MassSpecGym, which would introduce severe train-to-test leakages of data. We agree with the reviewer’s assessment that a suitable labeled training or validation dataset may not exist in some real-world settings.

To evaluate zero-target-label generalization from a general reference corpus (MassSpecGym) to a new specialized source (NPLIB1) without leakage, we isolated the 32 NPLIB1 test spectra whose structures are completely absent from all MassSpecGym splits and synthetic pools, evaluating fixed MassSpecGym checkpoints directly with their MassSpecGym validation thresholds (zero NPLIB1 training or validation data):

Encoder settingMIST No-aug.MIST 4M/2M-aug.Frozen MISTSOUR\text{MIST}^{\text{SOUR}}
MassSpecGym → NPLIB1 (cross-source, zero NPLIB1 labels)0.2600.3870.354
NPLIB1 → NPLIB1 (in-domain reference, 9,181 labeled train spectra)0.3350.3600.401

Real-only MIST No-aug. transfers poorly across sources (0.260 vs. 0.335 in-domain). In contrast, synthetic-augmented MassSpecGym encoders generalize well with zero target labels: both MIST 4M-aug. (0.387) and even the unadapted frozen MSG MISTSOUR\text{MIST}^{\text{SOUR}} checkpoint (0.354) outperform the in-domain NPLIB1 No-aug. encoder (0.335) trained directly on 9,181 labeled NPLIB1 spectra (and on the broader 63-spectrum subset disjoint from MassSpecGym and 4M-aug., cross-source MIST 4M-aug. reaches 0.421, matching in-domain NPLIB1 No-aug. at 0.421). Notably, MSG MISTSOUR\text{MIST}^{\text{SOUR}} performs slightly worse than MSG 4M-aug. here because its weights were adapted specifically to MassSpecGym's test queries QMSG\mathcal{Q}_{\text{MSG}}. In real deployment on an unlabeled new data source Qnew\mathcal{Q}_{\text{new}}, SOUR's test-time adaptation requires no target training or validation labels; one only needs the GLACIER simulator and source training set Dreal\mathcal{D}_{\text{real}} to adapt directly to the unlabeled target spectra Qnew\mathcal{Q}_{\text{new}}.

(to be continued)

Why does NPLIB1 use up to 2,048 candidate molecules per query, while MassSpecGym uses only 64? Could you evaluate both datasets with (K = 64, 256, 2048), using the same collision energies and keeping the encoder update budget and real-data mixing fixed within each dataset?

This difference in the molecule allowances for Phase 1 can be attributed to the differing query spectrum counts of both benchmarks: NPLIB1 consists of 819 test spectra, while MassSpecGym contains 17,556 test spectra (about 21x more). To achieve a reasonably similar order of total synthetic data points to utilize similar compute for adaptation, we allot a higher number of candidates to be accrued per query spectrum for NPLIB1 versus MassSpecGym. We do not have sufficient compute to consult and adapt on significantly more additional queries for MassSpecGym, since e.g. 2048 candidates per query x 17,556 query spectra x 5 collision energies would amount to 179 million synthetic pairs. Our updated Figure 2, which now breaks down the SOUR-selected augmentation into Phase 1 alone and Phase 1+2, highlights the log-linear trend in Phase 1 scaling that would be commensurate with the chosen value of KqryK_{\text{qry}}. The reported SOUR utilized for the MassSpecGym benchmark uses Kqry=64K_{\text{qry}} = 64 and utilizes 5 million spectra; the lower points leftward are run with smaller values of KqryK_{\text{qry}}, which cement the log-linear relationship we observe. We do observe from the rightmost point that utilizes 10 million spectra (which could roughly extrapolate as Kqry=128K_{\text{qry}} = 128) that performance saturates (with encoder accuracy going from 0.507 to .505, and de novo Acc@1 going from 11.711 to 11.626). We have similar trends observed for NPLIB1, where scaling the number of synthetic pairs also shows log-linear improvements on encoder reconstruction accuracy:

Synthetic pairsMean Tanimoto
1,0000.448
5,0000.456
10,0000.454
62,5000.485
250,0000.520
1,000,0000.558
5,000,0000.599
13,000,0000.606

Here, since the reported NPLIB1 encoder utilizes 13 million synthetic pairs corresponding to a Kqry=2048K_{\text{qry}} = 2048, we can estimate Kqry=64K_{\text{qry}} = 64 as roughly matching 500K pairs and Kqry=256K_{\text{qry}} = 256 matching 1.6 million; the values of mean Tanimoto at the closest runs, 0.520 for 250K and 0.558 at 1 million (versus 0.606 for the selected NPLIB1 encoder) support the conclusion that increasing KqryK_{\text{qry}} can indeed log-linearly improve the encoder.

Could you compare random and selected synthetic data, both with and without Phase 2, using different data sizes, such as 10K, 100K, and 1M synthetic pairs?

Thank you for the suggestion. We present a comparison of these two phases of encoder-based augmentation for the MassSpecGym benchmark in the updated Figure 2 and Figure 5 (in PDF), which now demonstrate two separate trends for the improvement in SOUR-based augmentation (where Phase 1 alone is denoted in light pink and Phase 1 and 2 together are denoted in dark pink). This plot has also been updated to include evaluations of the same models for de novo structural elucidation accuracy, in addition to fingerprint reconstruction accuracy. We provide a tabular documentation of these results below (also Table 8 in the updated PDF):

Table: Numerical values for the MassSpecGym synthetic-data scaling study. All metrics are reported to three decimal places. Random augmentation and SOUR Phase 1 use the Phase 1 synthetic budget N. Series incorporating Phase 2 additionally include 17,556 query–pseudo-label pairs and use N + 17,556.

Phase 1 pairs NWith Phase 2 N+17,556MIST Tanimoto: Random aug.MIST Tanimoto: Random + SOUR Phase 2MIST Tanimoto: SOUR Phase 1MIST Tanimoto: SOUR Phase 1+2SOUR-base Top-1 accuracy (%): Random aug.SOUR-base Top-1 accuracy (%): SOUR Phase 1SOUR-base Top-1 accuracy (%): SOUR Phase 1+2
1,00018,5560.2430.2390.2540.3090.0000.5183.389
5,00022,5560.2410.2820.2660.3380.0230.4734.004
10,00027,5560.2460.2920.2820.3600.0231.4355.263
62,50080,0560.2420.2890.3450.4170.0003.3216.943
250,000267,5560.2620.3250.3770.4320.0915.5888.863
1,000,0001,017,5560.2970.3380.4150.4780.4619.55212.298
5,000,0005,017,5560.3280.3740.4410.5071.07711.71115.448
10,000,00010,017,5560.2910.3100.4420.5050.32511.62615.379

(to be continued)

Both plots indicate that Phase 1 augmentation by itself improves over random augmentation at all steps, with the most significant improvements being accrued with larger amounts of synthetic datapoints (e.g. at 100K synthetic pairs and beyond), whereas random augmentation does not change performance drastically. In fact, for de novo elucidation accuracy, it stays largely unchanged at close to 0%. Applying Phase 2 to every SOUR Phase 1 model improves its encoder accuracy and de novo Acc@1. We further highlight the smaller, but still observable benefit of the Phase 2 augmentation to randomly augmented encoders in the MIST Tanimoto: Random + SOUR Phase 2 column, where the increases from MIST Tanimoto: Random are marginal in most cases, except for at 1,000 pairs where encoder accuracy is instead worsened (from 0.243 to 0.239). We therefore credit the strength of the SOUR-augmented encoder to the complementary nature of both augmentation phases. The value of Phase 2 does improve, but to a lesser extent, suggesting that Phase 1, spectral-side alignment, is a critical first step to getting Phase 2 to further improve the encoder significantly.

MassSpecGym has more than 10,000 test spectra, yet the adaptation process uses 20 million synthetic pairs. This raises concerns about the computational and storage costs of generating and maintaining such a large synthetic dataset, which may limit practical use.

Thank you for raising this concern. First, the 20 million figure for MassSpecGym was incorrect and has been updated to the correct 5 million figure. Generating and storing these 5 million spectra takes only ~1 NVIDIA RTX PRO 6000 GPU hour and 2 GB, which we consider not burdensome. We provide total bulk and amortized per-query compute costs in Table 12 and Appendix D.5, which in part account for the regeneration of selected synthetic neighbors (Synthetic generation) and Phase 1 and 2 training to update the encoder. We consider the biggest compute and storage bottleneck to be building the ~100M PubChem synthetic library, which takes 19.5 GPU hours and 40 GB when compressed per acquisition condition. We believe these are not extremely demanding storage or compute costs, and we plan to make the precomputed synthetic libraries available upon deanonymization to remove the recomputation burden for users adapting to new query spaces.

Could you report accuracy by the maximum similarity to the decoder pretraining set and the synthetic-source molecules, using bins below 0.5, 0.5-0.7, 0.7-0.9, and at least 0.9? This would help assess how well the method generalizes to molecules that are less similar to the training data.

Thank you for the suggestion. We now include a table that stratifies de novo accuracy (Acc@1 and Acc@10) based on their maximum similarity to the decoder pretraining set and the Phase 1 (P1) synthetic-source molecules (Table 9, Appendix D.1.3).

For MassSpecGym:

Max similarity bin (num. of molecules: decoder / P1)Decoder pretrainingP1 synthetic sources
Acc@1Acc@10Acc@1Acc@10
<0.5 (57 / 300)1.01.00.20.8
0.5–<0.7 (636 / 968)5.79.23.95.8
0.7–<0.9 (1,587 / 1,420)7.013.814.420.7
≥0.9 (717 / 309)31.640.036.550.1

For NPLIB1:

Max similarity bin (molecules: decoder / P1)Decoder pretrainingP1 synthetic sources
Acc@1Acc@10Acc@1Acc@10
<0.5 (8 / 6)0.00.00.00.0
0.5–<0.7 (99 / 82)14.328.63.48.0
0.7–<0.9 (352 / 332)21.133.219.229.7
≥0.9 (255 / 294)29.146.132.052.2

The decoder trends we quantify here, which are near-0 for <0.5 Tanimoto similarity, track with the generalization relationships observed by other decoders such as FRIGID ([1], Table 8); this is reasonable given that the fixes we made to the decoder are less about improving its generalization across chemical space and more to adapt to the errors of the particular encoder we train. On MassSpecGym, test structures with close synthetic neighbors included that end up having high structural similarity see a greater improvement in accuracy relative to the decoder pretraining similarity (e.g. at 0.7-0.9, for MassSpecGym, Acc@1 for decoder neighbors is 7.0% while for synthetic neighbors is 14.4%), which would suggest the SOUR-based exposure of the encoder to similar structures improves structural generation metrics versus exposure of the structure during decoding pretraining.

(to be continued)

Thank you once again for your constructive and detailed suggestions, which have strengthened both our empirical analysis and the clarity of the manuscript. We hope that clarifying the updated FRIGID benchmark results, adding the zero-label cross-dataset and similarity-stratified evaluations, and detailing the practical compute and storage costs fully resolve your concerns, and we would greatly appreciate it if you would consider raising your score toward acceptance. Please feel free to let us know if you have any follow-up questions.

Citations:

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

This paper proposes a new framework, SOUR, for identifying molecule structures from tandem mass spectra. To address the issue of scarce training data, the work augments the scarce experimental training data with spectra generated by a learned simulator. The synthetic mass spectra can be used to improve the training of the spectrum-to-fingerprint encoder. The fingerprint-to-molecule decoder is pretrained on a large molecule corpus, mid-trained on fingerprints perturbed to mimic encoder errors, and post-trained to revise candidate structures using simulator feedback. Experiments on multiple datasets report state-of-the-art retrieval and de novo generation accuracy.

Critical strengths

The performance is strong. SOUR outperforms competing methods on all datasets.

Critical weaknesses
  1. The writing is unclear. The method has many components, but the main text describes several of them only loosely and defers the details needed to understand them to the appendix. Here are a list of confusions I have:
  1. Line 191: "train the encoder on their simulated spectra with exact fingerprints." what does the encoder encode? What's the input, and what is the output? This paragraph does not have an explanation at all.

  2. Line 194: "D_real" contains labeled training spectra. Are these molecule-spectrum pairs?

  3. Line 199 to 200: is there any relatioinship between E_\theta and E_\theta^prob?

  4. Line 218, s indexing these three sources -- which three?

  5. Eq 2: what is l_fp?

  6. Line 231, "Phase 2 combines these two signals to construct fingerprint targets" --> which two signals?

  7. Line 239, the method for obtaining z_* is unclear.

  8. What does Eq 3 try to achieve?

...

  1. The work is more like building a system than proposing new machine learning methods. SOUR combines components that largely exist already. The contribution lies mainly in how these pieces are engineered together for this task. Because so many components are combined, it is also hard to tell which ones drive the gains. The main value of the paper is a strong application system for MS/MS structure elucidation. ICLR might not be the best venue for this work.
Questions for authors

Please check the questions from "Critical Weaknesses"

5 comments

We thank the reviewer for acknowledging our strong performance, the careful review of the methodology specifications, and the evaluation of the methodology. We believe there are several misunderstandings; we respond to your raised concerns below.

The writing is unclear. The method has many components, but the main text describes several of them only loosely and defers the details needed to understand them to the appendix.

Thank you for the suggestion. We have substantially revised most of the textual discussion of the paper, including the mathematical notation and elaboration of the methods section, in an effort to improve the clarity and reproducibility of the method. Amongst other earlier concerns, we hope the updated discussion of the methods section resolves the confusions you noted, but nonetheless address each of your specific questions below.

Here are a list of confusions I have: Line 191: "train the encoder on their simulated spectra with exact fingerprints." what does the encoder encode? What's the input, and what is the output? This paragraph does not have an explanation at all.

We followed the convention in papers on de novo generation from mass spectra. The encoder EθE_\theta aims to translate the spectra, a set of (m/z, intensity) tuples (y), into ECFP fingerprints (prediction) z^\hat{z}, a linearized representation of a structure that can be deterministically calculated from the structure xx [1]. We have elaborated upon the encoder input and output explicitly in this same section (Sec. 3, line 204-215).

Line 194: "D_real" contains labeled training spectra. Are these molecule-spectrum pairs?

Yes, D_real consists of experimentally observed molecule-spectrum pairs (as provided by each evaluation benchmark); we utilize “labeled” to indicate that there are experimentally known molecule labels available. We also denote D_real to distinguish from the other “unlabeled” datasets which involve synthetic spectra (e.g. D_nbr, D_qry,spec), whose spectra are simulated by GLACIER rather than experimentally measured. All sets D\mathcal D in Phase 1 and 2 augmentation are ultimately constructions of (molecule, spectrum) pairs. We did not originally make each candidate set explicit in the original submission, but the revised methodology now explicitly defines how each candidate set throughout Phase 1 and Phase 2 are constructed.

Line 199 to 200: is there any relationship between E_\theta and E_\theta^prob?

There is no distinction between the two. For simplicity, we have removed any references to E_\theta^prob. We maintain Eθ0E_{\theta_0}, Eθ1E_{\theta_1}, and Eθ2E_{\theta_2} to distinguish the starting encoder weights and weights after each phase of augmentation.

Line 218, s indexing these three sources -- which three?

In Phase 1, the three candidate sets utilized (and that this previously defined ss indexes from) are D_real (all training data points), D_nbr (synthetic neighbors of the training data), and D_qry,spec (the query-identified synthetic neighbors). We have updated the specification of the Phase 1 augmentation to explicitly define each of these three sets, how they are constructed, and additionally define D_Phase 1 as the union of these three sets (Sec 3.1, line 218-238). We have also now removed this particular usage of ss in favor of a simpler (y,z)∼DPhase 1(y, z)\sim D_{\text{Phase 1}} notation in the loss definition (Eq. 4).

Eq 2: what is l_fp?

In the previous Eq. 2, which has been updated to Eq. 4 in the new revision, l_fp is the fingerprint loss that aligns the output of the encoder E_\theta^1(y), \hat{z}, with the ground-truth fingerprints of the structures, z. For Phase 1, the fingerprint loss, which is elaborated upon in Appendix B.1, combines cosine similarity loss, Tanimoto similarity loss, and a cosine folded-fingerprint loss to maximize the fingerprint prediction reliability (since both cosine and Tanimoto similarity can be used to rank structures in candidate retrieval). For clarity, since the fingerprint loss is simplified in Phase 2 to only depend on soft Tanimoto distance, we have distinguished these two fingerprint losses as l_{fp, Phase 1} and l_{fp, Phase 2} in the methods section.

(to be continued)

Line 231, "Phase 2 combines these two signals to construct fingerprint targets" --> which two signals?

and

Line 239, the method for obtaining z_* is unclear.

The two signals being combined in Phase 2 are the spectral similarity and predicted fingerprint similarity; the latter signal requires the encoder prediction to be reasonably useful, which occurs much more strongly after Phase 1 adaptation. We specifically combine these two signals via reciprocal rank fusion, to utilize both ranks for retrieving candidates for this step. This defines a set of ranked candidates for each query spectrum, from which the medoid structure is identified and its corresponding fingerprint is collected as a representative target zi∗z_i^* to label each query spectrum, as defined in Section 3.2 (lines 257–261; full form in Eq. 12, Appendix B.2). We define the resulting training set for Phase 2, Dqry,fp(Q)\mathcal{D}_{\text{qry,fp}}(Q), as the test spectra with these pseudo-labels. We have revised the description of Phase 2 augmentation in Section 3.2 (lines 246–269) to clarify this process.

What does Eq 3 try to achieve?

Eq. 3, which has been revised to Eq. 6 in the updated revision, formulates the encoder training loss for Phase 2 as minimizing the fingerprint prediction distance to the representative pseudo-labels identified by this joint ranking process, as well as the original training structure-spectrum pairs. We use λreal\lambda_{\text{real}} to control the balance between real and synthetic pairs.

The work is more like building a system than proposing new machine learning methods. SOUR combines components that largely exist already. The contribution lies mainly in how these pieces are engineered together for this task. Because so many components are combined, it is also hard to tell which ones drive the gains. ICLR might not be the best venue for this work.

Thank you for your comment; we would like to clarify the technical contribution that may have been overlooked: as highlighted in the title, synthetic data simulation is an important but currently underemphasized factor in inverse task performance in MS/MS elucidation. Therefore, our key contribution is a new strategy focused on considering the target data distribution when utilizing synthetic data in various stages of model training and inference. For encoder training, both of our strategies of (1) first utilizing the target/query spectra to identify relevant synthetic data and (2) upon training with the data from (1), utilizing the improved prediction to identify additional synthetic neighbors, are novel strategies not found in other scientific inverse problem modeling to the best of our knowledge. On the decoder side, the mid-training to adapted fingerprints has been explored before in MS-GPT [3] and MSNovelist [2] but has not been framed as learning the errors of a synthetically-adapted model component; and the novel fragment-based conditioning we introduce unifies and streamlines the initial test-time regeneration strategy introduced by FRIGID [4]. We would thus respectfully disagree with the assessment “SOUR combines components that largely exist already”.

We also have differing thoughts on “Because so many components are combined, it is also hard to tell which ones drive the gains.” As the central technical contribution in our paper, simulated MS/MS data drives the performance gain. It is used in the following building blocks: Phase 1 and Phase 2 encoder training with augmented spectra, mid-training starting with synthetic-data-trained checkpoints, and test-time guidance with simulation. The level of gains contributed by each component is also supported by ablation studies in our submission:

  • Table 2: retrieval metrics, applying SOUR to different encoders
  • Table 3: de novo generation metrics, showing improvement from decoder choices (second block of rows) and from the encoder (third block of rows)
  • Figure 2, Figure 5, Table 8: isolated benefits of Phase 1 and Phase 2 to encoder improvement
  • Table 10: improvement from decoder mid-training on de novo generation accuracy
  • Table 11: isolated improvements from reranking, correction, and both combined

(to be continued)

ICLR might not be the best venue for this work.

We respectfully disagree. SOUR sits directly at the intersection of three active ICLR research threads: (i) query-conditioned test-time learning, test-time compute scaling, and learned self-correction [5–8]; (ii) forward-simulator-guided scientific inverse problems [9]; and (iii) de novo molecular structure elucidation from MS/MS and spectroscopic data [10–12]. We believe this paper fits within the theme of ICLR.

Beyond advancing MS/MS elucidation, SOUR addresses a general representation-learning question across asymmetric inverse problems (y→x^y \to \hat{x}) where learned forward simulators (x→y^x \to \hat{y}) are available while experimental labels remain scarce: how to scale imperfect synthetic supervision when naive synthetic expansion stalls or degrades (Figure 2). Reviewer BYUo commented that “the paper addresses a relevant problem of using imperfect synthetic spectra” when “simply increasing synthetic training data does not necessarily improve... performance.” Reviewer VFFR similarly noted that our formulation “follows quite naturally from the data imbalance problem” between a sample-starved representation encoder and a large-scale generator, with our adapted representations substantially improving unchanged external generative models (FRIGID and MS-GPT; Table 3); and Reviewer bA57 pointed out that “the approach is general and model-agnostic, with encoder adaptation improving performance across multiple backbones” (MIST, MVP, and FLARE; Table 2). Moreover, we identify other scientific inverse problems that also rely on synthetic data and that therefore stand to benefit from applying these same principles, such as inverse protein folding and other spectroscopy elucidation. Demonstrating that observation-guided synthetic selection, calibrated error adaptation, and simulator-feedback revision transfer across distinct encoder architectures and generative paradigms makes this work directly relevant to the ICLR community.

Thank you again for your detailed line-by-line reading, which has helped us substantially improve the clarity and presentation of our manuscript. We hope that the thoroughly revised Methods section, the component-level ablations, and our clarification of SOUR's algorithmic contributions and relevance to ICLR resolve your concerns, and we would be sincerely grateful if you would consider re-evaluating our work and raising your score. Please let us know if you have any further questions during the discussion period.

Citations:

  • [1] Rogers and Hahn, "Extended-Connectivity Fingerprints," JCIM 2010.
  • [2] Stravs et al., "MSNovelist: De Novo Structure Generation from Mass Spectra," Nature Methods 2022.
  • [3] Zhao et al., "MS-GPT: Rethinking MS/MS De Novo Structure Elucidation as Spectrum-Induced Posterior Querying of a Molecule-Language Model," arXiv:2607.23607, 2026.
  • [4] Bohde et al., "FRIGID: Scaling Diffusion-Based Molecular Generation from Mass Spectra at Training and Inference Time," ICML 2026.
  • [5] Hübotter et al., "Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs," ICLR 2025.
  • [6] Snell et al., "Scaling LLM Test-Time Compute Optimally Can Be More Effective than Scaling Parameters for Reasoning," ICLR 2025.
  • [7] Kumar et al., "Training Language Models to Self-Correct via Reinforcement Learning," ICLR 2025.
  • [8] Bushuiev et al., "One Protein Is All You Need," ICLR 2026.
  • [9] Levy et al., "Solving Inverse Problems in Protein Space Using Diffusion-Based Priors," ICLR 2025.
  • [10] Wang et al., "MADGEN: Mass-Spec Attends to De Novo Molecular Generation," ICLR 2025.
  • [11] Su et al., "SpectraLLM: Uncovering the Ability of LLMs for Molecular Structure Elucidation from Multi-Spectral Data," ICLR 2026.
  • [12] Noh et al., "IR-Agent: Expert-Inspired LLM Agents for Structure Elucidation from Infrared Spectra," ICLR 2026.
Reviewer_VFFROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies molecular identification from mass spectra (measurements of molecular fragmentation). The existing approaches typically first predict a “fingerprint” (a vector of structural features of the molecule) from the spectrum, and then use that fingerprint and atom counts in that molecule to predict the molecular structure. Basically, there are two maps here which compose: spectrum →\to fingerprint (encoder), fingerprint →\to molecular structure (generator). Training data for the generator is abundant, but the spectrum encoder needs paired spectrum-fingerprint data that is much scarcer. The authors propose SOUR to address this bottleneck.

At a high level, SOUR uses a simulator to create spectra for known molecules, whose fingerprints can be computed directly, supplying additional training samples for the encoder without requiring new experimental measurements. Then, the generator learns to handle imperfect fingerprints and to revise its predicted structures by checking how well their simulated spectra match the observations. The complete system improves molecular identification on two public benchmarks.

Critical strengths

From my understanding, this paper’s design and experimental progression follows quite naturally from the data imbalance problem -- the sample-starved encoder can induce errors which the downstream generator cannot fix alone. So the intervention is to improve the encoder mapping which has scarce experimental supervision, and then train the generator to handle the errors that remain.

The experiments then examine whether better fingerprint predictions (input to the generator) actually lead to better molecular identification. Indeed, improving the spectrum encoder benefits several unchanged generators (Table 3), showing that the bottleneck on fingerprint quality is significant. The selected synthetic data also improve performance with other encoder architectures (Table 1), providing further evidence that the approach is useful across different models.

Critical weaknesses

The authors argue that, in the SOUR pipeline generating data for the encoder, selecting simulated spectra that resemble the measured spectra makes the data more useful for the encoder. This is central to their proposed solution to the shortage of experimental training pairs. However, from my understanding, the controlled comparison supporting this claim changes both data selection and training: SOUR receives an additional training stage that the baseline using randomly selected simulations does not (Appendix B.3 points to this). So the results support the composite procedure for data creation, but leaves unclear how much the training-sample selection process contributes to the final success.

Questions for authors

SOUR uses simulations of known molecules to compensate for scarce experimental training data. Does its improvement over the baseline persist when the database contains few structures similar to the target? If possible, a breakdown of existing test results could help inform how well SOUR generalizes to molecules poorly represented in the database.

1 comment

We appreciate your recognition of our paper addressing the bottleneck of molecular generation, the quality of writing, and our thorough experiments. We address your concerns regarding the training setup and the possibility of over-relevance of the synthetic data as follows.

Training stages

The authors argue that, in the SOUR pipeline generating data for the encoder, selecting simulated spectra that resemble the measured spectra makes the data more useful for the encoder. This is central to their proposed solution to the shortage of experimental training pairs. However, from my understanding, the controlled comparison supporting this claim changes both data selection and training: SOUR receives an additional training stage that the baseline using randomly selected simulations does not (Appendix B.3 points to this). So the results support the composite procedure for data creation, but leaves unclear how much the training-sample selection process contributes to the final success.

Thank you for raising this important point. We agree that the benefits of the second additional training stage, the Phase 2, fingerprint-side augmentation, were not decoupled from the gains garnered from the Phase 1, spectral-side augmentation, and this might distort the data budget-accuracy relationships observed relative to random augmentation as reflected in Figure 1. In the updated PDF, Figure 2 now demonstrates the benefits of Phase 1 alone, showing a marked improvement over random augmentation. A separate Phase 1+2 augmentation trend further improves both the encoder objective (fingerprint reconstruction objective) and the downstream de novo accuracy metric. To further emphasize the value of Phase 2, fingerprint (prediction)-side alignment, we show in Figure 5 the effects of utilizing this augmentation approach to a randomly augmented encoder; the value of Phase 2 does improve, but to a lesser extent, suggesting that Phase 1, spectral-side alignment, is a critical first step to getting Phase 2 to also work.

Relationship between known structural similarity and ease of augmentation

SOUR uses simulations of known molecules to compensate for scarce experimental training data. Does its improvement over the baseline persist when the database contains few structures similar to the target? If possible, a breakdown of existing test results could help inform how well SOUR generalizes to molecules poorly represented in the database.

Thank you for this insightful suggestion. For both benchmarks, we have retrospectively binned the query spectra by the number of structural neighbors in the D_Phase 1 set (Figure 6, Appendix D.1.2). Note that the synthetic databases we have constructed exclude all test structures to prevent data leakage. The table below summarizes the same numbers:

Count of similar molecules (Tanimoto ≥0.5)NPLIB1: greedyNPLIB1: retrievedMassSpecGym: greedyMassSpecGym: retrieved
00.0290.0820.1590.068
1–90.0590.1130.1880.158
10–990.1300.1600.2620.351
100–9990.1700.1430.3210.444
≥1,0000.1280.0920.427†0.740†
†: fewer than 10 molecules observed in bin.

We encouragingly observe a positive improvement in encoder prediction for all bins, including even the set of query spectrum molecules that have no neighbor structures with Tanimoto similarity over 0.50 (first row), though the improvements are most pronounced for query spectrum molecules with more similar neighbors retrieved. This is observed for both neighbors found by Dqry,specD_{qry, spec} or DnbrD_{nbr} in Phase 1. Weaker relationships are observed for NPLIB1; this may reflect its original random training-validation-test split, which would permit greater structural similarity between training and test and subsequently easier generalization, compared to the graph-edit split in MassSpecGym.

We sincerely appreciate your insightful comments once again, and we hope that decoupling the two training stages and analyzing generalization across database similarity levels fully address your concerns. We would greatly appreciate your consideration in raising your score, and we remain happy to discuss further.

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

This paper presents SOUR, a framework for improving molecular structure elucidation from tandem mass spectra using synthetic data. The method selects query-relevant synthetic spectra to adapt spectral encoders, trains a molecular decoder to handle imperfect fingerprints, and uses forward-simulator feedback for iterative structure correction. Experiments on MassSpecGym and NPLIB1 show improvements in fingerprint prediction, molecular retrieval, and de novo structure generation, with SOUR-full achieving 17.98% and 27.35% Top-1 accuracy on the two benchmarks, respectively.

Critical strengths
  1. The paper addresses a relevant problem of using imperfect synthetic spectra to improve molecular structure elucidation when experimental annotations are limited.
  2. The experiments cover two benchmarks and multiple encoder architectures, with several ablation studies examining the proposed training and refinement strategies.
Critical weaknesses
  1. The synthetic-data scaling experiments do not fully isolate the benefit of query-conditioned data selection, since the SOUR setting includes additional pseudo-label adaptation that is not included in the reported synthetic-data budget.
  2. The contribution of the proposed iterative correction is not sufficiently separated from candidate reranking and additional inference computation, while the reported efficiency relies on amortized test-time training costs rather than actual per-query latency.
Questions for authors
  1. The motivation is reasonable, particularly the observation that simply increasing synthetic training data does not necessarily improve structure elucidation performance.
  2. The paper provides fairly extensive experiments, including comparisons with recent generation methods and tests of the encoder adaptation strategy across different architectures.
2 comments

We appreciate your thoughtful review and recognition of our motivations and thorough experiments. To address the general concerns with presentation, we have made revisions to the textual content of the paper, which can be found in the updated PDF. We acknowledge and address your two concerns in isolating the contributions of certain model design choices below.

Encoder-side scaling experiments

The synthetic-data scaling experiments do not fully isolate the benefit of query-conditioned data selection, since the SOUR setting includes additional pseudo-label adaptation that is not included in the reported synthetic-data budget.

Thank you for your suggestion. We have isolated the contributions of both Phase 1 and Phase 2 augmentation stages, including applying the same pseudo-label adaptation step of Phase 2 to the random augmentation settings to also assess its individual improvement. Below is an abbreviated table of encoder accuracies at differing levels of molecule counts (from 1K to 10M) (This table has been added as Table 8 in the updated PDF, and these trends have additionally been plotted in an updated version of Figure 2 and a newly added Figure 5):

Table: Numerical values for the MassSpecGym synthetic-data scaling study. All metrics are reported to three decimal places. Random augmentation and SOUR Phase 1 use the Phase 1 synthetic budget N. Series incorporating Phase 2 additionally include 17,556 query–pseudo-label pairs and use N+17,556.

Phase 1 pairs NWith Phase 2 N+17,556MIST Tanimoto: Random aug.MIST Tanimoto: Random + SOUR Phase 2MIST Tanimoto: SOUR Phase 1MIST Tanimoto: SOUR Phase 1+2SOUR-base Top-1 accuracy (%): Random aug.SOUR-base Top-1 accuracy (%): SOUR Phase 1SOUR-base Top-1 accuracy (%): SOUR Phase 1+2
1,00018,5560.2430.2390.2540.3090.0000.5183.389
5,00022,5560.2410.2820.2660.3380.0230.4734.004
10,00027,5560.2460.2920.2820.3600.0231.4355.263
62,50080,0560.2420.2890.3450.4170.0003.3216.943
250,000267,5560.2620.3250.3770.4320.0915.5888.863
1,000,0001,017,5560.2970.3380.4150.4780.4619.55212.298
5,000,0005,017,5560.3280.3740.4410.5071.07711.71115.448
10,000,00010,017,5560.2910.3100.4420.5050.32511.62615.379

Across all budgets N, the optimization budget is held fixed at 30,000 Phase 1 updates (plus 500 Phase 2 updates for series incorporating Phase 2), with the same batch composition for random and SOUR-selected augmentation (Appendix B.3), so the differences should only reflect synthetic data selection and volume.

SOUR benefits greatly from both phases, increasing near-monotonically while random augmentation suffers as larger numbers of synthetic data are utilized. The benefits of Phase 1 alone over random augmentation are observed from as early as only 1,000 structures. Applying Phase 2 to every SOUR Phase 1 model improves its encoder accuracy and de novo Acc@1. We further highlight the smaller, but still observable benefit of the Phase 2 augmentation to randomly augmented encoders in the MIST Tanimoto: Random + SOUR Phase 2 column, where the increases from MIST Tanimoto: Random are marginal in most cases, except for at 1,000 structures where encoder accuracy is instead worsened (from 0.243 to 0.239). We therefore credit the strength of the SOUR-augmented encoder to the complementary nature of both augmentation phases. The value of Phase 2 does improve, but to a lesser extent, suggesting that Phase 1, spectral-side alignment, is a critical first step to getting Phase 2 to further improve the encoder significantly.

(to be continued)

Benefits of iterative correction

The contribution of the proposed iterative correction is not sufficiently separated from candidate reranking and additional inference computation..

Thank you for this suggestion. In our original submission, Appendix D.3 (Correction and Candidate Selection; previously Table 9, now Table 11, lines 1498–1527) already compared SOUR-base with + GLACIER hybrid ranking only and with full iterative regeneration (SOUR-full), where each step adds inference-time compute on either GLACIER simulation or SOUR decoding. In the updated manuscript, we directly resolve your concern by expanding Table 11 with + one-round refinement only, which decodes GLACIER-guided revisions and ranks candidates solely by MISTSOUR\text{MIST}^{\text{SOUR}} fingerprint similarity without forward simulator reranking. Revision alone consistently improves exact recovery on both benchmarks, while the + second round control shows that spending further refinement compute brings diminishing additional gain. We hope this breakdown of Table 11 and the newly added ablation entry clarify the standalone contribution of forward simulator reranking and iterative revision, and could resolve your concern.

Compute costs

… the reported efficiency relies on amortized test-time training costs rather than actual per-query latency.

Actual untargeted LC-MS/MS workflows routinely acquire thousands of tandem mass spectra per sample run, making batch processing standard in practice. Moreover, the number of retrieved synthetic candidates per test spectrum follows the same per-query cap in both bulk and single-query modes, and in bulk adaptation both the synthetic pool size and training update budget scale up with the number of test queries (Appendix A.2 and Table 5, page 19), making the per-spectrum amortization in Figure 4 and Table 12 (Appendix D.5, lines 1529–1609) a fair reflection of per-sample cost.

For settings with only one or a few isolated queries (∣Q∣=1|\mathcal{Q}|=1), we also evaluated strict single-query adaptation with no cross-query sharing in Table 1 (page 8) and Appendix B.4 (Table 6, lines 1124–1166), retaining comparable accuracy. Below, we compare SOUR-full under bulk-query (amortized) and single-query (per-sample, unamortized) settings on the same NVIDIA RTX PRO 6000 GPU (in seconds per spectrum); even without amortization, single-query remains substantially faster than FRIGID (162.40 s on NPLIB1; 232.68 s on MassSpecGym) and faster than DiffMS on MassSpecGym (55.59 s):

DatasetSettingInference and correction (s)Synthetic generation (s)Phase 1 and Phase 2 training (s)Total (s)
NPLIB1Bulk-query (amortized, Table 12)2.995.525.6014.11
NPLIB1Single-query (per-sample, |Q|=1)2.994.7432.7140.44
MassSpecGymBulk-query (amortized, Table 12)3.610.0970.704.41
MassSpecGymSingle-query (per-sample, |Q|=1)3.610.2215.4719.30

Thank you again for your constructive feedback; we hope these clarifications, new ablations, and manuscript revisions resolve your concerns regarding soundness and presentation, and we would be grateful if you would consider raising your score or let us know if any questions remain.

Comments on the paper

2 comments

We thank all reviewers for their insightful feedback. Our paper identifies how the particular selection of synthetic data is an important but currently underemphasized factor in scientific inverse problems, including MS/MS elucidation. Naive augmentation scales poorly. We investigate this phenomenon and develop SOUR, a framework for efficient synthetic data utilization accommodating train-test distribution shift, intermediate representation errors, and test-time self-correction. It achieves 111% improvement over the next best de novo structural elucidation model on Acc@1 and 24% improvement over the next best candidate retrieval model on fingerprint reconstruction accuracy on MassSpecGym.

We appreciate reviewers’ positive feedback recognizing the novelty, clarity, and merits of our method. The importance of the problem, its relevance to the machine learning community, and the resulting approach is well-identified by reviewer BYUo, (“[our paper] addresses a relevant problem of using imperfect synthetic spectra”... “the motivation is reasonable, particularly the observation that simply increasing synthetic training data does not necessarily improve structure elucidation performance”); and reviewer VFFR, (“[our formulation] “follows quite naturally from the data imbalance problem”). SOUR’s significant improvements to the accuracy-efficiency frontier of mass-spec-based structure elucidation are recognized by all reviewers: BYUo (“Experiments on MassSpecGym and NPLIB1 show improvements in fingerprint prediction”); jdPm (The performance is strong. SOUR outperforms competing methods on all datasets); VFFR (“The complete system improves molecular identification on two public benchmarks”); and bA57 (“Comprehensive evaluations on the MassSpecGym and NPLIB1 benchmarks demonstrate the effectiveness of SOUR”). Reviewers BYUo, VFFR, and bA57 further appreciate the thoroughness of our experiments to isolate contributions: we “provide fairly extensive experiments (BYUo) to demonstrate that our encoder adaptation is "general and model-agnostic", consistently improving "multiple [encoder] backbones" (bA57) as well as "benefit[ting] several unchanged [decoder] generators” (VFFR).

We appreciate all the insightful suggestions from the reviewers, which we have addressed in rebuttal:

  • Reviewers BYUo, VFFR, and bA57 all asked for a separation of the improvements gained from each of the two phases of encoder augmentation, as well as the change observed from applying the second phase to a random augmentation control. In the rebuttal our updated PDF, we show that both phases of SOUR bring measurable improvements to the encoder accuracy and that the second phase of prediction-side sourcing of synthetic neighbors also helps random augmentation controls, though less strongly (updated PDF, Figure 1 and, Figure 5).
  • To observe generality, reviewers VFFR and bA57 both asked for a stratification of test molecule performance by ground-truth structural similarity to the several training sets utilized, including the synthetic neighbors collected and the decoder pretraining set. We have included these analyses in Figure 6 and Table 9, which highlight that performance expectably improves with exposure to additional close synthetic neighbors or to molecules in the decoder pretraining set. Most notably, even when the query spectra have no synthetic source neighbors found (as defined by 0.5 Tanimoto similarity), the SOUR adaptation nonetheless improves fingerprint reconstruction accuracy, which reinforces the generalization attained from applying SOUR.
  • Reviewer BYUo asked for controlled ablation of one-round refinement without simulator reranking to separate GLACIER-guided structural revision from candidate reranking on both benchmarks. Our expanded ablation in Appendix D.3 (Table 11) confirms that revision alone consistently improves exact-match recovery on both benchmarks, and combines complementarily with hybrid reranking in SOUR-full.
  • One reviewer, reviewer jdPm, focuses on several notational confusions, which we have addressed through a substantial revision of the method discussion. The methods section was also substantially revised following reviewer jdPm’s suggestion, which is also recognized in their follow-up comment.

All new results and our response to reviewer suggestions are reflected in the updated PDF.

Besides the suggestions made by the reviewers, we have made the following changes and additional results to the updated PDF:

  • Updated training recipes and total molecule counts consulted for both benchmarks (Appendix A.2).
  • Added stratified generalization analyses on MassSpecGym and NPLIB1, binning encoder performance improvements by Phase 1 synthetic neighbor coverage and end-to-end de novo accuracy by maximum structural similarity to the decoder pretraining corpus and Phase 1 synthetic sources (Appendices D.1.2–D.1.3, Figure 6, Table 9). (included already in last paragraph)
  • Added a one-round refinement control without simulator reranking, to separate the contribution of performance gains of GLACIER-guided structural revision from GLACIER+MIST hybrid candidate ranking on both benchmarks (Appendix D.3, Table 11).

In summary, our method, SOUR, identifies and addresses a critical criterion around the selection of synthetic data for utilization in scientific inverse problems. We do so by adapting the model components reliant on synthetic data to the query data distribution, then correcting the other components to adjust to the error and perform test-time correction to the prediction errors that arise. The improvements the SOUR framework brings to the encoders in MS/MS elucidation, alongside the novel decoder we introduce, complement existing models and the SOUR encoder-decoder combination achieves state-of-the-art performance on two benchmarks (MassSpecGym and NPLIB1) while also collecting notable speedups in inference costs. We believe our method brings significant insight and methodological novelty to addressing best practices for utilizing synthetic data in scientific inverse problems, and would thus be well-appreciated by the machine learning community and to the readers of ICLR.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.