acceptodds
Withdrawn from ICLR 2027

CoRe-Gen: Robust Spectrum-to-Structure Generation under Imperfect Fingerprint Conditions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Molecular structure elucidation from tandem mass spectra (MS/MS) remains challenging, particularly for de novo generation beyond database coverage. A common approach decomposes the task into spectrum-to-fingerprint prediction followed by fingerprint-to-structure decoding, enabling the use of large-scale molecular corpora. However, at deployment, the decoder relies on predicted rather than oracle fingerprints, introducing structured errors that propagate into generation. Moreover, common autoregressive decoders threshold fingerprint probabilities and serialize active bits as discrete input tokens, blocking molecular generation gradients from the spectrum encoder. We present CoRe-Gen, which improves the intermediate condition through synthetic-spectrum pretraining, matches deployment-time noise through frequencyaware corruption, and introduces a differentiable fingerprint-to-memory bridge for joint encoder–decoder finetuning. The bridge preserves soft bit confidences and conditions a structure-aware autoregressive decoder without discrete fingerprint tokenization. CoRe-Gen achieves 21.49%/35.42% Top-1/Top-10 exact-match accuracy on NPLIB1 and, under encoder-aligned evaluation, 20.02%/22.45% on MassSpecGym, while retaining efficient autoregressive inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spectrum-to-Structure Generation, Tandem Mass Spectrometry, Molecular Fingerprints, Robust Molecular Generation, Conditional Molecular Generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dEnSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces CoRe-Gen, a framework for spectrum-to-structure generation under imperfect fingerprint conditions. It combines synthetic-spectrum encoder pretraining, corruption-aware decoder pretraining, and a differentiable fingerprint-to-memory bridge for joint encoder–decoder finetuning. The authors also incorporate structural information into autoregressive decoding.

Critical strengths
  • Jointly updating the spectrum encoder through the molecular generation loss is a useful idea.
  • Addressing the mismatch between clean fingerprints used during training and predicted fingerprints used at inference is well motivated.
Critical weaknesses
  • The data separation and potential leakage issues in large-scale synthetic-spectrum pretraining need more explanation. In Section 4.1, the authors state that molecules overlapping with the test sets are removed from both pretraining corpora. However, I strongly suspect that the current pretraining data contain molecules highly similar to the test molecules. I suggest reporting the similarity distribution between current using pretraining dataset and the test set, and pretraining both the encoder and decoder on the standard pretraining dataset released with MassSpecGym1.5 [1]. Table 5 shows that removing encoder or decoder pretraining causes the largest performance drops further support the concern, while the other modules have only minor effects, suggesting substantial over-engineering.

  • Some terms and symbols are unclear. What does “encoder-aligned evaluation” mean in the abstract? In Equation (2), gg is used without being introduced.

  • Updating the encoder through the decoder is an interesting idea, but the descriptions in the Introduction and Related Work do not clearly explain how backpropagation works. On page 4, why are the fingerprint dimensions divided into C=64C=64 groups, and what is the rationale behind the definition of cic_i? A direct MLP connection would also allow backpropagation. The authors should include this ablation.

  • Section 3.2 is hard to follow. Which dataset does “training corpus” refer to? Why is grouping necessary, and has an ablation without grouping been conducted? What is the rationale behind ϵ\epsilon and α\alpha in Equation (4)? Grouping is introduced earlier, but its connection to the subsequent equations is unclear, as is the use of kk and pjp_j in Equations (6)–(8). The authors should include ablations with random bit flips without grouping, using both occurrence-based and fixed probabilities.

  • In Section 3.3, how was the five-component tuple (pt,at,bt,qt,dt)(p_t,a_t,b_t,q_t,d_t) in Equation (9) chosen? Are there ablation experiments supporting this choice?where does the rule-dependent mask mtm_t in Equation (11) come from?

[1] MassSpecGym in the Wild

Questions for authors

My biggest concern is potential data leakage during pretraining. In addition, the proposed method appears overly engineered and offers limited insights.

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

The paper proposes CoRe-Gen, a two-stage MS/MS-to-structure pipeline, focusing on inaccurate predicted fingerprints and the mismatch between decoder training and deployment conditions. CoRe-Gen combines synthetic-spectrum encoder pretraining, frequency-aware fingerprint corruption, a differentiable fingerprint-to-memory bridge, and structure-aware autoregressive decoding. The paper reports substantial improvements on NPLIB1 and strong exact-match performance on MassSpecGym.

Critical strengths
  • The problem framing is clear and well motivated. Treating predicted fingerprints as an imperfect condition is sensible, and the bit-frequency analysis is informative.
  • The NPLIB1 results are promising: CoRe-Gen achieves 21.49%/35.42% Top-1/Top-10 exact-match accuracy, improves structural similarity metrics, and maintains efficient inference.
  • Autoregressive decoding is much faster than diffusion-based alternatives.
Critical weaknesses

Limited and partly overstated novelty. Continuous probabilistic fingerprint conditioning is not new; prior works such as MSNovelist already consume real-valued fingerprint vectors directly. The other components (simulated-spectrum pretraining, condition noise injection, factorized token embeddings, formula-based masking) are each well-known techniques, so the contribution is mostly a careful engineering combination. The joint-finetuning gain enabled by the bridge is also modest (Top-1 19.54% → 21.49%).

  • MassSpecGym headline result does not reflect the proposed method. The 20.02% Top-1 on MassSpecGym relies on the released DiffMS encoder. With the authors’ own pretrained encoder, Top-1 drops to 1.23% (Table 7), below DiffMS (2.30%), so the encoder-side contribution does not transfer to this benchmark. Placing this result in Table 2 alongside methods using their own encoders is misleading. In addition, Top-1 MCES worsens (15.19 → 16.09) as exact match rises 16×, and Top-1/Top-10 are nearly identical (20.02 vs 22.45). This needs explanation.

  • The decoder-level claim is not tested under controlled conditions. If the key claim is that the soft bridge plus structure-aware decoder beats thresholded tokenization, the natural experiment is to plug the same DiffMS encoder into MS-BART or a MolForge-style decoder. In the current comparison, MS-BART uses its own weaker encoder (Table 6), which confounds encoder quality with decoder design. Likewise, “w/o joint finetune” in Table 5 does not isolate soft versus hard conditioning. An ablation that keeps the decoder and replaces the bridge with thresholded active-bit tokens is missing.

  • Missing comparison with the latest baseline. The paper omits FRIGID (Bohde et al., ICML 2026), the most recent state-of-the-art method in this area. FRIGID uses the same setting (MIST-predicted fingerprint plus molecular formula as the condition) and reports results on both NPLIB1 and MassSpecGym, so it is directly comparable.

Questions for authors
  • Could the authors clarify how the proposed fingerprint-to-memory bridge differs from the continuous fingerprint conditioning already used in prior works such as MSNovelist?
  • Could the authors provide an ablation that keeps the encoder and decoder fixed and replaces the bridge with thresholded active-bit tokens, to isolate the contribution of the bridge itself?
  • Could the authors include FRIGID as a baseline on both NPLIB1 and MassSpecGym under the same formula and candidate-budget settings, and discuss the main differences between the two approaches?
  • Despite synthetic pretraining, the authors' own encoder remains clearly weaker than the released DiffMS checkpoint on MassSpecGym. Could the authors offer any analysis of why the pretraining does not transfer to this benchmark?Besides, please report the result of the full pipeline (with your own encoder) in the main table, or at least clearly mark that the MassSpecGym row uses a different encoder from the other methods.
  • When switching to the released encoder, exact match rises from 1.23% to 20.02% while Top-1 MCES increases from 15.19 to 16.09. Could the authors explain this and provide the per-sample MCES distribution?
Reviewer_XuQTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a new method of generating molecules based on mass spectra. It has made improvement over an architecture of spectrum -> fingerprint -> SELFIES (molecules). The improvement including: pretraining the fingerprint encoder on a large dataset and training the decoder by injecting noise to fingerprints. The new system shows strong performance on NPLIB1 and MassSpecGym, ranking the best or the second best in the comparison against a list of baselines.

Critical strengths

The proposed research work is well motivated. The investigation of fingerprint entries on MassSpecGym reveals that some entries cannot be predicted well due to insufficient training. Pretraining the encoder on a large dataset is reasonable.

The method of adding noise to fingerprints to robustify the decoder is also reasonable since the decoder should not safely depend on the predicted fingerprints.

Critical weaknesses
  1. Can you clarify the motivation for using fingerprints: "they (recent neural models for de novo generation) remain constrained by limited paired supervision and are often unstable under real-world spectral variability. A practical compromise is fingerprint-conditioned generation".

Why is using fingerprints as an intermediate step a practical compromise? Several previous methods, including MADGEN and DiffMS, do not use fingerprints. What's the advantage of predicting the fingerprint first?

  1. Does the fingerprint contain all information from the mass spectrum about the molecule? Is it possible that some information about the molecule is lost? Does the proposed decoder receive less information than a generative model conditioned on a mass spectrum?

  2. "Our independently trained MIST encoder did not reproduce the fingerprint quality of the publicly released DiffMS checkpoint despite following the released code, configurations, and reported hyperparameters." I think this is not acceptable because this is not a fair comparison against other methods.

  3. The innovation is mainly in the pipeline, not the methodology. I am not sure that ICLR is the best publication venue.

Questions for authors

Please check weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.