acceptodds
Under review as a conference paper at ICLR 2027

ESDM: Efficient Spectrum-Conditioned Molecular Diffusion Models for molecular structure elucidation

Abstract

De novo molecular generation from tandem mass spectrometry (MS/MS) has become an important direction for unknown compound identification and molecular discovery. Diffusion-based methods have demonstrated strong molecular reconstruction capability by modeling conditional graph generation processes. However, their inference procedure still relies on long sequential reverse diffusion, resulting in substantial computational overhead and limited practical applicability in high-throughput scenarios. To address this issue, we propose ESDM, an efficient spectrum-conditioned molecular diffusion framework integrating Sampling Acceleration and Sampling Corrector. During training, a lightweight Conditional Timestep Tuner is optimized jointly with the diffusion denoising objective to learn spectrum-conditioned timestep preferences. During inference, the learned timestep allocation guides Sampling Acceleration to construct reduced reverse diffusion trajectories through non-uniform timestep scheduling and jump-step transitions, while Sampling Corrector improves sampling stability via lightweight posterior refinement on edge logits. Experimental results on the NPLIB1 and MassSpecGym benchmarks demonstrate that ESDM significantly reduces inference cost, with inference-time reductions of up to 58.9% on NPLIB1 and 62.8% on MassSpecGym, while maintaining competitive molecular reconstruction performance. Compared with DiffMS, ESDM achieves a favorable balance between generation quality and computational efficiency under multiple accelerated sampling configurations, highlighting its practical value for large-scale molecular generation and high-throughput molecular annotation.

Primary area:
generative models
Keywords:
Mass spectrometry, molecular generation, diffusion model, accelerated sampling, graph generation, de novo structure elucidation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Z8ChOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies inference acceleration for mass-spectrum-conditioned molecular graph generation, building on DiffMS. The proposed ESDM framework proposes a reduced, non-uniform reverse-diffusion schedule as well as a lightweight sampling corrector that modifies edge logits through temperature scaling, prior biasing, clipping, and timestep-dependent activation. It also proposes a learned conditional timestep tuner, although its relationship to the fixed schedules is unclear. Experiments on NPLIB1 and MassSpecGym shows substantial inference time cost reductions, while maintaining or slightly improving reconstruction accuracy compared to the original DiffMS setup. It also provides ablations that shows the contribution of timestep schedules and the corrector.

Critical strengths
  1. The paper addresses a practically relevant computation efficiency problem and shows good efficiency–accuracy trade-off. Repeated denoising is expensive for molecular graph diffusion, especially when many candidates are generated per spectrum. The configurations in Table 1 reduce runtime by 58.9% on NPLIB1 and 62.8% on MassSpecGym while slightly improving Top-1 accuracy relative to the original DiffMS setup. These provide useful evidence that by trading off the number of generated candidates, a shorter sampling trajectory can improve the efficiency-accuracy front.

  2. The paper includes controlled comparisons supporting some benefit from the proposed sampling choices. In Table 2, quadratic scheduling improves NPLIB1 Top-1 accuracy from 5.230% to 5.604% relative to uniform scheduling at the same 100-candidate, 80-step budget. At the 10-candidate, 80-step setting, adding the corrector improves Top-1 accuracy from 2.739% to 3.860%. These comparisons support the practical value of investigating schedule design and inexpensive edge-logit adjustments beyond naively reducing the number of denoising steps.

Critical weaknesses
  1. The method being evaluated is not consistently specified, making its central contribution difficult to assess or reproduce. The abstract and implementation section describe inference-only modifications that preserve the original denoiser and training strategy. In contrast, Sections 3.2–3.3 describe a conditional timestep tuner trained jointly with the denoising model and used to produce spectrum-dependent schedules. Figure 1 labels the tuner as optional, but the paper does not identify which configurations use it when reporting the performance, while Algorithm 1 and the tables instead suggest predefined schedules. There is also a concrete inconsistency in the schedule definition: Equation (2) specifies tK=Tt_K=T and t0=0t_0=0, whereas Equation (4), ti=T[1−(i/K)2]t_i=T[1-(i/K)^2], gives the opposite endpoints. Traversing Equation (4) from i=0i=0 to KK also produces increasingly large timestep gaps toward the end of denoising, contrary to the stated motivation of finer late-stage transitions. These ambiguities affect the implemented algorithm and the interpretation of its results, rather than only its presentation.

  2. The experiments do not yet establish how much of the reported quality improvement is attributable to the proposed components, or how reliably it generalizes. The headline comparisons increase the candidate count from 100 to 150 while reducing the number of denoising steps. This is a legitimate way to explore an efficiency–accuracy trade-off, but it does not isolate the effect of the schedule or corrector on reconstruction quality. At the matched 100-candidate setting on NPLIB1, the accelerated configuration reduces Top-1 accuracy from 7.098% to 5.600%. Moreover, the matched corrector ablation on MassSpecGym shows only a small change, from 0.280% to 0.293%, without repeated-run uncertainty estimates. The paper also claims improved molecular validity and sampling stability without reporting direct quantitative measurements of these properties. Finally, Appendix A.1 describes Bayesian hyperparameter optimization in “test-only mode,” but does not identify the data partition used for selecting configurations. This leaves the independence of the final evaluation unclear; the wording alone does not establish that test labels were used for tuning.

  3. The overall model performance on MSG and CANOPUS seems primitive. Over the last 12 months, ms/ms structure elucidation has seen great progress in model development and models like FRIGID/MS-GPT and much more others have set much higher performance level, and their sequence-based architecture is also much more compute-efficient. Optimizing on graph diffusion steps seems incremental and not that of a central concern if better architectures could already alleviate both sides of the issue, not to mention the fact that this paper is not putting forward techniques that could fundamentally improve modeling capabilities built on the already widely-used DiffMS architecture.

  4. Citation errors and additional presentation inconsistencies. Several baseline citations in Section 4.1 do not identify the corresponding methods. MADGEN is attributed to Shrivastava et al. (2021b), which is the MassGenie paper, whereas MADGEN was introduced by Wang et al. (2025). Likewise, MIST+Neuraldecipher is cited to Hájek et al. (2023), a different study on identification from GC-EI-MS spectra, rather than the Neuraldecipher work by Le et al. (2020). These mismatches make it difficult to trace the baseline results and should be corrected. Beyond the schedule and tuner inconsistencies discussed above, several implementation details remain underspecified: the construction and normalization of the edge prior PedgeP_{\mathrm{edge}}, the correction interval nn, and whether the activation threshold tcorrt_{\mathrm{corr}} uses normalized or integer timesteps. Section 3.4 motivates correction during late-stage denoising, while Appendix A.1 emphasizes earlier activation without clearly reconciling the two descriptions. Together, these issues need substantial clarification to make the paper more verifiable.

Questions for authors
  1. Which method produced each reported result? Was the conditional timestep tuner enabled in any row of Tables 1–2? If so, please specify its architecture, training objective, training budget, and how its output enters schedule construction, and distinguish its contribution from that of predefined schedules. If it was not used, please reconcile the method description with the inference-only implementation.

  2. What timestep sequence was actually executed? Please provide an example of the implemented quadratic schedule, including traversal order, integer rounding, and endpoint handling, and reconcile Equations (2) and (4) with the claim of denser late-stage denoising. Clarifying whether this is a notation error or an implementation issue would materially affect my assessment.

  3. How robust are the component-level gains under controlled budgets? Can you provide the schedule-only and schedule-plus-corrector results at the same candidate count and step budget as the principal ESDM configurations, together with uncertainty across sampling seeds? A matched-runtime comparison against DiffMS would also help establish the practical trade-off. If validity and stability are central benefits, please report the corresponding quantitative measurements.

  4. Which data were used for hyperparameter selection? In Appendix A.1, does “test-only mode” refer to an inference/evaluation execution mode or to evaluation on the held-out test partition? Please specify the tuning and final-evaluation splits, explain the relationship between the CANOPUS tuning data and the NPLIB1 evaluation, and state whether any final-test labels or metrics informed the choice of schedules, sampling budgets, or corrector parameters.

Reviewer_3k1kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose ESDM, which speeds up diffusion models that generate molecular graphs from mass spectra. It skips timesteps with a quadratic schedule instead of running the full reverse process, and post-hoc intervenes the predicted edge logits with temperature scaling, a prior bias, and clipping. By making their components inference-only, they require no retraining and are claimed to transfer to other existing diffusion-based generators.

Critical strengths

Inference cost seems like a true problem here (DiffMS takes 348 hours on MassSpecGym) and an inference-only fix is the right kind of answer, since it transfers to other discrete graph diffusion models.

Critical weaknesses

The comparisons in Table 1 are not matched on any axis, so I cannot verify the authors' efficiency claims. The table 1 compares DiffMS at 100 samples, and in this scenario, the Top-1 performance rises with sample count by construction; the matched 100-sample row gives 5.600%, below the baseline’s 7.098%. Two things would make the claim checkable: accuracy at a fixed sample count across step budgets, and accuracy at matched wall-clock time or matched NFE. As it stands I can neither confirm nor reject that ESDM improves the speed-quality trade-off.

The methodological novelty is low. Jump-step transitions and non-uniform schedules are DDIM with respacing, and the corrector is temperature scaling plus a logit bias plus clipping. All these techniques are not new, especially in the broad diffusion domain. At this current point, the method seems more like a engineering problem rather than a research problem. I would like the authors to better specify as to why this modality requires this set of solutions, when there are numerous other possible ways to handle the exact same problem, demonstrated in various 'other' domains.

Questions for authors

Refer to the weakness.

Reviewer_8fNKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

An accelerated denoiser architecture: ESDM- is proposed for molecule generation through discrete graph diffusion, based on DiffMS. The molecular graphs are reconstructed by conditioning the denoiser with mass spectra. It reduces reverse denoising from 500 steps to 80–120 using non-uniform timestep schedules. An additional corrector modifies edge logits through temperature scaling, prior bias, and clipping. The authors evaluate on NPLIB1 and MassSpecGym, reporting roughly 59–63% lower inference time while maintaining competitive molecular reconstruction accuracy.

Critical strengths
  1. Reducing the compute load of discrete diffusion pipelines is an important problem to address in molecule generation. The reported runtime improvements are substantial at good accuracies.

  2. The sampling corrector is quite simple and does not require additional denoiser evaluations. The ablations show some improvement in reconstruction quality under accelerated sampling.

Critical weaknesses
  1. The justification for skipping intermediate steps is not convincing enough: The jump-step transition formula is provided, but there is no analysis of how skipped denoising evaluations affect reconstruction error. Intermediate steps might be important for recovering molecular topology, especially bonds. With 100 candidates, Top-1 accuracy decreases from 7.098% to 5.604% when reducing sampling to 80 steps without correction. The headline improvement also uses more generated candidates, making attribution difficult. Equation (4) appears inconsistent with the claimed finer timestep allocation near final reconstruction.

  2. The corruption process and evaluation need better characterization: The method considers node and edge corruption, but effective corruption rates at different timesteps are unclear- this should be reported. If corruption is relatively mild, reconstruction might be easier than expected.

  3. Benchmarking is also limited to two spectrum-conditioned datasets and primarily one denoiser architecture.

  4. Generalizability to other discrete graph diffusion models should be reported. Also, given the potential of this architecture, a sweep should be done with different time schedules, corruption levels and conditioning of mass spectra being varied.

Questions for authors
  1. Can the authors report atom- and bond-reconstruction errors across different sampling-step budgets, keeping the number of generated candidates fixed?

  2. What fraction of node and edge features is corrupted at different timesteps? Can the exact forward corruption schedules be specified?

  3. Can the authors clarify Equation (4) and whether the Conditional Timestep Tuner is actually used in the reported experiments?

  4. Can the authors provide a comparison against other accelerated samplers at matched inference budgets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.