acceptodds
Withdrawn from ICLR 2027

Cost-Aware Hierarchical Evidence Routing for Molecular Retrieval from Mass Spectra

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular retrieval from tandem mass spectra aims to rank candidate structures based on their compatibility with observed fragmentation patterns. Existing approaches based on global spectrum–molecule representations do not adequately distinguish candidates with subtle structural differences and therefore struggle to rank structurally similar molecules accurately. Fine-grained structural evidence can improve discrimination among these candidates, but uniformly evaluating such evidence for all candidates incurs substantial computational costs and overlooks query-specific refinement needs. To address this challenge, we propose MS-CHER, a two-stage molecular retrieval framework with Cost-aware Hierarchical Evidence Routing. The key idea is to tailor fine-grained evidence computation to the refinement needs of each query. We obtain an initial candidate ranking through global spectrum–molecule matching using structure-aware molecular representations. Building on this ranking, a hierarchical routing policy determines a query-dependent refinement scope and selectively activates structural evidence within this scope according to predicted ranking benefits and computational costs. Experiments on multiple MS/MS retrieval benchmarks demonstrate competitive retrieval performance across diverse candidate settings. Notably, compared with exhaustive evidence evaluation on MassSpecGym, MS-CHER reduces structural evidence evaluation runtime by approximately 79.2% and 82.4% in the mass and formula settings, respectively, while maintaining comparable retrieval accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Retrieval, Mass Spectrometry, Adaptive Computation, AI for Science

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

This paper presents MS-CHER, a two-stage framework for molecular retrieval from tandem mass spectra, aiming to rank candidate molecular structures based on their compatibility with observed fragmentation patterns. Existing approaches based on global spectrum–molecule representations often struggle to distinguish candidates with subtle structural differences, while uniformly evaluating fine-grained structural evidence for all candidates incurs substantial computational costs and overlooks query-specific refinement needs.

The proposed framework consists of three main components.

(I) It performs global spectrum–molecule matching using structure-aware molecular representations learned through self-supervised pretraining to obtain an initial candidate ranking.

(II) It introduces two fine-grained structural evidence modules: LOCAL, which models correspondences between spectral peaks and molecular substructures, and FRAG, which evaluates the compatibility of observed spectra with candidate-specific fragmentation paths.

(III) It employs cost-aware hierarchical evidence routing, consisting of a Scope Router that selects a query-dependent prefix of the initial ranking for refinement and an Evidence Router that determines which structural evidence modules to activate within the selected prefix based on their predicted ranking benefits and computational costs.

Experiments on three MS/MS retrieval benchmarks, MassSpecGym, SpectraVerse, and NPLIB1, demonstrate competitive retrieval performance. On MassSpecGym, MS-CHER reduces structural evidence evaluation runtime by approximately 79.2% and 82.4% under the mass and formula settings, respectively, compared with exhaustive evidence evaluation, while maintaining comparable retrieval accuracy.

Critical strengths

(I) The paper combines several recent techniques, including JEPA-based molecular pretraining, fine-grained structural matching, and cost-aware routing. The authors also carefully design the multi-stage training procedure and the selection of molecular fragments. Although most of the individual techniques are based on existing methods, their combination is reasonable and useful for molecular retrieval.

(II) The proposed method achieves large improvements over the existing methods, especially in the mass setting. For example, on MassSpecGym, Hit@1 improves from 12.23% for the best baseline to 30.88%. In addition, the method reduces the runtime for structural evidence evaluation by about 79–82% compared with full evaluation, while maintaining similar retrieval accuracy. These results show the effectiveness of the proposed method.

Critical weaknesses

(I) The Related Work section introduces several existing methods, but it does not sufficiently discuss their limitations or explain how the proposed method improves upon them. In particular, the relationship between MS-CHER and previous studies on cost-aware ranking and fragmentation-based molecular identification is not clearly explained. Since the proposed method combines several existing techniques, a more detailed discussion of related work is needed to clarify the novelty of the method.

(II) The proposed method consists of many components, including structure-aware molecular pretraining with several loss functions, LOCAL, FRAG, and two routing modules. Although the paper provides some ablation studies, it is still unclear which components are important for the performance improvements. For example, the contributions of individual pretraining objectives are not evaluated. Also, the experimental results show that FRAG provides most of the retrieval accuracy improvements, but the proposed routing method is not compared with a simpler approach such as fixed Top-K reranking using FRAG. Therefore, it is difficult to understand whether all the proposed components are necessary to achieve the reported accuracy and computational efficiency.

Questions for authors

(I) Table 13 shows that FRAG gives most of the improvement in retrieval accuracy. However, it is unclear how much the proposed hierarchical routing contributes to the performance. Could the authors compare MS-CHER with simpler methods, such as fixed Top-K reranking using FRAG and confidence-based reranking using Stage-1 scores? I think these methods should be compared under similar computational costs to show the effectiveness of the proposed routing method.

(II) The proposed pretraining method uses several techniques, including structure-aware masking, latent prediction, and additional loss functions. However, Table 14 only compares the full pretraining method with a model without pretraining. Could the authors provide more detailed ablation studies to show which components are important for the performance improvement? In particular, I would like to know how much structure-aware masking and the additional loss functions contribute to the results.

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

This paper proposes MS-CHER, a two-stage molecular retrieval framework that refines global spectrum–molecule rankings using local structural matching and fragmentation-path matching. Two routers select the candidate scope and evidence types to reduce structural-evidence computation.

Critical strengths
  • Well-motivated problem framing. The observation that not every query needs fine-grained structural evidence is sensible and clearly illustrated , and cost-aware evidence allocation is a natural fit for large candidate sets.
  • Strong results under a leakage-aware protocol. The paper follows the MassSpecGym v1.5 protocol and excludes formula information from the mass setting. Stage 1 alone reaches Hit@1 of 17.92% and the full system reaches 30.88%.
Critical weaknesses
  • Insufficient validation of the central routing contribution The ablation lacks simpler baselines under matched compute, such as fixed top-K refinement, Stage-1-margin-based truncation, or FRAG-only reranking. Since FRAG alone accounts for most of the gain and nearly matches Full-Eval, it is unclear whether the dual-router dual-expert design beats a simple FRAG reranker at the same runtime. Moreover, Stage 1 uses pretrained DreaMS features not used by compared baselines, so the improvement cannot be attributed to the proposed design.
  • Cross-dataset generalization lacks an overlap audit Section 4.2 presents results on SpectraVerse and NPLIB1 as evidence of cross-dataset transfer, but does not report molecular overlap or duplicate and near-duplicate spectra between these evaluation sets and the training and validation data.
Questions for authors
  • At matched measured computational budgets, does MS-CHER outperform fixed top-K refinement, Stage-1 confidence-based truncation, and FRAG-only reranking, and do these advantages persist across multiple budget levels?
  • How much molecular and spectral overlap exists between the SpectraVerse/NPLIB1 evaluation sets and the training and validation data, and how does each method’s performance change after removing these overlaps?
  • What are MS-CHER’s end-to-end latency, peak GPU memory, and training costs after accounting for Stage-1, routing, fusion, and necessary preprocessing, and how much computation does it save on the larger candidate pools of SpectraVerse and NPLIB1?
  • What are the ground-truth coverage rates of the selected scopes, expert activation frequencies, and agreement between predicted utilities and realized ranking gains, and do these quantities remain stable under cross-dataset transfer?
Reviewer_XG98OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes MS-CHER, a two-stage framework for retrieving molecular structures from MS/MS spectra. Stage 1 ranks all candidates by a global spectrum-molecule score. Stage 2 seems to refine the ranking with two structural evidence modules. The experiment results on MassSpecGym, SpectraVerse, and NPLIB1 show strong retrieval performance.

Critical strengths
  1. The method performs well. On MassSpecGym it achieves the best results on all metrics in both the mass and formula settings, with a large margin in the mass setting.

  2. The high-level idea seems sound. Adding a second-stage reranking module is a common and effective design for retrieval problems, and allocating the reranking computation per query is a reasonable way to reduce its cost.

Critical weaknesses

The paper is VERY hard to read.

  1. Section 3.2: What does the context graph CC contain? The main text only says that TT is selected "with reference to its Murcko scaffold." Does CC contain only peripheral pieces? Does it provide much information for predicting the target region?

  2. fθf_\theta and fθˉf_{\bar{\theta}} are not well defined.

  3. The rationale of Section 3.2 is hard to understand. I assume the objective in Eq. 3 is minimized. If so, what is Laux\mathcal{L}_{\text{aux}}? It is essential to the model and should be described in the main text rather than only in the appendix.

  4. What is the relationship between dqd_q and bqb_q in Eq. 5?

  5. Many notations in Eq. 7 are not explained.

  6. Is the query-dependent refinement scope trained? If so, what is the training loss?

  7. Line 237 defines E={L,F}\mathcal{E} = \{L, F\}. What are LOCAL and FRAG? They are used in the routing equations before they are defined.

Questions for authors

I feel that the paper needs a thorough revision to be understandable. Every symbol should be clearly defined.

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

Related papers

Discussion (0)

Sign in to comment.