acceptodds
Under review as a conference paper at ICLR 2027

MUSE: Mixture-to-Substructure Elucidation via Disentangled Set Prediction

Abstract

Infrared (IR) spectroscopy is a crucial technique for molecular characterization in chemistry and materials science. Existing deep learning-based IR spectral analysis methods are largely confined to the idealized single-compound scenario, with limited exploration of more realistic multi-component mixtures. This challenge stems from two main issues: signal coupling and attribution ambiguity caused by overlapping absorption peaks, and the entanglement between a component's chemical structure and its concentration in the spectral features. To overcome these challenges, we introduce **MUSE**, an end-to-end framework that reformulates **M**ixt**u**re-to-**S**ubstructure **E**lucidation as a one-shot set prediction problem. By formulating a permutation-equivariant decoding scheme via learnable slot queries and cross-attention, the model simultaneously captures the physical additivity and global competition among components in a single forward pass. This joint modeling significantly mitigates the error accumulation that plagues traditional single-target methods. To overcome the intrinsic entanglement between chemical identity and abundance, MUSE partitions each slot embedding into structural and concentration subspaces, penalizing their empirical cross-covariance to encourage linear decorrelation. Building upon this latent disentanglement, it incorporates conditional flow matching during the generative process to provide physically interpretable supervision through generative reconstruction. Extensive experiments on real and simulated datasets demonstrate that MUSE significantly outperforms existing methods, establishing its effectiveness and robustness in complex mixture analysis. Our code can be found at https://anonymous.4open.science/r/MUSE-7716.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Infrared spectroscopy, Mixture-to-substructure elucidation..

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_mbhtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MUSE (Mixture-to-Substructure Elucidation), an end-to-end framework for identifying component-level functional groups from mixed infrared (IR) spectra. The work addresses two key challenges in mixture analysis: spectral overlap among different components and the entanglement between chemical identity and concentration. To tackle these challenges, the authors formulate mixture analysis as a one-shot set prediction problem and introduce a slot-based decoder with cross-attention to jointly predict the functional groups of multiple components. Each slot representation is further divided into structural and concentration subspaces, with a cross-covariance regularization term to encourage disentanglement. Additionally, conditional flow matching (CFM) is employed to reconstruct individual component spectra, providing generative supervision for spectral separation. Extensive experiments on simulated datasets (QM9S and Alberts et al.), NIST-derived mixtures, and physically measured mixture spectra demonstrate that MUSE consistently outperforms existing baselines. The method also shows improved robustness to increasing mixture complexity, low-concentration components, spectral noise, and imperfect concentration priors. Ablation studies further validate the contributions of the proposed components. Overall, the paper provides a promising approach to automated substructure elucidation in complex chemical mixtures.

Critical strengths
  1. Important and practical problem formulation. The paper addresses component-level substructure elucidation from mixed IR spectra, extending beyond the simplified single-compound setting. Formulating this task as one-shot set prediction is well motivated and aligns with the permutation-invariant nature of chemical mixtures.

  2. Well-motivated framework design. MUSE integrates slot-based set prediction, structural–concentration disentanglement, and conditional flow matching into a unified framework. These components address distinct challenges in mixture analysis and provide a coherent solution.

Critical weaknesses
  1. Limited methodological novelty. The main components, including slot-based set prediction, cross-covariance regularization, and conditional flow matching, are largely based on existing techniques. The authors should better clarify the unique methodological contributions beyond their integration and provide stronger evidence for the necessity of CFM over simpler reconstruction objectives.

  2. Limited real-world validation. Most experiments rely on synthetically mixed spectra, while the physically measured mixture datasets contain only 30 and 29 samples, respectively. Larger and more diverse experimental datasets, particularly involving mixtures with more than two components, are needed to establish practical generalization.

Questions for authors

See Weaknesses.

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

This paper studies per-component substructure elucidation from infrared (IR) spectra of multi-component mixtures. Given a mixture spectrum, the number of components K, and optional priors (per-component concentrations and molecular formulas), the task is to predict a multi-label vector over 37 predefined functional groups for each component. The authors propose MUSE, which formulates the task as set prediction, which includes an 1D-CNN and transformer encoder, a slot-query-based slot decoder through cross attention and Hungarian matching, a disentanglement module with cross-covariance penalty encourages linear decorrelation between the two subspaces, and an auxillary reconstruction module to reconstructs each pure-component spectrum along a linear interpolation path from the mixture spectrum. Training mixtures are synthesized by linearly superimposing pure spectra according to the Beer–Lambert law. Pure spectra come from two simulated sources (QM9S and Alberts et al.) and from about 3K experimental NIST spectra. MUSE is compared with XGBoost, 1D-CNN, Transformer, Wu et al., and the mixture-specific model of Alberts et al., and achieves the best F1 in all 2- and 3-component settings (Table 1).

Critical strengths
  1. A science-driven and practically important problem.

The paper targets the elucidation of individual components from mixture IR spectra. This is a real need in chemical analysis, for example in reaction monitoring, formulations, and environmental samples. Existing deep-learning IR methods are largely limited to pure compounds, and only a few works address mixtures. The problem formulation itself is a meaningful step toward making spectral ML useful in realistic settings, independent of the specific architecture.

  1. Consistent improvements over both generic and mixture-specific baselines.

MUSE achieves the best F1 in every 2- and 3-component setting across three data sources: QM9S, Alberts et al., and NIST experimental pure spectra mixed synthetically (Table 1). The gains are about 4–5 F1 points over the strongest, mixture-specific baseline (Alberts et al.), averaged over 5 runs with reported significance. The lead also holds on the 59 physically measured mixtures (Obs. 3). The improvements are largest on rare functional groups (Fig. 3d), which suggests that the model is more robust to class imbalance.

Critical weaknesses
  1. Restricted problem setting: K is given, and the output is functional groups rather than molecules.

MUSE requires the number of components K as input. It is evaluated only with 2–4 components, and it predicts a fixed set of 37 functional groups per component rather than generating the full set of (molecule, concentration) pairs end to end. In real discovery pipelines, chemists often face a crude reaction mixture with an unknown and possibly large number of components. The paper does not show whether MUSE gives chemically sensible predictions when the true mixture is much more complex than the preset K. The only test is an overestimate of K by one (K̂ = K + 1); underestimation is not evaluated. It also does not show whether the model scales to larger K (e.g., 5, 10, or more), which the authors acknowledge as a limitation. Moreover, recovering a molecule from its functional-group set is itself an underdetermined problem, and the paper notes this for isomer pairs in Appendix F. The current output therefore leaves the hardest part of component identification unresolved.

  1. The claimed unmixing ability is not properly evaluated.

The paper claims that the CFM branch performs spectral unmixing (Contribution 3 in Sec. 1: “We integrate CFM into generative reconstruction, where its continuous probability paths mirror signal unmixing and root latent disentanglement in the spectral generation space.”). However, reconstruction quality is reported in a single sentence (Obs. 7: cosine 0.833 / 0.673, MAE 0.011 / 0.014). No baselines are given, such as using the mixture spectrum itself as the estimate, classical unmixing (MCR-ALS / NMF with K given), or neural unmixing methods. Without such comparisons, the unmixing claim is not supported.

  1. Reliance on the concentration prior is not benchmarked against baselines.

The concentration prior carries substantial information. MUSE uses it both as an input and in the Hungarian matching cost (Eq. 11). In real-world settings, per-component concentrations are usually unknown. The settings without concentration (K + formula, K only; Table 2) report only MUSE’s own results, with no baselines. It is therefore unclear whether MUSE’s advantage holds when the concentration prior is unavailable, which is the more realistic case.

I will change my assessment if these weaknesses are answered properly, especially 1 and 3.

Questions for authors

Please also refer to weaknesses.

  1. How does MUSE behave when the true number of components exceeds the preset K, for example with 2 slots on a 4-component mixture? Can the model infer K itself, for example by adding a “no-object” class to the slots as in DETR-style set prediction?
  2. In the K̂ = K + 1 experiment, unmatched slots are ignored when computing F1 (Appendix J). What is the result if the extra slot’s predictions are counted as false positives?
  3. Could the authors compare the CFM reconstructions against (a) using the mixture spectrum itself as the estimate of each pure spectrum, and (b) classical blind unmixing with K given (e.g., MCR-ALS or NMF), and break the reconstruction quality down by component concentration?
  4. Could the authors report baseline results in the settings without concentration (K + formula and K only; Table 2), and an ablation that removes the concentration term from the Hungarian matching cost (Eq. 11)?
  5. Fig. 3(b–c) reports components with concentration < 0.1 (2-component) and < 0.05 (3-component), but Table 1 states that concentrations are sampled from [0.2, 0.8]. How were these low-concentration test samples generated?
Reviewer_7xoCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MUSE for predicting each component’s functional groups from mixed infrared spectra, combining slot-based set prediction, latent separation, and flow-based reconstruction. It addresses a relevant problem and reports encouraging gains. However, the main novelty beyond combining established techniques remains unclear, and stronger justification of the matching design and more controlled ablations are needed to establish what drives the improvements.

Critical strengths

The paper addresses a relevant chemical-analysis problem, where overlapping spectral signals make individual components difficult to characterize. results report consistent improvements across three spectral sources, supported by repeated runs. The measured-mixture experiments provide preliminary evidence of practical value.

Critical weaknesses
  1. The method combines established techniques, including slot queries, Hungarian matching, decorrelation, and flow matching. The authors should clarify the main technical contribution beyond combining these components.
  2. Although the full model performs well, the slot-only variant does not outperform the strongest baseline. More controlled ablations are needed to distinguish the contributions of individual components and determine whether their interaction explains the improvement.
  3. Table 3 shows that direct CNN reconstruction already recovers much of the improvement. The authors should compare the two reconstruction approaches under matched capacity and training budgets, reporting reconstruction quality, classification performance, and computational cost. Separate ablations of the curriculum and class weighting would also help identify what drives the gains on low-concentration components and rare groups.
  4. Hungarian matching is reasonable for unordered component predictions, but the paper does not establish the benefit of its concentration-aware cost (Eq. 11). The authors should clarify this choice, report the matching weight, and compare classification-only and concentration-assisted matching. Comparisons with greedy or soft Sinkhorn matching could further show whether the assignment strategy materially affects performance.
Questions for authors

See the weaknesses

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

MUSE predicts component-level functional groups from mixture infrared spectra using slot-based set prediction, structural–concentration subspace separation, and conditional flow matching for auxiliary spectral reconstruction. The method achieves improvements over the selected baselines, but its practical relevance remains insufficiently established because the evaluation relies heavily on synthetic mixtures and assumes prior information whose availability in realistic workflows is unclear.

Critical strengths

The paper addresses an interesting challenge: assigning functional groups to individual components rather than identifying their presence in the mixture as a whole. MUSE outperforms the selected baselines across multiple spectral sources and maintains advantages under noise, baseline drift, and low component concentrations. The ablations include useful controls, while representation probes provide empirical support for the proposed subspace separation.

Critical weaknesses

The practical task formulation is insufficiently justified. The paper does not explain how an accurate component count, molecular formulas, and concentrations could be obtained in realistic workflows without substantial prior analysis. Furthermore, the evaluation with an overestimated component count excludes unmatched predictions from F1, leaving spurious components unpenalized and providing insufficient evidence of robustness to unknown component counts.

The evaluation does not adequately represent realistic analytical workflows. It relies heavily on simulated pure spectra, and most mixtures are artificially constructed through linear superposition, including those based on experimental NIST spectra. The two small datasets of measured mixtures, totaling 59 spectra, provide preliminary validation but do not establish reliability in production workflows or more complex chemical systems.

The methodological novelty is limited. The framework primarily combines established set prediction, covariance regularization, and flow matching, while the slot-only variant does not outperform the strongest baseline, weakening the attribution of the gains to the proposed set-prediction formulation.

Several implementation details require clarification, particularly how concentration priors are used in forward prediction and whether the concentration noise experiment perturbs model inputs or only the Hungarian matching cost. The manuscript also does not fully specify how endpoint predictions are converted into the vector field used for flow integration, limiting reproducibility of the sampling procedure.

The mechanistic claims exceed the available evidence. Cross-covariance regularization and lightweight probes support partial representation separation but do not establish independent, physically identifiable factors. The paper also does not verify whether the reconstructed pure spectra reproduce the input mixture when recombined at the specified concentrations.

Questions for authors

What concrete application provides reliable component counts, formulas, and concentrations while still requiring component-level functional-group prediction, and what additional value does MUSE provide over separation-based analysis? How does performance change when unmatched predictions are counted as false positives, and can the model infer component presence without knowing the true count? Were the measured-mixture datasets strictly independent of training and model selection, including their constituent molecules, and what uncertainty accompanies the reported results? How exactly are concentration priors incorporated during training and inference, and does the concentration-noise experiment perturb predictions, matching, or both? Can the authors provide the complete sampling procedure and quantify the consistency between the observed mixture and the concentration-weighted reconstructed spectra?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.