acceptodds
Under review as a conference paper at ICLR 2027

SpecReact: Role-Aware Spectrum-to-Spectrum Reaction Modeling with Unbalanced Optimal Transport

Abstract

Chemical reactions are fundamental to functional molecule synthesis and play a central role in drug discovery, materials design, and chemical manufacturing. Spectra provide experimentally accessible and machine-readable observations of chemical transformations, motivating direct reaction modeling in the spectral domain without intermediate molecular-structure reconstruction. However, product spectrum prediction remains challenging because reactions contain an unordered and variable number of inputs with distinct chemical roles, while reactant-to-product spectra exhibit non-conservative changes such as peak shifts, intensity variations, and peak emergence or disappearance. We propose SpecReact, a spectrum-to-spectrum framework for predicting product spectra from reactant and reagent spectra. SpecReact employs a role-aware set encoder for variable-sized inputs and introduces an unbalanced optimal transport (UOT)-based spectral transformation module to model spectral correspondences and provide structured supervision for reaction-induced changes. We also construct a new spectrum-based reaction dataset from USPTO reaction records paired with simulated IR spectra. Experiments on simulated and experimentally measured spectra show that SpecReact consistently outperforms representative reaction baselines and remains robust to spectral perturbations. On experimental spectra dataset, SpecReact achieves a CosSim of and a peak F1 of , demonstrating its effectiveness in practical spectroscopic scenarios. Our code can be found at https://anonymous.4open.science/r/SpecReact-A30C/

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Chemical Reaction Prediction, Infrared Spectroscopy, Spectrum-to-Spectrum Modeling, Unbalanced Optimal Transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qN6FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a spectrum-to-spectrum framework for directly predicting product IR from reactant and reagent spectra without intermediate molecular-structure reconstruction. Experiments on simulated and experimentally measured spectra show that the proposed method consistently outperforms reaction baselines across diverse reactant settings and remains robust to spectral perturbations.

Critical strengths
  • The motivation is straightforward. The method appears to be effective.
Critical weaknesses
  • The generalization capability requires further validation. For example, by directly transferring from simulated spectra to real-world scenarios.
Questions for authors
  • The baseline models rely on methods from 2019 to 2022. It is difficult to support the method effectiveness.

  • Please explain the core differences between simulated and real spectra. I wonder if the method can extract the ability from the simulated spectra that are applicable to real data.

  • I think it's a good way to test the model generalizable ability by training on simulated spectral data and directly transferring to real scene data. But, note that there is no data leakage issue between the two settings.

5 comments

W1. & Q3. Zero-shot sim-to-real transfer and data leakage issue.

We thank the reviewer for this constructive suggestion. To directly evaluate the zero-shot generalization capability of our model trained solely on simulated spectra, we test the simulated-trained checkpoint directly on the experimental NIST test split without any fine-tuning (Zero-Shot), and compare it with the fine-tuned setting (SpecReact):

MethodCosSimF1
Zero-Shot0.481(0.019)0.433(0.021)
SpecReact0.561(0.013)0.517(0.015)
  • Zero-Shot Transfer Analysis: SpecReact boosts the performance to 0.561 in CosSim (+16.6%) and 0.517 in Peak F1 (+19.4%), demonstrating that models perform worse when directly transferred compared to when they are fine-tuned before testing, highlighting the importance of fine-tuning with experimental data.

  • No Data Leakage Between Simulated and Experimental Settings: We strictly enforced cross-setting product and reaction disjointness during dataset construction. Specifically, all target product molecules and reaction records appearing in the experimental val and test splits were filtered out from the simulated train and val splits prior to pretraining and reaction training. Thus, there is no data leakage.

Q1. Comparison with more recent baseline methods (2023–2025).

We add three recent reaction prediction architectures adapted to the spectral domain using the same pretrained spectrum encoder and decoder: Deductive-Spec [1], RXNG-Spec [2] and ReactionT5-Spec [3]. The results across reactant topologies and reactant counts are reported below:

Simulated Spectra

MethodAROARO+LCLCLC+ALRALR
CosSimF1CosSimF1CosSimF1CosSimF1CosSimF1
RXNG-Spec0.641(0.013)0.611(0.009)0.633(0.009)0.607(0.010)0.608(0.011)0.579(0.011)0.674(0.010)0.594(0.010)0.647(0.010)0.578(0.011)
Deductive-Spec0.664(0.012)0.621(0.010)0.643(0.009)0.617(0.013)0.626(0.012)0.572(0.013)0.675(0.010)0.588(0.009)0.658(0.014)0.577(0.012)
ReactionT5-Spec0.671(0.012)0.625(0.012)0.649(0.008)0.625(0.011)0.631(0.009)0.585(0.011)0.688(0.015)0.596(0.009)0.668(0.009)0.590(0.013)
SpecReact0.691(0.008)0.642(0.009)0.667(0.007)0.641(0.009)0.649(0.009)0.598(0.006)0.713(0.011)0.611(0.007)0.684(0.008)0.593(0.008)

Experimental Spectra

MethodAROARO+LCLCLC+ALRALR
CosSimF1CosSimF1CosSimF1CosSimF1CosSimF1
RXNG-Spec0.512(0.013)0.467(0.018)0.489(0.014)0.460(0.019)0.465(0.015)0.416(0.016)0.517(0.020)0.433(0.016)0.501(0.019)0.404(0.017)
Deductive-Spec0.514(0.018)0.473(0.016)0.490(0.015)0.467(0.016)0.466(0.014)0.414(0.018)0.527(0.020)0.435(0.017)0.500(0.017)0.417(0.018)
ReactionT5-Spec0.528(0.017)0.491(0.014)0.511(0.018)0.487(0.014)0.481(0.014)0.431(0.019)0.543(0.016)0.457(0.015)0.521(0.019)0.439(0.021)
SpecReact0.567(0.013)0.526(0.017)0.539(0.009)0.519(0.014)0.514(0.010)0.462(0.015)0.579(0.018)0.489(0.015)0.551(0.017)0.454(0.013)

Performance Across Reactant Counts on Simulated Spectra

Method1 Reactant2 Reactants3 Reactants≥4 Reactants
CosSimF1CosSimF1CosSimF1CosSimF1
RXNG-Spec0.652(0.007)0.621(0.008)0.655(0.007)0.622(0.010)0.578(0.018)0.583(0.013)0.540(0.020)0.559(0.017)
Deductive-Spec0.663(0.007)0.625(0.010)0.666(0.008)0.623(0.009)0.587(0.010)0.571(0.013)0.573(0.016)0.569(0.020)
ReactionT5-Spec0.670(0.008)0.628(0.007)0.671(0.009)0.627(0.008)0.598(0.017)0.588(0.011)0.581(0.022)0.561(0.019)
SpecReact0.683(0.005)0.640(0.006)0.687(0.004)0.638(0.007)0.618(0.010)0.616(0.012)0.590(0.015)0.594(0.016)

While these newer baselines improve over earlier 2019–2022 baselines, SpecReact still consistently outperforms the strongest baselineReactionT5-Spec across all reactant categories and reactant counts on both simulated and experimental spectra.

Q2. Core differences between simulated and experimental spectra, and transferability.

Simulated IR spectra reflect idealized isolated-molecule harmonic vibrations with clean baselines, whereas experimental NIST spectra contain anharmonic shifts, intermolecular interactions, instrumental noise, baseline drift, and heterogeneous measurement conditions[4,5]. Despite these discrepancies and the much smaller experimental sample size, the underlying functional group vibrational characteristics and reaction-induced spectral transformation rules remain physically consistent across both domains.

As shown in the table below, SpecReact achieves 0.680 CosSim and 0.637 Peak F1 on simulated spectra, and maintains 0.561 CosSim and 0.517 Peak F1 on experimental spectra, resulting in a moderate performance gap of only 0.119 in CosSim and 0.120 in Peak F1 (smaller than the 0.129–0.170 CosSim and 0.134–0.174 Peak F1 drops observed across all seven baselines). Although experimental variations and limited real-world data inevitably increase the difficulty of fine-grained peak alignment and cause a slight performance drop, this modest gap supports that SpecReact effectively extracts transferable molecular and reaction features applicable to real-world spectra.

SettingCosSimPeak F1
Simulated Spectra0.680(0.005)0.637(0.007)
Experimental Spectra0.561(0.013)0.517(0.015)
Performance Gap (Δ\Delta)-0.119-0.120

Thank you again for your careful review and constructive feedback. We hope our responses have clarified the key concerns, and we would be happy to address any further questions. If you find our explanations satisfactory, we would sincerely appreciate your consideration of a revised rating.


Reference

[1] Jin, T., Zhao, Q., Schofield, A. B., and Savoie, B. M. Deductive Machine Learning Models for Product Identification.

[2] Xu, L.-C., et al. A Unified Pre-trained Deep Learning Framework for Cross-task Reaction Performance Prediction and Synthesis Planning.

[3] Sagawa, T., and Kojima, R. ReactionT5: A Pre-trained Transformer Model for Accurate Chemical Reaction Prediction with Limited Data.

[4] Merrick, J. P., Moran, D., and Radom, L. An Evaluation of Harmonic Vibrational Frequency Scale Factors.

[5] Stein, S. E. Evaluated Infrared Reference Spectra. NIST Chemistry WebBook. https://webbook.nist.gov/chemistry/coblentz/

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer qN6F,

Thank you again for your careful review and for raising important questions regarding the sim-to-real generalization capability, recent baselines, domain discrepancies, and data leakage prevention.

In our response, we added a direct zero-shot sim-to-real transfer evaluation on the experimental NIST test set, comparisons against three recent reaction modeling baselines (Deductive-Spec, RXNG-Spec, and ReactionT5-Spec), a detailed analysis of the physical differences and transferability between simulated and experimental spectra, and clarification of our strict cross-domain product-disjoint filtering setting. Together, these additions demonstrate that SpecReact learns transferable spectral transformation representations without cross-setting data leakage and consistently outperforms recent baselines.

If time permits, we would be grateful if you could review these additional results and clarifications and let us know whether they address your concerns. Any remaining questions would be very helpful for us to clarify before the discussion period ends.

Thank you once again for your time and valuable feedback.

Best regards,

The authors

Reviewer_WoKhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

The paper studies direct prediction of product IR spectra from sets of individually measured or simulated reactant and reagent spectra, avoiding a cascade of spectrum-to-structure, structure-based reaction prediction, and structure-to-spectrum conversion. SpecReact pretrains a spectral encoder with masked reconstruction and peak-map prediction, uses role-aware set encoding to distinguish reactants from reagents, and refines product representations through parameter-shared iterations. During training, unbalanced optimal transport constructs four-channel spectral-change targets from reactant and ground-truth product spectra, providing auxiliary supervision for transported, disappearing, emerging, and displaced spectral components.

Critical strengths
  1. The task characteristics motivate identifiable architectural choices. Variable numbers of components, exchangeability within roles, and different functions across roles naturally motivate a role-aware set encoder.
  2. Peak displacement and intensity changes reasonably motivate an unbalanced-transport auxiliary objective.
  3. Tables 1/2 report standard deviations across five runs, and the authors attempt evaluation using experimentally measured molecular spectra.
Critical weaknesses
  1. The distinct contributions of UOT and iterative modeling are not yet separated from generic auxiliary supervision, representation refinement, and baseline-adaptation effects. Here UOT generates training targets; inference does not solve a transport problem against the true product. Improved prediction therefore does not itself demonstrate chemically meaningful peak correspondence or establish that unbalanced transport is the necessary source of improvement.
  2. The construction of reactant mass a, transport-cost scaling, entropy and marginal penalties, and displacement-channel normalization need specification.
  3. Products may preserve much of a substrate’s structure, so copying a major reactant spectrum or combining input spectra could already achieve useful global CosSim. The absence of these simple references and change-region-specific evaluation makes it difficult to separate learning a reaction transformation from preserving the dominant spectral profile.
Questions for authors
  1. Which parameters, losses, and feedback paths are removed in “w/o UOT”? When the ChangePredictor, iterative architecture, and training budget are retained, what happens if only the UOT loss is removed or replaced by simple difference-spectrum supervision?
  2. How are independently normalized reactant spectra aggregated into a? Please specify C, ε, τ_r, τ_p, solver accuracy, and scaling of the four change channels. As reactant count increases, do these choices alter the mass scale or the channel dominating the loss?
  3. Relative to copying a major reactant spectrum, learned input-spectrum averaging, and a budget-matched direct set model, how much improvement remains in genuinely changed spectral regions? Are reactant/reagent roles available before the experiment, or derived from product mapping/post hoc annotation?
  4. How many reactions and unique components remain after NIST matching, and what are the split and category counts? Are experimental spectra used for training, fine-tuning, or model selection, and how are molecule/product overlaps across splits handled?
8 comments

W1. & Q1. Details in UOT ablation.

We thank the reviewer for this insightful comment. Below we clarify the w/o UOT configuration and provide an ablation holding the iterative architecture and parameter budget strictly identical.

  1. Exact specification of w/o UOT (Figure 4(c)): In w/o UOT, the iterative reasoning module (ChangePredictor, ChangeMapHead, and feedback path C(t+1)=[C(0);E(t)]\mathcal{C}^{(t+1)} = [\mathcal{C}^{(0)}; E^{(t)}]) is removed. The model performs a single-pass mapping $ \mathcal{C}_{\text{rxn}}^{(0)} \to Z_{\text{rxn}} \to A_p \to \tilde{A}_p \to \hat{y} $, trained solely with Lspec\mathcal{L}_{\text{spec}}.

  2. Retaining the iterative architecture while removing or replacing LUOT\mathcal{L}_{\text{UOT}}: When the 3-iteration architecture (T=3T=3), ChangePredictor, feedback path, and training budget are fully retained:

    • Removing only LUOT\mathcal{L}_{\text{UOT}} corresponds to λUOT=0\lambda_{\text{UOT}} = 0 in Figure 4(d), where ChangePredictor still generates refinement tokens E(t)E^{(t)} supervised end-to-end by Lspec\mathcal{L}_{\text{spec}}.
    • Simple Difference-Spectrum Supervision (Ldiff\mathcal{L}_{\text{diff}}): We retain the full iterative architecture and ChangePredictor, but replace UOT targets with point-wise difference targets between the pooled reactant envelope aa and product spectrum bb: persistence pdiff=min⁡(a,b)p^{\text{diff}} = \min(a, b), disappearance ddiff=[a−b]+d^{\text{diff}} = [a - b]_+, and emergence ndiff=[b−a]+n^{\text{diff}} = [b - a]_+.
ModelIterative Arch. (T=3T=3)ChangePredictor & FeedbackAuxiliary TargetCosSimF1
1. w/o UOT×\times×\timesNone0.6310.576
2. w/o LUOT\mathcal{L}_{\text{UOT}} (λUOT=0\lambda_{\text{UOT}}=0)✓\checkmark✓\checkmarkNone0.6330.579
3. w. Diff-Spectrum Supervision✓\checkmark✓\checkmarkPoint-wise (a,b)(a, b) diff0.6470.595
4. w. Balanced OT✓\checkmark✓\checkmarkBalanced OT0.6560.608
5. w. UOT (SpecReact)✓\checkmark✓\checkmarkUOT0.6800.637
  • As shown above, retaining the iterative loop without auxiliary supervision (Row 2 vs. Row 1) brings negligible gains. w. Diff-spectrum supervision (Row 3) improves performance through supervision, but falls short of OT-based variants. W. Balanced OT (Row 4) captures shifts but enforces strict mass conservation, introducing false transport when bands genuinely vanish or emerge, and w. UOT (Row 5) jointly models non-conservative intensity changes and spectral shifts, yielding the primary gain.

W2. & Q2. Specification of UOT Target Construction and Scaling.

We appreciate the reviewer's thoroughness and will add the full mathematical specification to Section 2.3.3 and Appendix B.

  1. Construction of Source aa and Target bb: Each input spectrum is normalized by its maximum intensity to [0,1][0, 1] on the L=1,600L=1,600 grid (400–4000 cm−1400\text{--}4000\text{ cm}^{-1}). We take the pointwise maximum envelope over valid reactants, el=max⁡m∈{1..Nr}xr,m,le_l = \max_{m \in \{1..N_r\}} x_{r,m,l}, and apply non-overlapping 16-point max-pooling to obtain a∈[0,1]100a \in [0, 1]^{100}. The same pooling on the product spectrum yields b∈[0,1]100b \in [0, 1]^{100}. Renormalizing ∑iai=1\sum_i a_i = 1 would artificially dilute peak intensities in multi-band molecules or multi-reactant mixtures, whereas pointwise max-pooling preserves the [0,1][0, 1] peak prominence scale and represents the upper-bound spectral availability at each frequency bin.

  2. UOT Hyperparameters (C,ε,τr,τpC, \varepsilon, \tau_r, \tau_p) and Solver Accuracy: For bin i∈{0,…,N−1}i \in \{0, \dots, N-1\} (N=100N=100), we set normalized coordinates zi=i/(N−1)∈[0,1]z_i = i/(N-1) \in [0, 1] and cost Cij=∣zi−zj∣∈[0,1]C_{ij} = |z_i - z_j| \in [0, 1]. Here, ε\varepsilon controls the entropic smoothing of the transport plan across neighboring wavenumber bins, while τr\tau_r and τp\tau_p control the KL penalties for reactant peaks disappearance and product peaks emergence, respectively (where ε→0\varepsilon \to 0 recovers unregularized OT and τr,τp=0\tau_r, \tau_p = 0 removes marginal constraints, causing T→0T \to 0). We set ε=0.05\varepsilon = 0.05 and τr=τp=0.20\tau_r = \tau_p = 0.20. The sensitivity across [0,0.25][0, 0.25] is shown below (CosSim / F1)), varying one hyperparameter while fixing the others at default.

    HyperparameterMetric00.050.100.150.200.25
    ε\varepsilonCosSim0.6580.6800.6760.6710.6650.659
    Peak F10.6110.6370.6320.6260.6190.612
    τr\tau_rCosSim0.6240.6590.6700.6750.6800.677
    Peak F10.5710.6130.6240.6310.6370.633
    τp\tau_pCosSim0.6230.6550.6680.6760.6800.678
    Peak F10.5690.6150.6250.6330.6370.635

The transport plan TT is solved via 30 alternating unbalanced Sinkhorn iterations initialized with unit scaling vectors. Because Cij∈[0,1]C_{ij} \in [0, 1] and ε=0.05\varepsilon = 0.05, the contraction factor guarantees rapid exponential convergence, where the marginal ℓ∞\ell_\infty drift between iteration 25 and 30 is already <10−4< 10^{-4}, making 30 fixed iterations a fast and accurate choice.

  1. Scaling of the Four Change Channels: With ri=∑jTijr_i = \sum_j T_{ij} and qj=∑iTijq_j = \sum_i T_{ij}, the implemented targets are:

    • Persistence: pj=qjp_j = q_j;
    • Disappearance: di=[ai−ri]+d_i = [a_i - r_i]_+;
    • Emergence: nj=[bj−qj]+n_j = [b_j - q_j]_+;
    • Displacement: sj=clip(∑iTij(zj−zi)qj+ϵfloor,−1,1)s_j = \text{clip}\big(\frac{\sum_i T_{ij}(z_j - z_i)}{q_j + \epsilon_{\text{floor}}}, -1, 1\big).

    ChangeMapHead applies Sigmoid to (p,d,n)(p, d, n) and Tanh to ss. Although pj=qjp_j = q_j is not algebraically bounded by 11, the penalty τpDKL(q∥b)\tau_p D_{\text{KL}}(q \parallel b) tightly regularizes qjq_j toward bj≤1b_j \le 1, so Sigmoid aligns well with the effective target range without clipping distortion. LUOT\mathcal{L}_{\text{UOT}} averages Smooth L1 over (p,d,n)(p, d, n) and then plus ss.

  2. Effect of Increasing Reactant Count NrN_r: Because aa uses the pointwise maximum rather than summation, increasing NrN_r never exceeds the upper bound (1.01.0) of any individual bin; it only fills bins where existing reactants had weak signals, moderately increasing ∑iai\sum_i a_i. Untransported excess bands populate di=[ai−ri]+∈[0,1]d_i = [a_i - r_i]_+ \in [0, 1]. Thus, larger NrN_r mildly increases the active support of dd without altering the intensity scale or dominating the equally weighted channel loss.

W3. & Q3. Comparison Against Copy/Averaging References, Change-Region Evaluation, and Role Availability.

  1. Performance in Genuinely Changed Regions and Across Similarity Strata: To verify that SpecReact learns true chemical transformations rather than copying dominant substrates, we compare against three simple/budget-matched references on genuinely changed spectral regions Mchg=l∣∣yl−al∣>0.10\mathcal{M}_{\text{chg}} = \\{l \mid |y_l - a_l| > 0.10\\}:

    MethodRegional CosSimRegional F1
    Copying major reactant spectra0.3120.246
    Learned Input Averaging0.385(0.021)0.324(0.022)
    Budget-Matched Direct Set0.528(0.014)0.482(0.014)
    SpecReact0.614(0.011)0.573(0.013)

    As shown in the table above, while simply copying or averaging reactant spectra fails severely in genuinely changed regions, SpecReact reaches 0.614 CosSim and 0.573 F1, substantially outperforming the budget-matched direct set baseline (0.528 / 0.482).

    We did further study across Low, Medium, and High reactant-product similarity subsets (CosSim). The table below shows the high prediction similarity is genuinely driven by SpecReact modeling reaction transformations rather than inheriting pre-existing substrate similarity.

    MethodLow (<=0.250)Medium (>0.250, <=0.400)High (>0.400)
    Input–Target0.2140.3460.504
    ReactionT5-Spec0.612(0.010)0.658(0.012)0.684(0.011)
    SpecReact0.640(0.007)0.683(0.006)0.719(0.006)
  2. Availability of Reactant/Reagent Roles: Our dataset preserves the standard reactant/reagent separation from USPTO records as a prior experimental recipe specification, and no product structure or post-hoc atom mapping is ever used at inference time. Moreover, among multiple reactants in Xr\mathcal{X}_r, the model receives no annotation of which molecule is the major substrate. Instead, the Role-Aware Set Encoder and Eq. (11) learn to weight dominant structural contributors directly from the input spectra.

Q4. Details about experimental data and fine-tuning.

  1. NIST Dataset Statistics: After matching USPTO reaction components against experimental IR spectra from the NIST Chemistry WebBook and retaining complete reactions where all reactants, reagents, and products have valid experimental spectra, the exact counts across splits and categories are:
StatisticTrainValTestTotal
Reactions3,5411672723,980
Molecules2,4132853232,567
  1. Training and Sim-to-Real Fine-Tuning Protocol: Models are first trained on the simulated train split, with the best checkpoint selected on the simulated val split. For experimental spectra, we initialize from the best simulated checkpoint, fine-tune on the experimental train split, select the best checkpoint via the experimental val split, and report final results on the held-out experimental test split. All baselines follow the exact same protocol.

  2. Molecule/Product Overlap Handling: Reactions are partitioned with no product molecule appears in more than one split (train, val, or test) in either dataset. While common reagents/solvents and standard building-block reactants naturally recur across splits which reflects realistic synthesis, target products and complete reaction tuples are strictly non-overlapping.

We greatly appreciate the time and thoughtful effort you have devoted to reviewing our work. Your insightful comments have helped us strengthen both the clarity and quality of our manuscript. We remain open to any further questions or discussions and would be glad to address any remaining concerns. Should you find our clarifications satisfactory, we sincerely hope you might consider updating your rating accordingly.

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

The paper solves for a very important yet under-researched task: predicting spectra of the product based on spectra of substrates. The paper constructs a new benchmark: comparing some models derived from literature on reaction prediction, to the proposed SpecReact, on simulated and real IR. The proposed architecture and training recipe clearly outperforms other models. The novelty and comparison to the broader literature seem to be the weakest points of the paper. A substantial revision of the comparison to both specialized and frontier LLMs could impact my rating.

Critical strengths
  1. The modeling choices are strong (pretraining, reasoning, UOT)
  2. The task is of high practical importance, while somewhat under-researched.
Critical weaknesses
  1. No comparison to an LLM. How would GPT-Astra or similar model perform at this task?
  2. Limited novelty. Related models for other spectra (e.g. for NMR) have been proposed. The paper instead constructs its own benchmarking approach.
  3. The approach is complex and it would be helpful to see ablation of other components like reasoning or pretraining.
Questions for authors
  1. How would the model work in case of measuring IR spectra of reaction mixture after LC? Or that is not done normally?
  2. I would imagine that PDA spectra or NMR of substrates is far from that of product, and as such is optimal transport the right formulation broadly? Are IR spectra of products usually similar?
  3. Could you compare to models for other spectra like NMR (e.g. “Predicting Discrete Structural Transformations in Small Molecules from Tandem Mass Spectrometry”)?
  4. How could we apply this model downstream? Perhaps some illustration of that could be helpful?
4 comments

W1. No comparison to an LLM. Performance on LLMs.

Thank you for this valuable suggestion. We additionally evaluate GPT-6 Astra and DeepSeek-V4.1-Flash under zero-shot and 3-shot settings. Both LLMs receive the exact same normalized reactant and reagent IR spectra as SpecReact and are instructed to directly predict the product IR spectrum without any molecular structure or SMILES information. In addition we use 3 training-set demonstrations for the 3-shot setting.

ModelSettingCosSimF1
GPT-6 AstraZero-shot0.3220.358
GPT-6 Astra3-shot0.3560.386
DeepSeek-V4.1-FlashZero-shot0.3040.337
DeepSeek-V4.1-Flash3-shot0.3320.361
SpecReact0.6800.637

As shown above, SpecReact still achieves the best performance on both metrics, highlighting the necessity of dedicated spectral representation learning and structured transformation modeling for reconstructing fine-grained product IR spectra.

W2. & Q3. Novelty and comparison with related models for other spectra.

  1. Practical Motivation & Core Synergy of SpecReact:

    • In practical chemical synthesis and autonomous laboratories, spectra serve as direct, machine-readable experimental observations for assessing reaction outcomes. To model chemical reactions, conventional approaches typically adopt a structure-to-structure paradigm operating on resolved molecular graphs or SMILES [5,6]. However, in real-world experimental workflows, intermediate spectrum-to-structure elucidation that easily introduces structural errors and error propagation [11]. Therefore, we perform direct reaction monitoring and outcome prediction in the spectral domain, establishing a spectrum-to-spectrum paradigm that aligns much more closely with practical laboratory scenarios [7–9, 12].
    • To achieve this, SpecReact tightly couples a Role-Aware Set Encoder with UOT-Guided Iterative Reaction Reasoning: the Role-Aware Set Encoder disentangles unordered reactants from reagents to construct both the initial reaction context Crxn(0)\mathcal{C}_{\text{rxn}}^{(0)} and an attention-aggregated reactant reference ArA_r; the iterative reasoning module then contrasts the evolving product representation Ap(t)A_p^{(t)} against ArA_r under UOT supervision to predict structured spectral changes E(t)E^{(t)}, which are fed back into Crxn(t+1)\mathcal{C}_{\text{rxn}}^{(t+1)} to progressively refine the product spectrum.
  2. Fundamental Differences from Existing Spectral & Reaction Models:

    • Comparison with STEP [1]: The suggested work [1] predicts discrete molecular graph-edit distances between a pair of already observed molecules from their respective MS/MS spectra. Its objective is pairwise structural similarity/edit assessment rather than generating an unknown product spectrum from multi-component reactant and reagent spectra. Thus, its input availability, prediction target, and task formulation are fundamentally different from SpecReact. We will discuss it in Related Work.
    • Comparison with NMR & Structure-Based Reaction Models: Following your advice, we also surveyed recent NMR foundation models such as NMRMind [2] and PRISM [3], which focus on single molecule structure elucidation or bidirectional spectrum-structure translation rather than multi-component reaction modeling. Meanwhile, recent reaction models such as Deductive Models [4], ReactionT5 [5], and RXNGraphormer [6] still rely on structure-to-structure translation, which suffers from the intermediate structure-resolution bottleneck in spectroscopic workflows as we mentioned above. To strengthen our empirical evaluation, we adapted [4–6] into spectral baselines denoted asDeductive-Spec, ReactionT5-Spec and RXNG-Spec, and SpecReact consistently outperforms all of them.

Simulated Spectra

MethodAROARO+LCLCLC+ALRALR
CosSimF1CosSimF1CosSimF1CosSimF1CosSimF1
RXNG-Spec0.641(0.013)0.611(0.009)0.633(0.009)0.607(0.010)0.608(0.011)0.579(0.011)0.674(0.010)0.594(0.010)0.647(0.010)0.578(0.011)
Deductive-Spec0.664(0.012)0.621(0.010)0.643(0.009)0.617(0.013)0.626(0.012)0.572(0.013)0.675(0.010)0.588(0.009)0.658(0.014)0.577(0.012)
ReactionT5-Spec0.671(0.012)0.625(0.012)0.649(0.008)0.625(0.011)0.631(0.009)0.585(0.011)0.688(0.015)0.596(0.009)0.668(0.009)0.590(0.013)
SpecReact0.691(0.008)0.642(0.009)0.667(0.007)0.641(0.009)0.649(0.009)0.598(0.006)0.713(0.011)0.611(0.007)0.684(0.008)0.593(0.008)

Experimental Spectra

MethodAROARO+LCLCLC+ALRALR
CosSimF1CosSimF1CosSimF1CosSimF1CosSimF1
RXNG-Spec0.512(0.013)0.467(0.018)0.489(0.014)0.460(0.019)0.465(0.015)0.416(0.016)0.517(0.020)0.433(0.016)0.501(0.019)0.404(0.017)
Deductive-Spec0.514(0.018)0.473(0.016)0.490(0.015)0.467(0.016)0.466(0.014)0.414(0.018)0.527(0.020)0.435(0.017)0.500(0.017)0.417(0.018)
ReactionT5-Spec0.528(0.017)0.491(0.014)0.511(0.018)0.487(0.014)0.481(0.014)0.431(0.019)0.543(0.016)0.457(0.015)0.521(0.019)0.439(0.021)
SpecReact0.567(0.013)0.526(0.017)0.539(0.009)0.519(0.014)0.514(0.010)0.462(0.015)0.579(0.018)0.489(0.015)0.551(0.017)0.454(0.013)

Performance Across Reactant Counts on Simulated Spectra

Method1 Reactant2 Reactants3 Reactants≥4 Reactants
CosSimF1CosSimF1CosSimF1CosSimF1
RXNG-Spec0.652(0.007)0.621(0.008)0.655(0.007)0.622(0.010)0.578(0.018)0.583(0.013)0.540(0.020)0.559(0.017)
Deductive-Spec0.663(0.007)0.625(0.010)0.666(0.008)0.623(0.009)0.587(0.010)0.571(0.013)0.573(0.016)0.569(0.020)
ReactionT5-Spec0.670(0.008)0.628(0.007)0.671(0.009)0.627(0.008)0.598(0.017)0.588(0.011)0.581(0.022)0.561(0.019)
SpecReact0.683(0.005)0.640(0.006)0.687(0.004)0.638(0.007)0.618(0.010)0.616(0.012)0.590(0.015)0.594(0.016)

While these newer baselines improve over earlier 2019–2022 baselines, SpecReact still consistently outperforms the strongest baselineReactionT5-Spec across all reactant categories and reactant counts on both simulated and experimental spectra.

We have not found prior models that directly predict product spectra from reactant and reagent spectra, and we would be grateful if the reviewer could point us to any additional specific references and we would be happy to discuss and compare them in the revised manuscript.

W3. Ablation of other components like reasoning or pretraining.

Thank you for the suggestion. We would like to clarify that the submitted manuscript already includes comprehensive ablations on both spectral pretraining and iterative reasoning:

  • Effect of Pretraining (Appendix G, Table 3): Pretraining the spectrum encoder improves CosSim by +0.039 and Peak F1 by +0.049, confirming the benefit of learning generic peak-aware spectral representations prior to reaction training.
SettingCosSimF1
w/o Pretraining0.6410.588
w. Pretraining (SpecReact)0.6800.637
  • Effect of Iterative Reasoning (Appendix H, Figure 5): Varying the reasoning iterations from T=1T=1 to T=5T=5 shows consistent gains up to T=3T=3 (+0.034 CosSim and +0.045 F1 over T=1T=1), after which performance saturates:
MetricT=1T=1T=2T=2T=3T=3T=4T=4T=5T=5
CosSim0.6460.6670.6800.6770.673
Peak F10.5920.6190.6370.6330.627
  • Together with the UOT ablation in Figure 4(c) (improving CosSim from 0.631 to 0.680) and the qualitative iterative refinement trajectories in Appendix H (Figures 6–7), these results thoroughly validate each core component of SpecReact.

Q1. How would the model work in case of measuring IR spectra of reaction mixture after LC?

Thank you for this practical question. SpecReact currently operates on component-resolved IR spectra rather than unseparated crude mixture spectra. In a standard LC-coupled workflow, liquid chromatography first separates the reaction mixture into individual fractions, whose resolved IR spectra can then be measured and directly matched against SpecReact's predicted product spectrum for outcome verification. Extending SpecReact to direct mixture-level spectral modeling is an active direction of our ongoing work, which we will clarify in the revised manuscript.

Q2. Applicability of UOT to different spectroscopic modalities & source of CosSim improvement.

  1. Applicability of UOT to IR and Other Modalities: We focus on IR spectroscopy because in situ FTIR is widely deployed for real-time reaction monitoring with minimal sample preparation [7–9], and reaction-induced bond changes naturally manifest as localized band persistence, disappearance, emergence, and frequency shifts. Rather than assuming global similarity between substrate and product spectra, UOT relaxes marginal constraints via KL penalties to model these non-conservative spectral changes through soft correspondence supervision. Similarly, in modalities such as NMR, UV-Vis/PDA, and MS/MS, chemical transformations are also reflected in spectral peak shifts, multiplet variations, or fragment emergence/disappearance [1, 10]. Because UOT generally captures non-conservative transport between positive distributions, our framework can be naturally extended to these modalities by adapting the domain-specific Encoder and transport cost CijC_{ij}, which we consider an important direction for future work.

  2. CosSim Gains Stem from Model Learning Rather Than Substrate Similarity: To verify that high CosSim is driven by SpecReact rather than inherent reactant-product similarity, we partition the test set by baseline Input–Target similarity (Low <=0.250, 0.250 < Medium <=0.400, High >0.400):

    MethodLow (<=0.250)Medium (>0.250, <=0.400)High (>0.400)
    Input–Target0.2140.3460.504
    ReactionT5-Spec0.612(0.010)0.658(0.012)0.684(0.011)
    SpecReact0.640(0.007)0.683(0.006)0.719(0.006)

    Even when reactants and products are highly dissimilar (Low <=0.250, Input–Target = 0.214), SpecReact achieves 0.640 CosSim (+199.1%) and consistently outperforms ReactionT5-Spec across all strata, confirming that our model actively learns reaction transformations.

Q4. Downstream task illustration and evaluation.

Thank you for this valuable suggestion. We evaluate a downstream product functional group identification task across 36 functional groups, where a spectral classifier predicts the presence of functional groups in the reaction product directly from the product IR spectrum generated by each reaction model. Below, we report the F1 scores on the 11 functional groups (frequency >0.10> 0.10) alongside Deductive-Spec and ReactionT5-Spec. The complete results across all 36 functional groups will be included in the Appendix of the revised manuscript.

MethodAlkaneAreneEtherAmineHaloalkaneAmideAlcoholEsterSulfideAlkeneCarbamate
Frequency0.9350.8010.5710.4520.4520.2920.2650.1820.1380.1200.105
Deductive-Spec0.9180.8610.7240.6780.5840.5760.5010.6020.1940.2420.654
ReactionT5-Spec0.9290.8780.7410.6940.6020.5980.5200.6230.2060.2610.679
SpecReact0.9500.8940.7650.7100.6260.6220.5410.6460.2310.2840.702

As shown in the table above, SpecReact consistently achieves the highest downstream F1 scores across all functional groups compared with Deductive-Spec and ReactionT5-Spec. These results demonstrate that SpecReact preserves chemically informative spectral features beyond global intensity reconstruction, supporting automated product characterization and downstream experimental analysis.

We sincerely thank you again for your time and effort in reviewing our manuscript. Your constructive feedback has been invaluable in helping us improve our work. We would be happy to provide further clarification or engage in additional discussion if any concerns remain. If our responses have adequately addressed your concerns, we would be grateful if you could reconsider your initial rating.

References

[1] Wang, X., et al. Predicting Discrete Structural Transformations in Small Molecules from Tandem Mass Spectrometry.

[2] Xue, X., et al. NMRMind: A Transformer-Based Model Enabling the Elucidation from Multidimensional NMR to Structures.

[3] Hoque, A., and Glorius, F. PRISM, a Bidirectional NMR Foundation Model: From Peak List to Structure and Back.

[4] Jin, T., Zhao, Q., Schofield, A. B., and Savoie, B. M. Deductive Machine Learning Models for Product Identification.

[5] Sagawa, T., and Kojima, R. ReactionT5: A Pre-trained Transformer Model for Accurate Chemical Reaction Prediction with Limited Data.

[6] Xu, L.-C., et al. A Unified Pre-trained Deep Learning Framework for Cross-task Reaction Performance Prediction and Synthesis Planning.

[7] Cherniienko, A., et al. IR-EcoSpectra: Exploring Sustainable Ex Situ and In Situ FTIR Applications for Green Chemical and Pharmaceutical Analysis.

[8] Mezzetti, A., et al. Time-resolved Infrared Absorption Spectroscopy Applied to Photoinduced Reactions: How and Why.

[9] Madani, A., and Pieber, B. In Situ Reaction Monitoring in Photocatalytic Organic Synthesis.

[10] Bazzoni, M., et al. Fast Multidimensional Flow Nuclear Magnetic Resonance at High Field for Real-Time Reaction Monitoring and Flow Synthesis.

[11] Alberts, M., Laino, T., and Vaucher, A. C. Leveraging Infrared Spectroscopy for Automated Structure Elucidation.

[12] Bayley, O., Savino, E., Slattery, A., and Noël, T. Autonomous Chemistry: Navigating Self-Driving Labs in Chemical and Material Sciences.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.