acceptodds
Under review as a conference paper at ICLR 2027

Reconstruction Is Not Transportability: Transition-State-Aware Tokenization for Reaction Pathway Flow Matching

Abstract

Mapping the pathway that carries reactants through the transition state to products is a bottleneck in mechanistic analysis and reaction screening, and the fastest learned methods now produce a complete pathway in one pass from the endpoint geometries alone. Generating it in a learned latent space separates geometric encoding from transport, but a latent trained only to reconstruct is not the one a flow transports best. We call this the representation–transport mismatch: a pathway is recovered from its latent to 0.052 Å while that latent stays hard for the flow to predict. We propose REACFLOW, a two-stage framework that shapes the pathway latent for transport rather than for reconstruction alone. Its autoencoder is pretrained with a climbing-image NEB force-response head, which keeps search-relevant motion readable after compression, and an interval-completion objective, whose extreme case is the endpoint interpolation the generator starts from. We then freeze it and train a conditional flow-matching model to carry the interpolation latent to the target latent; generation uses four velocity-field evaluations followed by the frozen decoder, with no energy or force evaluations. We test REACFLOW on two benchmarks, Transition1x and Halo8, and it attains the lowest geometry error of any method compared on both (0.215/0.175 Å TS-RMSD and 0.155/0.165 Å IRC-RMSD), improving generation by 24.6–36.7% over reconstruction-only pretraining while clean reconstruction stays flat at 0.052–0.056 Å.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Reaction Pathway Generation, Flow Matching, Transition State, Latent Representation, Molecular Geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SXSEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper proposes ReacFlow, a two-stage framework that generates complete reaction pathways from reactant and product structures. Its main idea is that accurate reconstruction alone does not mean that latent representations are easy to predict from endpoint conditions. To address this, the tokenizer uses reconstruction together with CI-NEB-inspired search-response supervision and endpoint-anchored interval completion. After the tokenizer is frozen, a latent flow model learns to turn the endpoint-linear pathway representation into the reference pathway representation. It uses flow-matching and decoded-geometry losses for training. Inference needs four velocity-network evaluations and no force calculations. Tests on Transition1x and Halo8 include geometry and energy metrics, controlled pretraining ablations, latent-error diagnostics, and a comparison with direct completion. Conditional error bounds provide theoretical support for the design.

Critical strengths
  • A key strength is that the paper recognizes that accurate autoencoder reconstruction alone does not mean that latent representations are easy to predict from endpoint conditions. The paper builds its pretraining strategy around this problem. It combines CI-NEB-inspired search-response supervision with interval completion that includes the endpoint-linear input used during generation. Together, these objectives are designed to help the tokenizer keep physically relevant information and support downstream conditional prediction. I find the clear link between the problem and the proposed training objectives especially convincing. The framework also offers efficient inference, generating complete pathways with four velocity-network evaluations and no force calculations.
Critical weaknesses
  • The Drift-React results in Table 1 do not match those in the cited paper, while the other baseline results are exactly the same. The paper states that the baseline results were taken directly from that work, but does not clearly explain this difference. This unexplained difference greatly reduces my confidence in the reported comparison. In its current form, I cannot rely on Table 1 as evidence that ReacFlow performs better than existing methods.
Questions for authors

I see value in the proposed approach, but the unclear baseline comparison raised in the first question strongly affects my current assessment. A clear explanation supported by evidence could change my recommendation. Answering the endpoint-alignment question would further increase my confidence that the method can be used in practice and strengthen the case for acceptance.

  • Please explain why the Drift-React results in Table 1 differ from those in the cited paper, while the other baseline values remain unchanged. This seems to conflict with the statement that all baseline results were taken directly from that work. Where exactly did the reported Drift-React values come from, and were any changes made to them? Please provide supporting calculations or evaluation code to show that these values are correct. Also, please clarify whether the claimed improvements still hold when the same evaluation rules are applied.
  • Please clarify the endpoint alignment assumptions. Does the method require a specific relative alignment between the reactant and product structures? Is any rigid alignment performed before building the endpoint-linear path? If the method is meant to work with independently prepared endpoint structures, could you provide a small sensitivity test that rotates or translates only the product structure? The evaluation should account for the changed endpoint coordinates. Otherwise, please state the required alignment and explain how the benchmark inputs meet this requirement.
8 comments

W / Q1. DriftReact reporting and common evaluation

“Where exactly did the reported Drift-React values come from, and were any changes made to them?”

“Also, please clarify whether the claimed improvements still hold when the same evaluation rules are applied.”

Thank you for pointing out the discrepancy. DriftReact's public benchmark and code provided the basis for our comparison. Our statement that all baseline values were copied directly was inaccurate: we applied an additional normalization conversion to its geometry means and standard deviations when preparing Table 1.

For fixed aligned coordinates, per-atom and per-coordinate RMSD satisfy

dN=3,d3N.d_N=\sqrt{3}\\,d_{3N}.

When preparing Table 1, we applied a factor of 3\sqrt{3} to the means and standard deviations of all three DriftReact geometry metrics, incorrectly treating them all as per-coordinate (3N3N) statistics. Inspection of the pinned evaluation implementation confirmed that this conversion was appropriate for the historical TS-RMSD calculation, whereas IRC-RMSD and Fréchet were already normalized per atom (NN). We therefore restore the original IRC-RMSD and Fréchet statistics while retaining the converted TS-RMSD. We acknowledge that the original table caption did not clearly disclose this conversion. The remaining baseline geometry entries and all baseline energy statistics were copied without this conversion.

Static inspection of repository commit 67eb280 traces the TS evaluation call from rmsd_coords, through pymatgen_rmsd and rmsd_core, to KabschMatcher.fit. The locked pymatgen 2025.10.7 implementation averages squared residuals over all N×3N\times3 coordinates, giving the historical 3N3N normalization.

In the same pinned release, the separate evaluator for Linear, IDPP, Geodesic, and NeuralNEB uses NN-normalized, proper-rotation TS-RMSD through _kabsch_rmsd, so no additional 3\sqrt{3} conversion is applied to those rows. Their TS alignment excludes reflections, whereas the DriftReact TS evaluator also considers mirror matching. These literature-reported results remain distinct from the common-protocol comparison of rerun DriftReact and ReacFlow predictions in Table R10.

Table R9. Source values, submitted values, and the reporting correction.

DatasetMetricSource-reported valueValue in our submissionTS converted to N; other metrics restored
Transition1xTS-RMSD0.151 ± 0.0930.262 ± 0.1610.262 ± 0.161
Transition1xIRC-RMSD0.223 ± 0.1230.386 ± 0.2130.223 ± 0.123
Transition1xFréchet0.460 ± 0.2580.797 ± 0.4470.460 ± 0.258
Halo8TS-RMSD0.107 ± 0.0860.185 ± 0.1490.185 ± 0.149
Halo8IRC-RMSD0.123 ± 0.1090.213 ± 0.1890.123 ± 0.109
Halo8Fréchet0.212 ± 0.1890.367 ± 0.3270.212 ± 0.189

The same factor scales means and SDs. The restored IRC and Fréchet values come directly from the source table. Table R9 records the historical arithmetic correction; Table R10 separately reports predictions generated in our reruns and scored under the shared protocol.

Normalization correction in the DriftReact evaluator. For commit 67eb280 with its locked pymatgen==2025.10.7, the following minimal patch to scripts/evaluate.py converts TS-RMSD from 3N3N to NN normalization before aggregating reaction-level means and SDs:

- ts_rmsds.append(rmsd_coords(p[ts], t[ts], nums, same_order=True))
+ ts_rmsds.append(np.sqrt(3.0) * rmsd_coords(p[ts], t[ts], nums, same_order=True))

The existing valid-atom filtering, reference-peak selection, and TS mirror handling are retained. IRC-RMSD and Fréchet already use NN-normalized point distances and receive no additional factor.

Tables prefixed “R” belong to this response; unprefixed table numbers refer to the submission. Geometry errors are in Å, with lower values better.

Shared evaluation protocol. Generated-geometry results follow the benchmark conventions in manuscript Appendix C.1, including Tables R1–R6, the common evaluation of reruns in Table R10, the input-placement experiment in Table R14, and the additional diagnostics in Tables R16–R18. Each metric uses the same scoring rule across compared methods and tables, with metric-specific alignment as specified below.

ItemRule
Evaluation inputsThe same benchmark reaction IDs, reference paths, and ten-frame grids within each dataset comparison; padding is excluded using valid-atom masks. Any restricted subset is identified explicitly.
Atom correspondenceFixed according to the benchmark atom mapping; no atom rematching.
AlignmentTS-RMSD takes the smaller aligned RMSD obtained with and without mirror reflection. IRC-RMSD and Fréchet use centered, proper-rotation Kabsch alignment (determinant +1), without reflections.
NormalizationPer atom: the squared coordinate residuals are summed over the three coordinates and valid atoms, divided by the number of valid atoms N, then square-rooted.
TS-RMSD / reference-peak RMSDFor a full path, compare generated and reference structures at the first maximum of the reference energies over all ten frames. Predicted energies are not used to select this frame. A TS-only method's predicted structure is compared with this same reference structure.
IRC-RMSDAverage the aligned corresponding-frame RMSDs over all ten frames, including both endpoints.
FréchetDiscrete Fréchet distance over complete paths, using proper-rotation, per-atom RMSD as the point distance.
AggregationReactions are equally weighted, with no error-based trimming of geometry scores. Measured SDs describe variation across reactions.

“TS-RMSD” and “reference-peak RMSD” refer to the same scoring rule above. The released evaluator specifies TS mirror handling and proper-rotation path alignment separately. Clean reconstruction remains raw, per-atom RMSD over all ten frames in the common coordinate frame, before endpoint restoration and without independent alignment. Training losses retain their per-coordinate convention; the SO(3), eight-internal-frame Path diagnostic in Appendix C.2 is separate from benchmark IRC-RMSD.

Evaluation provenance. Table R10 evaluates predictions from our reruns of both models under the shared protocol; its ReacFlow scores agree with the manuscript values at the reported precision. Table R9 documents the historical normalization correction to source-reported statistics. MolGEN's single-sample, all-candidate mean, and reference-selected best-of-30 results are distinguished explicitly; its published energy-selected result is a separate literature value.

Reruns and common evaluation. We independently retrained DriftReact using its released implementation, retaining the original training data, configuration, and checkpoint-selection procedure, with training seed 42. We reran ReacFlow and evaluated both methods with a shared evaluator using the same reaction IDs, reference paths, ten-frame grids, and valid-atom masks. Metric-specific alignment and per-atom normalization follow manuscript Appendix C.1 and the shared protocol above. DriftReact's reproduced mean geometry errors are close to the corrected source values in Table R9, with relative deviations of approximately 0.8–1.9% across the six dataset–metric pairs, computed from the displayed rounded means. ReacFlow's scores agree with the manuscript values at the reported precision.

Table R10. Rerun results for both models under the shared evaluation protocol. Mean ± reaction-level SD.

DatasetMethodReference-peak RMSD ↓IRC-RMSD ↓Fréchet ↓
Transition1xDriftReact0.264 ± 0.1630.225 ± 0.1250.464 ± 0.262
Transition1xReacFlow0.215 ± 0.1430.155 ± 0.1180.260 ± 0.215
Halo8DriftReact0.188 ± 0.1520.125 ± 0.1110.216 ± 0.193
Halo8ReacFlow0.175 ± 0.1280.165 ± 0.1260.280 ± 0.231

Under these shared rules, ReacFlow has lower means for all three metrics on Transition1x. On Halo8, ReacFlow has lower reference-peak RMSD, while DriftReact has lower IRC-RMSD and Fréchet error. The paired 95% intervals in Table R11 exclude zero in these respective directions. The results do not support uniform geometric superiority across both datasets. The baseline energy statistics were copied without conversion, and ReacFlow's energy results are unchanged. In manuscript Table 1, ReacFlow retains the lowest reported barrier error on both datasets and the lowest reported per-image energy error on Halo8.

Tables R11–R13 quantify uncertainty using paired bootstrap analysis of the reaction-level errors underlying Tables R10, R4, and R1, respectively. For a target method A and comparator B, define

\bar{\Delta}=\frac{1}{M}\sum_{i=1}^{M}\Delta_i.$$ Negative differences favor A. For each comparison, we draw 10,000 bootstrap resamples, each containing M reaction IDs sampled with replacement. The two methods' saved errors are resampled together for each reaction, preserving their pairing. We compute the mean paired difference in each resample; its 2.5th and 97.5th percentiles define the 95% confidence interval. Transition1x comparisons contain 1,073 evaluation reactions and Halo8 comparisons contain 1,891. All differences and interval bounds below are in Å. The intervals in Tables R11–R13 characterize reaction-level sampling variability conditional on the fixed trained models. **Table R11. Paired-bootstrap comparison of ReacFlow and DriftReact under the common evaluation protocol, based on Table R10.** | Dataset | Metric | DriftReact | ReacFlow | Mean paired difference | 95% CI | | --- | --- | ---: | ---: | ---: | --- | | Transition1x | TS-RMSD | 0.264 | 0.215 | −0.049 | [−0.0576, −0.0407] | | Transition1x | IRC-RMSD | 0.225 | 0.155 | −0.070 | [−0.0769, −0.0632] | | Transition1x | Fréchet | 0.464 | 0.260 | −0.204 | [−0.2175, −0.1908] | | Halo8 | TS-RMSD | 0.188 | 0.175 | −0.013 | [−0.0216, −0.0044] | | Halo8 | IRC-RMSD | 0.125 | 0.165 | +0.040 | [+0.0348, +0.0452] | | Halo8 | Fréchet | 0.216 | 0.280 | +0.064 | [+0.0545, +0.0734] | *Note.* Means reproduce Table R10, which also reports reaction-level SDs. Differences are ReacFlow minus DriftReact; intervals use the paired reaction-level errors from that common evaluation and the bootstrap procedure above. The three Transition1x intervals favor ReacFlow. On Halo8, the TS-RMSD interval favors ReacFlow, while the IRC-RMSD and Fréchet intervals favor DriftReact. The tokenizer contribution is supported separately by the controlled comparisons in Tables R1, R5, and R13. **Changes.** We have described the reporting correction and the shared protocol in Appendix C.1, added the rerun comparison of Table R10 to the appendix, and updated Table 1, its caption and rankings, the abstract, and the corresponding claims in the introduction, Section 4.2, and Appendix C.1, removing the claim of lowest geometry error on both datasets. ## Q2. Endpoint alignment assumptions > “Does the method require a specific relative alignment between the reactant and product structures?” Yes. The model requires atom-mapped endpoints in a common relative coordinate frame, as supplied by the stored benchmark pathways. Preprocessing subtracts the reactant center from the whole path and does not independently align the product before forming $L_k=(1-s_k)R+s_kP$. The benchmark endpoints are the first and last structures of the same stored path, with fixed atom correspondence and a shared coordinate frame. A common rigid transformation preserves their relative placement; moving only the product changes the source interpolation and the model's conditioning. Table R14 reports a separate sensitivity experiment using the archived Full (R+S+C) checkpoint, with inference repeated for each input-placement condition and no weight updates. Its remeasured unperturbed scores agree with the manuscript's Full row when rounded to the manuscript's three decimal places. The checkpoint and replay details are recorded with Table R14. **Measured sensitivity (Table R14).** We evaluated all **1,073 Transition1x reactions** using the archived Full (R+S+C) checkpoint, with no retraining, no reaction filtering, and zero failed outputs in every condition. Product-only rotations of 5°, 15°, and 45° increase all three errors progressively; a 5° rotation changes them by less than 1.5%. At 45°, TS-RMSD, IRC-RMSD, and Fréchet increase by **27.2%, 24.5%, and 23.3%**, respectively; a 0.5 Å product translation increases them by **18.4%, 16.0%, and 15.4%**.
Reviewer_1g4SOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper investigates the problem of predicting the pathway of a chemical reaction, which is important for mechanistic analysis and barrier estimation of reactions. Since a reaction pathway is not an arbitrary sequence of geometries but the result of a saddle search, the authors aim at learning latent representations that satisfy two requirements: that it keep what that search moves, and that it be predictable from the endpoints alone. Their proposed method REACFLOW (1) trains a tokenizer on reaction relevance as well as conditional predictability and (2) freezes the tokenizer while training a conditional flow-matching model that learns to transport in the latent space. REACFLOW was compared experimentally against four baselines on two benchmarks, demonstrating significantly improved accuracy of transportability at comparable fidelity of reconstruction.

Critical strengths

Technical quality Major design choices of the proposed method are well-motivated, e.g. the choice of a shallow head to predict the projected-force response: "A pathway is not just a shape to reconstruct: it is the answer to a search problem. We therefore attach a deliberately shallow head to the latent and require it to predict a CI-inspired projected-force response to a candidate path." A theoretical analysis
provides a necessary latent-separation condition for accurate search response prediction and a bound for the error of an endpoint-conditioned latent predictor.

Supportive experimental results REACFLOW was compared experimentally against four baselines on two benchmarks (Transition1x and Halo8). The experimental results show significantly improved accuracy of transportability at comparable fidelity of reconstruction. An ablation study investigates the question "What makes a latent transportable?" and performs experiments for four tokenizer variants, which differ only in their pretraining objective.

Critical weaknesses

Choice of baselines for the experiments I am not (yet) convinced of the choice of the baselines. You used experimental results for the baselines from Schlama & Schwaller (2026), but that does justify the choice of the baselines. I would also like to see some results for a string method. Such results will provide an upper bound for the accuracy of ML-methods based on endpoints as well as evidence to support your claim that they do not scale for screening. Please see my questions below.

Questions for authors
  1. Why did you choose the given baselines? BaseIDPP and Geodesic are not discussed under Related Work. NeuralNEB (Schreiner et al., 2022b) is old. Why not use the similar, more recent method of (Ramakrishnan et al., 2025)?

  2. You mention that chain-of-states methods, in particular string methods such as the method of (Henkelman et al., 2000) remain the reference but are too inefficient to scale to screening applications. How does the reconstruction fidelity of REACFLOW compare to that of the (Henkelman et al., 2000) method? And how do the runtimes compare?

5 comments

W / Q1. Choice of baselines and related work

“Why did you choose the given baselines? BaseIDPP and Geodesic are not discussed under Related Work.”

We selected the baselines to test three complementary ways of constructing a complete path from endpoints: geometry-only interpolation, energy/force-driven refinement, and direct learned prediction. Their inclusion in a previous benchmark is a source of comparable results, but is not by itself a sufficient selection rationale.

  • Geometric baselines: Linear, IDPP, and Geodesic. These test how much pathway accuracy can be obtained from endpoint geometry without learning reaction-path examples or evaluating an electronic-structure energy surface. IDPP constructs an auxiliary objective from interpolated interatomic distances and optimizes an initial path on that objective. Geodesic interpolation constructs a path using a geometry-derived metric. Both provide more structured initial guesses than Cartesian linear interpolation; neither is a converged search on the physical potential-energy surface.
  • Energy/force-driven refinement: NeuralNEB. This tests iterative path optimization using a learned potential, including whether force-guided refinement improves on a geometric starting path. The submission uses the reported PaiNN/CI-NEB benchmark results; our additional runs in Tables R6–R8 report geometry, runtime, and convergence for the explicitly specified finite-budget setup.
  • Direct learned prediction: Drift-React. This tests complete-path generation without iterative energy/force evaluations at inference, and is the closest original comparator for the endpoint-conditioned output task. The corrected reporting, rerun comparison, and paired intervals are given in Tables R9–R11.

Expanded coverage. Table R2 adds MEPIN-L/G as transferable full-path predictors and React-OT and MolGEN as TS predictors. Appendix C.1 of the submission explains that TS-only methods were excluded from the full-path table because their outputs do not support the same path metrics, and that MEPIN required a separate comparison controlling the split and reference level. We now report the added methods on Transition1x with their training settings and output scope stated explicitly: TS-only methods receive TS-RMSD scores, with path metrics left unreported. Among the full-path methods in Table R2, ReacFlow has the lowest error on all three metrics. The TS-only comparisons with React-OT and MolGEN, including the distinction between MolGEN's single-sample, oracle, and published energy-selected results, are discussed in W2 / Q1 of our response to 6LJa.

Relation to Ramakrishnan et al. The submitted Related Work already discusses their implicit neural representation for path search; we have made the distinction more explicit in the revision. Their atomistic experiments optimize a continuous neural path separately for each system using energies and path-projected gradients, while their transferable prediction experiment is a proof of concept on two-dimensional systems. ReacFlow learns across atomistic reaction examples and predicts a path without energy/force calls at inference. Their atomistic optimizer therefore belongs to the same family as the search procedure in Tables R6–R8, which needs energy and force evaluations for every new system. A quantitative comparison would require a dedicated implementation on a common potential, reaction set, and convergence/budget protocol, so we discuss their method as related work and do not include it as a baseline.

Changes. We have added the IDPP and geodesic references and descriptions to Related Work, stated the above selection rationale and output distinctions in the experimental setup, expanded the discussion of Ramakrishnan et al., and updated Appendix C.1 to distinguish the original literature results from the additional evaluations.

Q2. Accuracy and runtime relative to CI-NEB

“How does the reconstruction fidelity of REACFLOW compare to that of the (Henkelman et al., 2000) method? And how do the runtimes compare?”

Transition1x reference paths come from DFT-based NEB/CI-NEB, and Halo8 uses a multi-level workflow with DFT labeling. Comparing a stored reference path with itself gives zero error by definition; an independent DFT-CI-NEB run need not reproduce it pointwise. Tables R6–R8 compare ReacFlow with finite-budget CI-NEB on the learned PaiNN potential-energy surface, using the same reaction sets, shared geometry metrics, and the solver settings below.

Tables prefixed “R” belong to this response; unprefixed table numbers refer to the submission. Geometry errors are in Å, with lower values better.

Shared evaluation protocol. Generated-geometry results follow the benchmark conventions in manuscript Appendix C.1, including Tables R1–R6, the common evaluation of reruns in Table R10, the input-placement experiment in Table R14, and the additional diagnostics in Tables R16–R18. Each metric uses the same scoring rule across compared methods and tables, with metric-specific alignment as specified below.

ItemRule
Evaluation inputsThe same benchmark reaction IDs, reference paths, and ten-frame grids within each dataset comparison; padding is excluded using valid-atom masks. Any restricted subset is identified explicitly.
Atom correspondenceFixed according to the benchmark atom mapping; no atom rematching.
AlignmentTS-RMSD takes the smaller aligned RMSD obtained with and without mirror reflection. IRC-RMSD and Fréchet use centered, proper-rotation Kabsch alignment (determinant +1), without reflections.
NormalizationPer atom: the squared coordinate residuals are summed over the three coordinates and valid atoms, divided by the number of valid atoms N, then square-rooted.
TS-RMSD / reference-peak RMSDFor a full path, compare generated and reference structures at the first maximum of the reference energies over all ten frames. Predicted energies are not used to select this frame. A TS-only method's predicted structure is compared with this same reference structure.
IRC-RMSDAverage the aligned corresponding-frame RMSDs over all ten frames, including both endpoints.
FréchetDiscrete Fréchet distance over complete paths, using proper-rotation, per-atom RMSD as the point distance.
AggregationReactions are equally weighted, with no error-based trimming of geometry scores. Measured SDs describe variation across reactions.

“TS-RMSD” and “reference-peak RMSD” refer to the same scoring rule above. The released evaluator specifies TS mirror handling and proper-rotation path alignment separately. Clean reconstruction remains raw, per-atom RMSD over all ten frames in the common coordinate frame, before endpoint restoration and without independent alignment. Training losses retain their per-coordinate convention; the SO(3), eight-internal-frame Path diagnostic in Appendix C.2 is separate from benchmark IRC-RMSD.

Setup. We use released held-out-fold PaiNN checkpoints for Transition1x and training-split PaiNN selected by validation force MAE for Halo8. Both use ten images, fixed endpoints, and IDPP. NEB/CI-NEB each allow 1,000 steps, excluding IDPP, with early stopping at 0.10/0.05 eV/Å, respectively. These additional runs use the checkpoints and budgets specified here; they are not a reproduction of manuscript Table 1's literature-sourced NeuralNEB row.

Table R6. Full-set and paired converged-subset geometry under the shared evaluation protocol. Mean errors in Å; lower is better.

Dataset / scopeMethodReference-peak RMSD ↓IRC-RMSD ↓Fréchet ↓
Transition1x, all (1,073)ReacFlow0.2150.1550.260
Transition1x, all (1,073)CI-NEB0.5710.3210.689
Transition1x, converged (889)ReacFlow0.2050.1490.249
Transition1x, converged (889)CI-NEB0.3390.2100.324
Halo8, all (1,891)ReacFlow0.1750.1650.280
Halo8, all (1,891)CI-NEB0.8590.5131.226
Halo8, converged (1,221)ReacFlow0.1690.1600.270
Halo8, converged (1,221)CI-NEB0.5040.3010.510

Both methods use identical reaction IDs within each scope; subsets are selected solely by CI-NEB convergence. Full-set results retain unconverged runs' last finite paths. ReacFlow's lower mean errors persist on the paired converged subsets.

Table R7. Same-machine generation time over the complete evaluation sets. Mean ms per reaction.

DatasetEvaluation reactionsPaiNN-driven CI-NEBFull ReacFlowRatio of mean times
Transition1x1,07328,794.0458.26494×
Halo81,89153,715.0053.97995×

Timing uses one RTX 5090, two CPU threads, and sequential execution, including budget-exhausted runs. CI-NEB includes IDPP and energy/force calls; ReacFlow includes encoding, four velocity evaluations, decoding, and endpoint restoration. Both exclude loading, file I/O, and subsequent DFT evaluation. The measured 494× / 995× speedups apply to this PaiNN-driven solver under the stated hardware, initialization, convergence thresholds, and optimization budgets.

Table R8. CI-NEB convergence on the complete evaluation sets.

DatasetEvaluation reactionsConverged, n (%)Unconverged: final-stage budget exhaustedNumerical failures
Transition1x1,073889 (82.85%)17.15%0
Halo81,8911,221 (64.57%)35.43%0

Convergence requires maximum NEB force ≤ 0.05 eV/Å on PaiNN, not DFT. Potential approximation and optimization errors remain entangled.

Scope. We did not perform an independent, matched DFT-CI-NEB comparison. Tables R6–R8 therefore establish neither a DFT-level accuracy bound nor a measured DFT speedup, and do not show that ReacFlow outperforms converged CI-NEB on the DFT potential-energy surface.

Changes. We have added the accuracy and runtime comparisons (Tables R6 and R7), together with the convergence rates, to the appendix, and revised the screening-cost statement in the introduction to refer specifically to the PaiNN-driven solver and settings tested here.

Reviewer_6LJaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents ReacFlow, a latent flow matching framework that encodes molecular geometry to learn reaction pathways in latent space by aligning the flow with the actual transition pathway using energies and forces from simulation data. On proposed benchmark datasets, ReacFlow recovers ground turth energies while capturing the geometry of the transition pathway.

Critical strengths
  • I like the disclaimers provided by the authors about the visual checks not signifying the exact transition state. The theorems seem to be provide upper bounds on performance as well, which contextualises the results more.
  • I like the model ablations provided in the paper, which explains how different components affect the overall framework (eg: Table 2, Table 3).
Critical weaknesses
  • Novelty is somewhat underwhelming here. Two-stage latent flows have been used quite a bit and unified tokenizers with multiple objectives have been explroed before, especially for molecular modeling.

  • Some existing baselines are excluded which should have been there, like MEPIN, React-OT, and MolGen. Perhaps the authors could provide these.

  • Calling it flow matching is a bit of a strecth since the flow loss weight is very low at 0.1 modulation while the geometric loss is upweighted. This just means there might be quite a bit of overfitting on the supplied TS path with a risk of poor generalisation. Could the authors please explain their logic for this design choice?

Questions for authors
  1. Can the authors please include the other baselines like MEPIN, React-OT, and MolGen as well? That'd make the analysis a lot fairer.

  2. Please explain the rationale of overpowering the underlying flow matching with the high geometry loss weight? It seems like latent regression at best rather than using the generative capabilities of flow matching. Is there a way to get multiple pathways? It seems like you're overfitting to the single pathway provided by the dataset(s) and this makes it hard to verify other possible paths.

8 comments

W1. Novelty of the tokenizer and latent transport

“Two-stage latent flows have been used quite a bit and unified tokenizers with multiple objectives have been explored before, especially for molecular modeling.”

We agree that two-stage latent generation and multi-objective tokenizers are established. Our contribution is shaping a reaction-pathway representation to preserve search-relevant information while supporting prediction from endpoints, and testing these objectives at fixed downstream architecture and update budget. Prior work observed the reconstruction–generation gap for image and video tokenizers; we quantify it for reaction pathways under these controlled conditions and identify reaction-specific objectives that reduce it. The specific distinctions are:

Prior workEstablished approachSpecific distinction in ReacFlow
l-DeTok (Yang et al.); PV-VAE (Zhao et al.)Image-tokenizer denoising under corruption; video-tokenizer prediction of future frames from past observations.Completion replaces an internal geometric interval by interpolation between retained boundaries, retaining both reaction endpoints. Full internal masking gives exactly the generator's endpoint-linear source.
EGInterpolator (Iyengar et al.); RigidSSL (Ni et al.)Conformer diffusion pretraining followed by temporal MD modeling; bidirectional rigid-body FM pretraining using perturbed structures and MD conformations.A full-path autoencoder learns projected search responses and endpoint-anchored completion, then is frozen for a separate latent flow. The learned object and pretraining targets differ.
MEPIN; Drift-ReactCoordinate-path prediction using geometric priors and an energy objective; one-step pathway generation trained with a drifting field.Transport operates between learned path codes. We vary how those codes are pretrained while keeping the downstream flow architecture and update budget fixed.

Why these objectives. Search supervision (S) trains a shallow head to recover transverse and climbing-image projected-force responses from the candidate-path latent and endpoints. This is a CI-inspired target without the NEB spring term, not a full optimizer update. Completion (C) trains the encoder and decoder on boundary-interpolated inputs, including the endpoint-only source. S addresses what survives compression; C addresses what can be inferred from partial context. The search head is discarded before transport training; generation requires no energies or forces.

Table R1. Tokenizer objectives on Transition1x, reproduced from manuscript Table 2.

Tokenizer pretrainingClean reconstruction ↓Reference-peak RMSD ↓IRC-RMSD ↓Fréchet ↓
Reconstruction only0.0520.2850.2450.410
Reconstruction + search response0.0540.2350.2050.340
Reconstruction + interval completion0.0560.2250.1900.310
Reconstruction + both objectives0.0550.2150.1550.260

With the same downstream architecture and update budget, reconstruction stays within 0.052–0.056 Å while generated IRC-RMSD improves by 36.7%. Manuscript Table 3 tests the specific choices beyond simply adding auxiliary objectives:

  • Search target: ordinary-force supervision gives IRC-RMSD 0.175 Å, versus 0.155 Å for CI-inspired responses, with the same head and target-RMS normalization procedure.
  • Completion structure: random-frame masking gives 0.180 Å, despite matching the hidden-frame count, full-mask probability, endpoints, and training budget. Thus the difference is not merely exposure to endpoint-only inputs.
  • Encoder supervision: blocking completion gradients to the encoder and shared parameters gives 0.175 Å. This supports the direct encoder-training route.

These controlled objective substitutions support the task-specific design under the tested setting; the matched-budget FM-versus-terminal-supervision comparison is addressed separately in W3 / Q2 (Table R4).

Paired uncertainty. Table R13 reports 95% paired-bootstrap intervals for R+S+C minus R. The IRC-RMSD difference is −0.090 Å [−0.0983, −0.0820]; the TS-RMSD and Fréchet intervals also lie below zero. These intervals describe the fixed models' reaction-level sampling uncertainty.

Changes. We have added these direct distinctions to the introduction and Related Work and linked each design to its control in manuscript Table 3. Our novelty claim concerns the reaction-specific objectives and their controlled evidence, not the general two-stage framework.

Tables prefixed “R” belong to this response; unprefixed table numbers refer to the submission. Geometry errors are in Å, with lower values better.

Shared evaluation protocol. Generated-geometry results follow the benchmark conventions in manuscript Appendix C.1, including Tables R1–R6, the common evaluation of reruns in Table R10, the input-placement experiment in Table R14, and the additional diagnostics in Tables R16–R18. Each metric uses the same scoring rule across compared methods and tables, with metric-specific alignment as specified below.

ItemRule
Evaluation inputsThe same benchmark reaction IDs, reference paths, and ten-frame grids within each dataset comparison; padding is excluded using valid-atom masks. Any restricted subset is identified explicitly.
Atom correspondenceFixed according to the benchmark atom mapping; no atom rematching.
AlignmentTS-RMSD takes the smaller aligned RMSD obtained with and without mirror reflection. IRC-RMSD and Fréchet use centered, proper-rotation Kabsch alignment (determinant +1), without reflections.
NormalizationPer atom: the squared coordinate residuals are summed over the three coordinates and valid atoms, divided by the number of valid atoms N, then square-rooted.
TS-RMSD / reference-peak RMSDFor a full path, compare generated and reference structures at the first maximum of the reference energies over all ten frames. Predicted energies are not used to select this frame. A TS-only method's predicted structure is compared with this same reference structure.
IRC-RMSDAverage the aligned corresponding-frame RMSDs over all ten frames, including both endpoints.
FréchetDiscrete Fréchet distance over complete paths, using proper-rotation, per-atom RMSD as the point distance.
AggregationReactions are equally weighted, with no error-based trimming of geometry scores. Measured SDs describe variation across reactions.

“TS-RMSD” and “reference-peak RMSD” refer to the same scoring rule above. The released evaluator specifies TS mirror handling and proper-rotation path alignment separately. Clean reconstruction remains raw, per-atom RMSD over all ten frames in the common coordinate frame, before endpoint restoration and without independent alignment. Training losses retain their per-coordinate convention; the SO(3), eight-internal-frame Path diagnostic in Appendix C.2 is separate from benchmark IRC-RMSD.

Tables R11–R13 quantify uncertainty using paired bootstrap analysis of the reaction-level errors underlying Tables R10, R4, and R1, respectively. For a target method A and comparator B, define

\bar{\Delta}=\frac{1}{M}\sum_{i=1}^{M}\Delta_i.$$ Negative differences favor A. For each comparison, we draw 10,000 bootstrap resamples, each containing M reaction IDs sampled with replacement. The two methods' saved errors are resampled together for each reaction, preserving their pairing. We compute the mean paired difference in each resample; its 2.5th and 97.5th percentiles define the 95% confidence interval. Transition1x comparisons contain 1,073 evaluation reactions and Halo8 comparisons contain 1,891. All differences and interval bounds below are in Å. The intervals in Tables R11–R13 characterize reaction-level sampling variability conditional on the fixed trained models. **Table R13. Paired-bootstrap comparison of reconstruction-only and full tokenizer pretraining on Transition1x, based on Table R1.** | Metric | Reconstruction only (R) | Full (R+S+C) | Mean paired difference | 95% CI | | --- | ---: | ---: | ---: | --- | | TS-RMSD | 0.285 | 0.215 | −0.070 | [−0.0790, −0.0613] | | IRC-RMSD | 0.245 | 0.155 | −0.090 | [−0.0983, −0.0820] | | Fréchet | 0.410 | 0.260 | −0.150 | [−0.1646, −0.1357] | *Note.* Mean errors reproduce manuscript Table 2 and rebuttal Table R1. Differences are R+S+C minus R, using the same 1,073 evaluation reactions from the original controlled evaluation. The paired differences and intervals favor full pretraining on all three generation metrics. The mean reductions are 24.6% in TS-RMSD, 36.7% in IRC-RMSD, and 36.6% in Fréchet distance, while clean reconstruction remains comparable at 0.052 Å for R and 0.055 Å for R+S+C. Thus the more accurate reconstruction-only tokenizer does not yield the better endpoint-conditioned generator in this controlled comparison.

W2 / Q1. MEPIN, React-OT, and MolGEN

“Can the authors please include the other baselines like MEPIN, React-OT, and MolGen as well?”

We rerun MEPIN-L, MEPIN-G, and MolGEN on this benchmark, retraining on the manuscript's training partition and evaluating on the same 1,073 Transition1x reactions as ReacFlow. React-OT uses its released checkpoint with the matching partition. MolGEN's released training list overlaps 1,008 evaluation reactions, motivating retraining; MEPIN's released training overlap is zero, and its published test set shares 316 reactions with ours. We retrain it to match the training reaction pool across methods.

We retain the native architectures and objectives. MEPIN uses GFN1-xTB energies and forces in its variational objective, with training-only geometric pretraining for MEPIN-L and geodesic initialization for MEPIN-G. MolGEN uses FM followed by the authors' modified JS loss. The retrained models use no released pretrained weights. The common evaluation reference is the DFT benchmark.

Table R2. Added baselines on Transition1x under the shared geometry scoring rules, with output and selection settings stated below. Mean ± reaction-level SD.

MethodEvaluated outputTS-RMSD ↓IRC-RMSD ↓Fréchet ↓
MEPIN-LComplete ten-frame path0.324 ± 0.2210.223 ± 0.1470.358 ± 0.249
MEPIN-GComplete ten-frame path0.316 ± 0.2140.218 ± 0.1450.347 ± 0.242
React-OTPredicted TS structure0.174 ± 0.126——
MolGEN (single sample)First TS candidate0.5434 ± 0.231——
MolGEN (all samples)Mean error over all 30 TS candidates0.5396 ± 0.217——
MolGEN (oracle)Reference-selected best-of-30 TS0.1573 ± 0.199——
ReacFlow (ours; manuscript Table 1)Complete ten-frame path0.215 ± 0.1430.155 ± 0.1180.260 ± 0.215

Evaluation. All methods use the shared geometry scoring rules above: TS-RMSD includes mirror matching, while IRC-RMSD and Fréchet use proper rotations only. Full-path methods are scored at the reference peak on the ten-frame grid; the structures predicted directly by React-OT and MolGEN are compared with that same reference structure. ReacFlow retains its original evaluation results, also reported in manuscript Table 1. All runs use fixed seeds; measured SDs describe variation across reactions.

MolGEN evaluation and candidate selection. We generate 30 independent TS candidates per evaluation reaction. Table R2 reports the first-candidate error, the mean error over all 30 candidates, and the reference-selected minimum, each averaged over the 1,073 reactions. This evaluation uses no energy calculations or energy-based reranking. For a fixed candidate set, reference-assisted best-of-30 gives a lower bound on TS-RMSD for any rule selecting one of those candidates, including energy ranking. It is an oracle candidate-quality diagnostic that requires the reference TS, so it is not a deployable predictor. MolGEN's published 0.106 Å uses lowest-energy selection among 30 candidates, which requires an energy evaluation of every candidate; ReacFlow returns one path and uses no energy evaluations. The single-sample row provides a one-candidate, energy-free comparison. The published number also comes from a different training partition, and we make no claim about MolGEN's energy-selected pipeline.

React-OT's published 0.103 Å corresponds to approximately 0.178 Å under a normalization-only conversion from its repository's historical per-coordinate (3N3N) convention to per-atom (NN) RMSD (evaluation code). The 0.174 Å in Table R2 is measured with our shared evaluator and is numerically close to this converted value.

Reading Table R2. Among the full-path methods in Table R2, ReacFlow has the lowest error on all three metrics: its TS-RMSD is 32–34% lower, its IRC-RMSD 29–30% lower, and its Fréchet distance 25–27% lower than those of MEPIN-G and MEPIN-L. ReacFlow's TS-RMSD is also 60% lower than MolGEN's in our retrained single-sample setting. React-OT and MolGEN's reference-selected best-of-30 oracle achieve lower TS-RMSD (0.174 and 0.1573 Å, respectively, versus ReacFlow's 0.215 Å). These methods return a TS structure, while ReacFlow generates the complete pathway; the oracle additionally requires the reference TS for selection.

Changes. We have added these comparisons and their settings. Table R2 covers Transition1x.

Comments on the paper

1 comment

Dear reviewers,

As the authors have posted their responses, I would encourage you to actively parcitipate in the dicussion.

Thank you.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.