acceptodds
Under review as a conference paper at ICLR 2027

Mechanistic Reaction Prediction via Discrete Flow Matching on Graph-Structured Electron Occupation

Abstract

Chemical reactions span diverse molecules and transformation types, but available data cover them unevenly, making prediction beyond the training distribution a central challenge. Chemists reason about reactivity through arrow-pushing: electrons move from a donor site to an acceptor site, and that direction helps determine which bonds form. Most reaction predictors instead generate products or edit molecular topology without explicitly representing these transfers. We introduce MAELLE, a discrete flow-matching model that predicts directional electron-pair edits on a graph of bonding, non-bonding, and hydrogen sites. Optimal transport infers an unordered set of edits from each reactant–product pair, allowing training without annotated elementary steps. The resulting stochastic trajectories retain the reactants' heavy atoms. MAELLE achieves 93.9% top-5 accuracy on USPTO-480K and 59.3% top-5 accuracy on the literature-derived AbSynth benchmark, the highest AbSynth result among the evaluated baselines. In a separate analysis of expert-annotated reactions, its electron sources and sinks agree with the annotations for 95.4% of recovered persistent bond-forming steps in trajectories reaching the recorded product. These results support endpoint-supervised directional electron flow as a useful representation for reaction prediction beyond pattern recognition.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Discrete Flow Matching, Graph Neural Network, Reaction Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vUW8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents MAELLE, a discrete flow-matching framework for reaction outcome prediction that represents chemical reactions as directional electron-pair edits. An optimal transport (OT) procedure constructs an unordered set of edits from reactant–product pairs, and a continuous-time Markov chain is trained to generate product structures without elementary-step annotations. The paper evaluates the method on USPTO-480K and AbSynth and analyzes agreement between predicted electron flows and expert-annotated mechanisms.

Critical strengths
  1. The paper proposes an interesting framework that combines directional electron-pair edits with discrete flow matching, enabling reaction prediction without elementary-step annotations.
Critical weaknesses
  1. Insufficient baseline comparisons. The paper omits stronger existing methods on USPTO-480K, including NERF* (91.5% Top-1, 95.1% Top-5), ReactionSink (91.3%, 94.5%), Chemformer-Large (92.8% Top-1 under the separated-input setting) [1], and RXNGraphormer[2]. These results exceed MAELLE's 87.2% Top-1 accuracy, raising concerns about its competitiveness.

  2. Mismatch between OT and chemical motivation. The OT objective minimizes graph-distance costs rather than chemical plausibility or mechanistic feasibility. The inferred edits are therefore optimal under a heuristic cost, but are not necessarily chemically meaningful. Since these edits define the training targets, this weakens the justification for interpreting generated trajectories as mechanistic electron flows.

[1]: Joonyoung F. Joung, Nicholas Casetti, Priyanka Raghavan, Connor W. Coley; An overview of reaction outcome prediction with physics-based and data-driven methods. Chem. Soc. Rev. 2026; 55 (12): 6768–6813. https://doi.org/10.1039/d6cs00079g [2]: Xu, LC., Tang, MJ., An, J. et al. A unified pre-trained deep learning framework for cross-task reaction performance prediction and synthesis planning. Nat Mach Intell 7, 1561–1571 (2025). https://doi.org/10.1038/s42256-025-01098-4

Questions for authors
  1. Why were stronger and more recent baselines, particularly NERF* and ReactionSink, excluded? Can the authors provide comparisons under consistent evaluation protocols?

  2. How do the authors justify the chemical validity of OT-inferred electron transfers when the transport cost only reflects graph distances? What shows that these edits correspond to chemically meaningful pathways rather than arbitrary endpoint alignments?

4 comments
AuthorsOpenReview ↗

Response to Reviewer vUW8

Thank you for your constructive feedback. We address your concerns below. Before we go into that, as reviewer L2ch points out, the performance of NERF is indeed lower than usual, and we identified and corrected bugs in our evaluation script. First, we would like to provide an updated main result:

Updated main results

ModelUSPTO Top-1Top-3Top-5AbSynth Top-1Top-3Top-5
Molecular Transformer0.8860.9350.9420.3540.4190.441
Graph2SMILES0.9030.9400.9480.3950.4520.471
MEGAN0.8630.9240.9400.4570.5610.595
NERF0.8980.9130.9160.3950.4290.437
MAELLE0.8830.9340.9420.4820.5810.607

The corrected NERF performance is substantially closer to its originally reported results. Our main conclusion remains unchanged: MAELLE is competitive on USPTO-480K and achieves the highest AbSynth accuracy among the evaluated baselines.

Why were stronger and more recent baselines excluded?

We selected representative template-free methods spanning three prediction modalities: SMILES generation (Molecular Transformer, Graph2SMILES), graph editing (MEGAN), and electron redistribution (NERF). Our goal was to compare these modeling representations under a consistent input setting rather than exhaustively benchmark architectural variants.

We agree that NERF* and ReactionSink are relevant. However, NERF* was reported in the initial version of the preprint. However, the authors subsequently replaced the NERF-based approach with ReactionTeam, a substantially different Graph2SMILES-based architecture, in the latest version. The NERF* method and its originally reported results are no longer presented in the revised paper. For ReactionSink, we could not identify publicly available source code for reproduction and benchmarking on AbSynth.

Moreover, our primary contribution is not a new state-of-the-art USPTO-480K score, but an end-to-end discrete flow matching reaction prediction model with chemically interpretable CTMC trajectories. We evaluate both its generalization to AbSynth and the agreement of its inferred electron-flow directions with expert annotations.

Chemical motivation for graph-distance OT?

We use graph distance to encourage local electron transfers, rather than to approximate reaction energetics. This provides an endpoint-consistent alignment motivated by the locality of bond rearrangements.

Consider the reaction:

CH3OH + CH3Br → CH3OCH3 + HBr

In our electron-occupation representation, the two source sites are O–H and C–Br, while the two sink sites are C–O and Br–H. There are exactly two endpoint-valid one-to-one couplings:

PlanElectron transfersTotal graph-distance cost
A (OT)O–H → C–O; C–Br → Br–H0 + 0 = 0
BO–H → Br–H; C–Br → C–OD + 0 = D

Here, D denotes a large distance assigned to disconnected sites. Plan A couples electron sites sharing an atom, while Plan B includes a transfer between disconnected sites. Thus, graph-distance OT selects the chemically meaningful assignment.

We further evaluated the inferred directions against expert annotations over 17,534 bond-forming events. Graph-distance OT achieves 90.0% directional agreement, compared with 27.7 ± 0.4% for random source–sink coupling (see table). MAELLE's predicted electron-flow directions also exhibit high agreement with expert annotations:

sample_modeGT bondscoveragebonds compareddirection agrees
correct1753412956 (73.9%)1295612297 (94.9%)
top11753414305 (81.6%)1430512933 (90.4%)
all1753417479 (99.7%)1747915589 (89.2%)
Graph-distance-based OT interpolation1753417534 (100.0%)1753415773 (90.0%)
random coupling (16 samples)1753417534 ± 0 (100.0 ± 0.0%)17534 ± 04857 ± 70 (27.7 ± 0.4%)

These results support graph distance as a chemically informative locality prior and show that MAELLE's inferred electron-flow directions frequently agree with expert annotations. We note that we only claim directional agreement, not the recovery of the complete elementary reaction mechanism. We will clarify this motivation and limitation in Section 3.3, and add the two additional rows for graph-distance-based OT and random coupling into Table 2 in our paper.

Thank you for the rebuttal and for correcting the NERF evaluation. I have two remaining concerns.

  1. Comparison with stronger baselines. My request was for a targeted comparison with the methods cited in my review that report higher accuracy on USPTO-480K. The corrected NERF results address one part of this concern. A comparison with a stronger baseline such as RXNGraphormer under the same input and evaluation protocol would help establish MAELLE’s competitiveness. Where a matched rerun is infeasible, please include the published results and identify any protocol differences.
  2. Evidence that graph-distance OT improves OOD prediction. On AbSynth, replacing graph-distance OT with random source–sink assignment changes Top-1 accuracy from 48.2% to 47.5% and Top-5 from 60.7% to 60.0%. The directional-agreement analysis supports graph distance as a locality prior. However, for OOD product prediction, the small accuracy gaps leave its contribution unclear. This weakens the empirical motivation for the graph-distance OT component as a way to improve generalization. For these reasons, I am keeping my original rating of 1.
AuthorsOpenReview ↗

About RXNGraphormer and graph-distance OT

Inclusion of RXNGraphormer into the baselines:

It is fortunate that RXNGraphormer comes with open sourced code and model weights, so it was quick to test the model on AbSynth. We would hereby include RXNGraphormer into our main table:

Top-k accuracy on USPTO-480K and AbSynth, mixed setting, best in bold, second best in italic

ModelPrediction modalityUSPTO Top-1USPTO Top-3USPTO Top-5AbSynth Top-1AbSynth Top-3AbSynth Top-5
Molecular TransformerSMILES0.886*0.935*0.942*0.3540.4190.441
Graph2SMILESSMILES0.903*0.940*0.948*0.3950.4520.471
RXNGraphormerSMILES0.906*0.940*0.949*0.4100.4710.491
MEGANGraph edits0.863*0.924*0.940*0.4570.5610.595
NERFElectron diff0.8980.9130.9160.3950.4290.437
MAELLEElectron flow0.8830.9340.9420.4820.5810.607

* Number reported in the original papers.

Similar to other SMILES-based model, RXNGraphormer performs well on USPTO-480K, but falls short on AbSynth.

The contribution of the graph-distance OT component

As our results show, removing the graph-distance OT component, by performing random coupling instead, only slightly decreases the model performance on end-to-end prediction. This is understandable because both coupling methods ensure end-to-end consistency.

What the graph-distance OT component constributes however, is the chemical interpretability of the CTMC trajectories, evidence by the correct direction of electron moves, using FlowER elementary steps as references.

Additionally, our abalation also shows that when either the OT component or the FLOW operation get removed, MAELLE still performs better than MEGAN (graph-edit-based) and NERF (oneshot electron redistribution) at most sample sizes, suggesting that it was the electron representation + discrete flow matching formulation that helps MAELLE. We would like to emphasize that, and we will do so in the revised version of the paper, these two components are also our novelty, as to the best of our knowledge, MAELLE is also the first discrete flow matching model on electrons for reaction prediction.

Thank you for the additional experiments. The RXNGraphormer comparison addresses my baseline concern. However, my concern about the motivation for graph-distance OT remains. I am keeping my original rating of 1.

Reviewer_1dVeOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MAELLE, a model that predicts reaction products by moving electron pairs between sites in a molecular graph. It uses optimal transport and discrete flow matching to learn from reactant–product pairs, without labeled intermediate steps.

Critical strengths
  1. The method learns directional electron edits without elementary-step labels. This reduces reliance on costly mechanism annotations. It also provides explicit electron sources and destinations.

  2. Fig.s 1 - 2 are helpful for understanding the paper.

Critical weaknesses
  1. The experiments do not isolate the benefit of directional electron edits. MAELLE and MEGAN differ in architecture and training, as well as their edit operations. A controlled comparison is needed to show whether pairing electron sources with destinations explains the generalization gains.

  2. The mechanism analysis lacks a comparison with the OT-derived training targets. The 95.4% agreement concerns recovered persistent steps in trajectories reaching the correct product. The paper does not report how well OT alone matches the expert directions. This makes the additional mechanistic information learned by the model unclear.

Questions for authors
  1. What drives MAELLE’s better generalization to unseen chemistry? What evidence links this improvement to directional electron flow?

  2. What mechanistic information does MAELLE learn beyond the net changes between reactants and products? How can this be distinguished from the information already supplied by optimal transport?

  3. Which reaction types remain difficult for MAELLE, and why? Do these failures mainly reflect limited training data or limitations of the electron-edit representation?

3 comments
AuthorsOpenReview ↗

What mechanistic information does MAELLE learn beyond the net changes between reactants and products? How can this be distinguished from the information already supplied by optimal transport?

We thank reviewer 1dVe for the insightful feedback. We would gladly address your concerns below. Before we go into that, as reviewer L2ch also noted, the performance of NERF is indeed lower than expected, and we identified and corrected a bug in our evaluation script. First, we would like to provide an updated main results:

0. Updated main results

ModelUSPTO Top-1Top-3Top-5AbSynth Top-1Top-3Top-5
Molecular Transformer0.8860.9350.9420.3540.4190.441
Graph2SMILES0.9030.9400.9480.3950.4520.471
MEGAN0.8630.9240.9400.4570.5610.595
NERF0.8980.9130.9160.3950.4290.437
MAELLE0.8830.9340.9420.4820.5810.607

The corrected NERF performance is substantially closer to its originally reported results. Our main conclusion remains unchanged: MAELLE is competitive on USPTO-480K and achieves the highest AbSynth accuracy among the evaluated baselines.

1. What mechanistic information does MAELLE learn beyond the net changes between reactants and products? How can this be distinguished from the information already supplied by optimal transport?

To answer this question, we would like to add two additional extra rows to Table 2:

sample_modeGT bondscoveragebonds compareddirection agrees
correct1753412956 (73.9%)1295612297 (94.9%)
top11753414305 (81.6%)1430512933 (90.4%)
all1753417479 (99.7%)1747915589 (89.2%)
Graph-distance-based OT interpolation1753417534 (100.0%)1753415773 (90.0%)
random coupling (16 samples)1753417534 ± 0 (100.0 ± 0.0%)17534 ± 04857 ± 70 (27.7 ± 0.4%)

To just simply learn the net change between reactants and products, there are more than one way to assign the sources (electron sites that lose electrons) to sinks (electron sites that gain electrons) other than just optimal transport: One could simply assign the pairs randomly, which significantly changes the CTMC trajectories but endpoints remain the same.

In the table, since bot OT and random coupling are endpoint-consistent, they recover all persistent bond formation. But endpoint-consistent source-sink coupling doesn't mean the source-sink assignment agrees with the annotated mechanism in FlowER: the direction of the electron moves only agree 27.7% of the time, compared with 90.0% of the OT. And since MAELLE is trained on the trajectory interpolated from OT, it learns direction of the electron flow as well.

AuthorsOpenReview ↗

What drives MAELLE’s better generalization to unseen chemistry? What evidence links this improvement to directional electron flow?

2. What drives MAELLE’s better generalization to unseen chemistry? What evidence links this improvement to directional electron flow?

To answer this question, we trained two other versions of MAELLE, one with the random coupling above, and one without the FLOW move entirely:

ModelUSPTO-480K top-1USPTO-480K top-3USPTO-480K top-5ABSynth top-1ABSynth top-3ABSynth top-5
Base MAELLE0.8830.9340.9420.4820.5810.607
Random source-sink coupling0.8740.9300.9390.4750.5750.600
without FLOW move0.8750.9280.9360.4850.5790.603
MEGAN0.8630.9240.9400.4570.5610.595
NERF0.8980.9130.9160.3950.4290.437

Removing graph-distance-based OT coupling or the directional FLOW move leads to only modest changes in predictive accuracy. Notably, all three MAELLE variants outperform MEGAN and NERF across all Top-k metrics on AbSynth.

These results suggest that MAELLE's improved generalization to unseen chemistry is not primarily driven by directional electron-flow modeling, but may instead arise from its electron-occupation representation and discrete flow matching formulation.

Nevertheless, directional electron-flow modeling serves a distinct purpose: producing chemically interpretable CTMC trajectories as seen in the first table.

AuthorsOpenReview ↗

Which reaction types remain difficult for MAELLE, and why? Do these failures mainly reflect limited training data or limitations of the electron-edit representation?

3. Which reaction types remain difficult for MAELLE, and why? Do these failures mainly reflect limited training data or limitations of the electron-edit representation?

To answer this question and to explain why MAELLE falls short on the USPTO-480K benchmark, we hypothesize that the number of moves of MAELLE could be biased toward shorter trajectories.

Top-5 accuracy stratified by the number of electron moves derived from OT (best in bolds, second best in italics)

n_moves (from OT)# reactionsMAELLEG2SMTMEGANNERFMAELLE − G2S
227,5960.9670.9650.9610.9710.939+0.002
34,5090.9290.9410.9310.9210.891−0.012
44,4490.9210.9290.9200.9120.868−0.008
51,0930.7770.8680.8490.7810.695−0.091
61,6110.8800.9060.8930.8710.849−0.026
72920.7430.7880.7840.6540.695−0.045
8+3440.6080.7330.7010.5840.558−0.125

MAELLE performs particularly well on short trajectories. For reactions requiring two electron moves, it achieves 96.7% Top-5 accuracy, slightly exceeding Graph2SMILES (96.5%). However, its relative performance generally deteriorates as the required number of moves increases. For reactions requiring eight or more moves, MAELLE achieves 60.8% accuracy, compared with 73.3% for Graph2SMILES.

One possible explanation is that SMILES-based models generate products directly, without explicitly executing a sequence of electron moves or graph edits, thus less sensitive to the number of edits required to transform the reactants into the products. MEGAN and NERF also generally exhibit reduced accuracy as the number of required electron moves increases, correlated with the number of graph edits / electron changes needed to be made.

A second possible explanation is the imbalance in trajectory lengths within the training data, where short electron-edit trajectories predominate. This may bias MAELLE toward generating too few moves.

We compare the number of moves fired by MAELLE with the number required by the OT-derived reference, separately for Top-5 hits and misses (see Figure 7, Appendix M, page 27 in our revised submission). For Top-5 hits, the generated move counts are strongly concentrated near the reference counts. In contrast, Top-5 misses show a tendency toward two-move rollouts, even when the reference requires more moves.

We note that the frequency of MAELLE firing a move can be controlled via the Poisson temperature τ\tau, described in our Algorithm 3:

peflow/del=1−exp(−(λeflow+λedel)Δt/τ);paadd=1−exp(−λaaddΔt/τ),p^{flow/del}_e = 1 - exp(-(\lambda^{flow}_e + \lambda^{del}_e) \Delta t / \tau); \quad p^{add}_a = 1 - exp(-\lambda^{add}_a \Delta t / \tau), thus a lower temperature would promote the model to sample more moves. All experiments reported here use a fixed temperature of (τ=0.6\tau=0.6), however scanning different temperatures at inference could help MAELLE sample longer or shorter trajectories.

Overall, these results identify reactions requiring longer electron-edit trajectories as an important failure mode. The observed pattern is consistent with both training-data imbalance and the difficulty of executing longer sequences of edits, rather than demonstrating an intrinsic expressivity limitation of the electron-occupation representation.

We will include this stratified evaluation, together with a more detailed analysis of electron-move counts in USPTO-480K, in the revised manuscript.

Reviewer_L2chOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MAELLE, a discrete flow-matching framework for chemical reaction prediction based on graph-structured electron occupations. The method represents chemical transformations through directional electron-pair edits, including electron transfers, additions, and deletions. Optimal transport is used to construct electron-edit targets from reactant–product pairs without requiring annotated elementary reaction steps. A discrete flow-matching model then learns to generate stochastic electron-edit trajectories for product prediction. Experiments on USPTO-480K and AbSynth evaluate reaction prediction accuracy and out-of-distribution generalization. The authors also analyze the agreement between predicted electron transfers and expert-annotated reaction mechanisms. The results suggest that explicitly modeling electron redistribution may offer benefits for reaction prediction, particularly under distribution shifts.

Critical strengths
  1. The paper explores an interesting direction by explicitly modeling directional electron transfers for chemical reaction prediction. Representing reactions through electron-pair movements provides a chemically motivated alternative to conventional molecular graph editing and product generation.

  2. The proposed framework learns electron-edit trajectories from reactant–product pairs without requiring explicit elementary-step annotations. The inclusion of both reaction prediction experiments and comparisons with expert-annotated electron transfers provides useful evidence for evaluating the proposed representation.

Critical weaknesses

1. Limited empirical competitiveness on the standard USPTO benchmark. The experimental results do not convincingly demonstrate the advantages of the proposed method over existing reaction prediction approaches. On the standard USPTO-480K benchmark, MAELLE achieves 87.2% Top-1 accuracy, substantially lower than Graph2SMILES (90.3%) and the reported NERF baseline (90.7%). Its main empirical advantage is observed on AbSynth, where the improvements over the strongest competing baseline, MEGAN, are relatively modest (46.6% vs. 45.6% Top-1 accuracy). Given the additional complexity introduced by graph-structured electron occupations, optimal transport, and discrete flow matching, these results provide limited evidence that the proposed modeling framework offers substantial practical benefits over existing approaches.

2. Insufficiently convincing and standardized OOD generalization evaluation. A central motivation of this paper is to improve reaction prediction beyond the training distribution. However, the empirical evidence supporting this claim remains insufficient. The main OOD evaluation relies on transferring models trained on USPTO-480K to AbSynth, where multiple factors, including molecular complexity, reaction types, and dataset preprocessing, differ simultaneously. Consequently, it is difficult to determine whether the observed improvements genuinely arise from the proposed electron-flow representation. Although the authors additionally construct OOD-Mass and OOD-Ester splits, these settings examine relatively specific forms of distribution shift and may not adequately establish general OOD generalization. A more systematic and reproducible evaluation would be valuable, such as Tanimoto-similarity-based splits on USPTO, which would allow the authors to examine performance as test reactions become increasingly dissimilar to the training distribution. Without such evaluations, the claimed generalization advantages remain insufficiently established.

3. Incomplete and potentially unfair comparison with NERF. NERF is one of the most relevant baselines for this work, as it also explicitly models electron redistribution for reaction prediction. However, the paper does not report the performance of the original NERF on AbSynth, arguing that NERF distinguishes reactants from reagents. This justification is not sufficiently convincing, as differences in input representations should not automatically exclude an important baseline from evaluation. More concerningly, the authors instead compare against a modified NERF with its reactant–reagent segment embeddings removed, which causes a dramatic decrease in USPTO Top-1 accuracy from 91.4% to 68.2%. The resulting model achieves only 9.1% Top-1 accuracy on AbSynth. Such a substantial performance degradation raises concerns about whether the modified NERF constitutes a sufficiently competitive baseline. The authors should report the original NERF's performance on AbSynth where its required inputs can be provided, explicitly acknowledge differences in input information, and investigate a more appropriate adaptation under comparable input conditions. Without these comparisons, MAELLE's advantage over existing electron-redistribution models remains unclear.

4. Insufficient evidence that the generated trajectories represent meaningful reaction mechanisms. The paper emphasizes mechanistic reaction prediction through directional electron-flow modeling. However, the proposed electron-transfer trajectories are inferred from reactant–product endpoints using optimal transport rather than supervised by actual elementary reaction mechanisms. While the authors report 95.4% donor–acceptor agreement with expert annotations, this result is conditional on recovered persistent bond-forming steps in trajectories that reach the correct product. The corresponding coverage is only 73.3%, and the reported coverage of transient bond-forming steps is substantially lower. Consequently, these experiments mainly demonstrate consistency with certain product-defining electron transfers rather than the ability to recover chemically faithful reaction mechanisms. The paper would benefit from a more comprehensive evaluation of intermediate steps, transient bond changes, and the chemical validity of generated trajectories. Otherwise, the extent to which the proposed approach genuinely learns reaction mechanisms, rather than endpoint-consistent electron edits, remains unclear.

5. Insufficient ablation studies to establish the contributions of the proposed modeling components. The framework combines several methodological components, including graph-structured electron occupations, optimal-transport-based edit construction, directional electron-pair transitions, and discrete flow matching. However, the paper does not sufficiently isolate the contribution of each component to reaction prediction accuracy and OOD generalization. In particular, it remains unclear whether the observed performance improvements arise from explicitly modeling electron donor–acceptor relationships, from the graph-edit representation itself, or from the discrete flow-matching framework. More controlled experiments comparing directional electron-flow modeling with conventional electron redistribution or graph-edit formulations under comparable architectures and training settings would help establish the actual methodological contribution of the proposed approach.

Questions for authors
  1. Given that OOD generalization is a central motivation of this work, why do the authors primarily rely on AbSynth and specifically constructed OOD splits rather than more systematic similarity-based evaluations? Could the authors provide additional experiments using Tanimoto-similarity-based splits on USPTO to demonstrate whether MAELLE consistently generalizes better than existing methods as reaction similarity to the training data decreases?

  2. Why do the authors not compare MAELLE with the original NERF that explicitly distinguishes reactants and reagents on AbSynth?** NERF is one of the most relevant baselines for this work, and simply using reactant–reagent separation does not seem to be a sufficient justification for excluding it. Instead, the authors remove NERF's segment embeddings, causing its USPTO Top-1 accuracy to drop dramatically from 91.4% to 68.2%. Why not report the performance of the original NERF on AbSynth with reactants and reagents separated? Such a comparison would be important for establishing whether MAELLE genuinely outperforms existing electron-redistribution approaches.

7 comments
AuthorsOpenReview ↗

Response to reviewer L2ch: Correct NERF results under mixed and separate settings

Thank you for your constructive feedback. We address your concerns below and report additional experiments conducted in response to your review.

1. Updated results under mixed and separate reactant–reagent settings

You correctly noted that our previously reported NERF performance was unexpectedly low. We have identified and corrected a bug in our evaluation script, which substantially affected NERF and led to smaller changes in the results for MEGAN and MAELLE.

We first report the corrected results under the mixed reactant–reagent setting.

Top-k accuracy on USPTO-480K and AbSynth (mixed setting; best in bold, second best in italics)

ModelPrediction modalityUSPTO Top-1USPTO Top-3USPTO Top-5AbSynth Top-1AbSynth Top-3AbSynth Top-5
Molecular TransformerSMILES0.886*0.935*0.942*0.3540.4190.441
Graph2SMILESSMILES0.903*0.940*0.948*0.3950.4520.471
MEGANGraph edits0.863*0.924*0.940*0.4570.5610.595
NERFElectron difference0.8980.9130.9160.3950.4290.437
MAELLEElectron flow0.8830.9340.9420.4820.5810.607

* Results reported in the original publications.

The corrected NERF performance is substantially closer to its originally reported results. Our main conclusion remains unchanged: MAELLE is competitive on USPTO-480K and achieves the highest AbSynth accuracy among the evaluated baselines.

Comparison with the original NERF under separate inputs

Because the original NERF reports results using separately identified reactants and reagents, we additionally evaluate MAELLE and the other applicable baselines under this setting.

Top-k accuracy on USPTO-480K and AbSynth (separate setting; best in bold, second best in italics)

ModelPrediction modalityUSPTO Top-1USPTO Top-3USPTO Top-5AbSynth Top-1AbSynth Top-3AbSynth Top-5
Molecular TransformerSMILES0.904*0.946*0.953*0.3050.3620.382
MEGANGraph edits0.893*0.944*0.956*0.3710.4480.473
NERFElectron differences0.907*0.933*0.937*0.3250.3520.355
MAELLEElectron flow0.9010.9440.9500.4180.5040.526

* Results reported in the original publications.

Graph2SMILES is omitted because it does not support the separate-input setting.

Under separate inputs, MAELLE achieves 41.8% Top-1 accuracy on AbSynth, compared with 32.5% for NERF. Thus, MAELLE outperforms NERF on AbSynth under both mixed and separate reactant–reagent settings, addressing the concern that our original comparison disadvantaged NERF through input modification.

Interestingly, reactant–reagent separation generally improves USPTO-480K accuracy but reduces AbSynth accuracy for all evaluated models. We attribute this to differences in how reactants and reagents are annotated across the datasets.

In USPTO-480K, the distinction is based on whether a molecule contributes atoms to the recorded product. In contrast, AbSynth provides chemist-assigned reagent labels that do not necessarily follow this criterion. For example, Br2 may be classified as a reagent even when bromine atoms are incorporated into the product.

Consequently, a model trained with USPTO-style separation may encounter a shift in the meaning of these labels when evaluated on AbSynth. Moreover, product-derived reactant/reagent assignments cannot generally be reconstructed from unannotated inputs at inference time. The mixed setting avoids this annotation mismatch and provides a consistent input protocol for cross-dataset evaluation.

We will report the corrected mixed-setting results in the main text and include the separate-setting comparison in the appendix.

AuthorsOpenReview ↗

Response to reviewer L2ch: Systematic similarity-based OOD evaluation

2. Systematic similarity-based OOD evaluation

To address your concern about the generality of our OOD evaluation, we constructed a new USPTO-480K split designed to assess two distinct distribution shifts:

  • OOD-Reaction: reactions whose structural transformations are dissimilar to those in the training set.
  • OOD-Mass: reactions producing substantially larger molecules than those in the training distribution.

For OOD-Reaction, we construct a reaction-difference fingerprint by subtracting the reactant fingerprint from the product fingerprint. We then calculate Tanimoto similarity between these reaction fingerprints. Each reaction is characterized by its maximum similarity to the training set, allowing us to identify reactions involving transformations that are relatively uncommon in training.

The resulting dataset statistics are:

Dataset statistics

SplitNumber of reactionsMean MWMedian MWMW rangeMean heavy atomsMean max. similarity to trainMedian max. similarity to train
Train371,313341.4336.370.1–577.923.6––
Validation20,628341.1335.3117.2–577.723.50.8160.879
ID test20,628342.2337.3106.1–577.823.60.8200.897
OOD-Reaction18,808327.1318.4106.1–577.822.40.3480.353
OOD-Mass23,659658.9632.7577.9–1746.845.80.7500.778

The OOD-Reaction split has substantially lower similarity to the training distribution than the ID test set (median maximum similarity of 0.353 versus 0.897), while retaining a broadly comparable molecular-weight distribution. Conversely, OOD-Mass contains much larger products (mean molecular weight 658.9 versus 342.2 for ID), allowing us to examine a different axis of generalization.

We retrained and evaluated the models under the mixed reactant–reagent setting.

Top-k accuracy on the new splits (best in bold, second best in italics)

ModelID Top-1ID Top-3ID Top-5OOD-Reaction Top-1OOD-Reaction Top-3OOD-Reaction Top-5OOD-Mass Top-1OOD-Mass Top-3OOD-Mass Top-5
Molecular Transformer0.8910.9430.9530.6620.7900.8240.6840.7260.737
Graph2SMILES0.9070.9280.9300.6890.7400.7480.8540.8770.880
MEGAN0.8650.9340.9490.6470.8100.8590.8550.9290.945
NERF0.8880.9050.9060.6890.7340.7380.8440.8700.872
MAELLE0.8750.9300.9400.6980.8120.8350.8740.9240.934

In-distribution performance. The results remain consistent with the original USPTO-480K evaluation. SMILES-based models achieve the highest Top-1 accuracy, while MAELLE remains competitive with graph-editing approaches.

OOD-Reaction. MAELLE achieves the highest Top-1 (69.8%) and Top-3 (81.2%) accuracy, while MEGAN achieves the highest Top-5 accuracy (85.9%). MAELLE exceeds MEGAN by 5.1 percentage points at Top-1 and the next-best models, Graph2SMILES and NERF, by 0.9 points.

OOD-Mass. MAELLE achieves the highest Top-1 accuracy (87.4%), followed by MEGAN (85.5%) and Graph2SMILES (85.4%). MEGAN performs best at Top-3 and Top-5, while MAELLE ranks second with relatively small margins.

Across both OOD test sets, MAELLE outperforms MEGAN at Top-1, while the two are competitive when more predictions are considered. SMILES-based models (Graph2SMILES and Molecular Transformer) also generally fall short of the leading graph-based methods on these OOD splits.

Except for ID Top-1, MAELLE outperforms NERF across all evaluated settings and Top-k metrics.

Overall, these experiments strengthen our evidence for generalization across two distinct forms of distribution shift: reaction-transformation dissimilarity and increased molecular size. We will replace the previous OOD-Ester evaluation with these results and document the new split-construction procedure in the revised manuscript.

AuthorsOpenReview ↗

Response to reviewer L2ch: Ablation study

3. Ablation study

Reviewer L2ch also noted the lack of ablation studies in the original submission. We conducted additional experiments to assess the contributions of graph-distance-based optimal transport (OT) and directional electron-flow moves. Specifically, we address two questions.

How does MAELLE perform with random source–sink coupling or without the FLOW move?

We trained two additional variants of MAELLE: (1) Random coupling, which replaces graph-distance-based OT with random source–sink assignments while preserving the same reaction endpoints; and (2) Without FLOW, which removes the directional FLOW operation from the model.

Top-k accuracy on USPTO-480K and AbSynth (mixed setting; best in bold)

ModelUSPTO-480K top-1USPTO-480K top-3USPTO-480K top-5ABSynth top-1ABSynth top-3ABSynth top-5
Base MAELLE0.8830.9340.9420.4820.5810.607
Random source-sink coupling0.8740.9300.9390.4750.5750.600
without FLOW move0.8750.9280.9360.4850.5790.603
MEGAN0.8630.9240.9400.4570.5610.595
NERF0.8980.9130.9160.3950.4290.437

Removing either graph-distance-based OT coupling or the directional FLOW move leads to only modest changes in predictive accuracy. Notably, all three MAELLE variants outperform MEGAN and NERF across all Top-k metrics on AbSynth. These results suggest that MAELLE's improved generalization to unseen chemistry is not primarily driven by directional electron-flow modeling, but may instead arise from the electron-occupation representation and discrete flow matching formulation.

So what's the point of the graph-distance-based OT framework?

The reactant and product electron occupations determine which sites lose or gain electrons, but do not uniquely specify how these source and sink sites should be paired. Multiple endpoint-consistent couplings can therefore yield the same final product while implying different electron-transfer directions.

To assess whether graph-distance-based OT provides chemically informative source–sink assignments, we compare its directional agreement with expert-annotated persistent bond-forming steps against random endpoint-consistent coupling.

sample_modeGT bondscoveragebonds compareddirection agrees
correct1753412956 (73.9%)1295612297 (94.9%)
top11753414305 (81.6%)1430512933 (90.4%)
all1753417479 (99.7%)1747915589 (89.2%)
Graph-distance-based OT interpolation1753417534 (100.0%)1753415773 (90.0%)
random coupling (16 samples)1753417534 ± 0 (100.0 ± 0.0%)17534 ± 04857 ± 70 (27.7 ± 0.4%)

In the table, since bot OT and random coupling are endpoint-consistent, they recover all persistent bond formation. But endpoint-consistent source-sink coupling doesn't mean the source-sink assignment agrees with the annotated mechanism in FlowER: the direction of the electron moves only agree 27.7% of the time, compared with 90.0% of the OT. And since MAELLE is trained on the trajectory interpolated from OT, it learns direction of the electron flow as well.

We shall add the two additional rows of "Graph-distance-based OT interpolation" and "random coupling" into table 2 of our manuscript.

As to why graph-distance based OT can induce the direction of electron flow, including a concrete reaction example, reviewer L2ch is more than welcome to visit our response for reviewer vUW8.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.