acceptodds
Under review as a conference paper at ICLR 2027

Neural Petri flows for chemical reactions

Abstract

Petri nets, in which transitions move tokens between places, have been used to describe chemical processes such as reactions. Petri net terms map well to chemistry: Places are the bonds between atoms and the free valence of each atom, a token is a unit of bond order, a transition forms or breaks a bond, the conserved quantities are the valence budgets of the atoms, and the enabling rule, that a transition needs tokens on its input places, is the valence rule. These semantics are not guaranteed by learned models of reactions or neural networks that are built on Petri nets, which use the net as a scaffold for message passing. Here, we ask what architecture remains a Petri net for every value of its weights. We find the answer in the theory, where all semantics of a net share the firing form , locality, as enabling reads only the inputs of a transition, and the enabling rule, and we prove that conservation forces the firing form and that non-negativity forces the enabling rule on local rate laws. What this leaves free is the rate law, which is the propensity of each transition to fire. We introduce Neural Petri Flow, which learns this rate law, or a readout for classification, and hard-wires the rest as parameter-free layers. On what we denote a valence net, atom mapping, reaction classification, and forward prediction become three tasks on one firing vector, which is an unordered list of bond changes. Without training, the minimum firing vector maps 88.8 % of the curated Golden set against 85.6 % for RXNMapper, and 88.7 against 77.9 % of the enzymatic reactions of EnzymeMap. On USPTO-480K, NPF trained on these firing vectors predicts 87.7 % of the products and 67.4 % when trained on a 1 % subset of the training reactions. EC numbers of ECREACT are predicted at the third level for 90.2 % of reactions, 5.6 points ahead of the best published method. With electrons as tokens, the same token game predicts 90.5 % of the elementary steps of FlowER first, ahead of the published baseline, and every top-1 prediction is a valid molecule without a filter.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
petri nets, chemistry, biology, resource flows

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_U23iOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds a Petri net formulation around chemical constraints on valence, hydrogen counts, and charge. It adapts this unified formulation to atom mapping, forward reaction prediction, reaction classification, and reaction mechanism prediction, yielding Neural Petri Flow (NPF). The Petri net constraints improve data efficiency in low-resource settings and the chemical structural validity of predictions.

Critical strengths
  1. Defined by invariant valence budgets and a predefined set of operations, the NPF provides a unified, interpretable formulation that can be incorporated into different downstream chemistry tasks, with competitive performance in training-free and low-resource settings.
  2. The paper systematically and fairly compares NPF and its variants with other baselines across multiple well-defined downstream tasks. It also evaluates the chemical validity of the generated outputs, demonstrating that NPF's constraints improve prediction validity without significantly affecting downstream performance.
Critical weaknesses
  1. As a network that enforces chemical priors, NPF could be used as a standalone component independent of the trainable architecture. The paper also evaluates combinations of NPF with different readout networks. This raises the question of whether NPF can be combined directly with current state-of-the-art models in the field to improve both predictive performance and validity. NPF's task-specific designs also complicate the comparison: the trainable backbones can themselves affect performance and output validity through their own inductive biases (e.g., molecular editing rather than AR for reaction prediction may itself constrain the structural search space). The ablations do not fully isolate the benefits of NPF under these conditions.
  2. Related to this concern, a substantial portion of the benchmarks in the main text are already close to saturation, including atom mapping, forward reaction prediction, and reaction classification on USPTO-50K. There is consequently little separation between NPF and the baselines in accuracy and output validity, which further calls into question the benefits attributable to NPF itself. Earlier baselines have, in fact, already achieved performance close to saturation on these tasks.
Questions for authors

Following the concerns above, it is possible to view NPF as a standalone network encoding prior rules that could restrict the allowed molecular edits at each step or be incorporated into a backbone as an embedding. Have the authors explored integrating it directly with existing backbones, rather than designing their own backbone / readout strategy from-scratch?

2 comments

Thank you for your feedback. Even though you declare low confidence, we would like to answer your questions.

W1. Combination with existing models

We absolutely agree that the backbone matters. The firing form, the mask, and the state equation do not have trainable parameters, which means that a stronger network can replace our rate law and keep the valence guarantee, whose proof does not use the rate law (see Proposition 4). We have not tried this with our forward model. Our controls change one component and keep the backbone and its size (see Tables 14 and 22). In forward prediction, the backbone without the mask is as accurate on Schneider50k, but breaks valence budget in 0.02 % of the predictions, while a one-shot model does so in 2.79 % (see Table 18). Editing alone already keeps 99.98 % of the predictions within the budgets and the mask makes this a guarantee. So what the net gives to the forward model is a guarantee and not accuracy (see our answer to reviewer TgvJ, W1e).

W2. Saturated benchmarks

We agree for USPTO-480K, where NPF is not ahead at top-1 (see Table 4). Also for organic atom mapping and Schneider50k with all labels. What we believe separates us there is efficiency and methodology. For example, the mapper needs no training and NPF reaches 67.4 % top-1 with 1 % of the USPTO-480K training data (see Table 16). However, harder benchmarks separate the models. NPF reaches 90.2 vs 84.6 % at EC level 3 on ECREACT (see Table 3), of which 1.6 points come from the form of the readout (see Table 14), and 70.2 vs 59.3 % at the full EC number on CARE against the best method that only uses the reaction (see Table 3). Furthermore, it maps 88.7 vs 77.9 % of the EnzymeMap reactions (see Table 11)., recovers 94.5 vs 92.5 % of FlowER pathways (see Table 5), and reaches 81.2 vs 65.0% for DRFP with 250 labels on Schneider50k (Table 13). We credit these leads to NPF as a whole (beyond the controls mentioned in W1). On FlowER, every single top-1 prediction of our net is a valid molecule, and every prediction conserves atoms, protons, and electrons. The sequence models only do this for 17 to 33% of the predictions (see Table 21).

Q1. Integration with existing backbones

That’s only in classification, where an existing backbone (Chemprop CGR D-MPNN) reads the maps of our mapper. With 1k labels on Schneider50k, its best setting reaches 89.2 vs 95.1 % for our classifier on the firing vector (see Table 13). For forward prediction, the mask and the order of the edits must be chosen jointly. MEGAN orders its targets with all new bonds first, and the valence rule blocks every order like this in 59.7 % of our correct USPTO-480K predictions (with two or more changes, see Appendix D.5). When trained without the mask on a fixed order with all new bonds first first, we lose 32 points once the mask is turned on (see Table 23). This means that a mask in MEGAN would require a different order.

Reviewer_U23iOpenReview ↗

Comment by U23i

Thanks authors for their response. The new controlled backbone experiments clarify where NPF provides benefits, and the additional comparisons with models of the same budget demonstrate the value of its constraints, or priors. In my view, NPF's self-consistent and interpretable prior design could be explored further as a source of priors for existing backbones, or as a first principle for designing domain-specific backbones.

Based on these considerations, I have increased my rating to 3. My confidence score primarily reflects my familiarity with this field and background (which appears to be how confidence is defined in the ICLR system), rather than my personal confidence in the rating.

Reviewer_kTyKOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Neural Petri Flow (NPF), a framework for modeling chemical reactions that incorporates explicit structural constraints into its architecture. Its central idea is to learn which chemical changes to make while enforcing conservation and enabling rules independently of the neural network’s weights. The framework represents reactions as firing vectors describing bond or electron changes and applies this representation to atom mapping, reaction classification, product prediction, and elementary reaction-step prediction. The experiments demonstrate competitive performance across these tasks, with particularly strong results on enzyme classification and elementary-step prediction, although several existing methods achieve higher accuracy on forward product prediction.

Critical strengths

Competitive performance across several tasks: The method achieves strong results on multiple chemical benchmarks, particularly enzyme classification and elementary reaction-step prediction. These results suggest that incorporating explicit structural constraints need not come at the expense of predictive accuracy.

Informative ablation studies: The authors investigate sequential prediction, the enabling mask, classification readouts, atom-mapping targets, and single-electron moves. These experiments help distinguish components that improve accuracy from those that provide structural guarantees. In particular, sequential prediction substantially outperforms one-shot prediction, while the enabling mask prevents violations of the encoded valence rule with little change in accuracy.

Critical weaknesses

Incomplete coverage of strong forward-prediction baselines. Table 3 omits Chemformer [1] and R-SMILES [2], despite citing these methods in the related work. On the USPTO-MIT mixed benchmark, R-SMILES reports 91.0% top-1 accuracy, and its comparison table lists 91.3% for Chemformer, compared with 87.7% for NPF. Differences in pretraining, augmentation, inference budgets, or evaluation protocols may affect comparability. Nevertheless, these results should be included with appropriate qualifications, or their exclusion should be explained. This would provide a more complete picture of NPF’s predictive performance.

The valence guarantee is weaker than a guarantee of molecular validity. For the bond-based model, Proposition 4 establishes reachability under augmented valence budgets that include a half-token tolerance. This does not ensure that every decoded output is a valid molecule: Table 17 reports predictions that fail RDKit sanitization and cases in which undecodable fragments are dropped. The paper should distinguish more consistently between satisfying the encoded valence budgets and producing chemically valid molecular structures. This concern does not apply in the same way to the electron model, which achieves 100% molecular validity on the reported benchmark.

Fixed valence budgets restrict the chemistry that the model can represent. Each atom receives a budget determined by its initial state, and bond changes cannot increase this budget. Consequently, some real transformations are excluded. Appendix B explicitly notes that a sulfide cannot become a sulfone in one reaction under the proposed representation. This illustrates a limitation of the hypothesis space rather than an error that additional training could resolve. The implications of these restrictions for reaction coverage and practical applicability deserve clearer discussion.

[1] Irwin, R., Dimitriadis, S., He, J., and Bjerrum, E. J. Chemformer: a pre-trained transformer for computational chemistry. Machine Learning: Science and Technology, 3(1), 015022, 2022.

[2] Zhong, Z., Song, J., Feng, Z., Liu, T., Jia, L., Yao, S., Wu, M., Hou, T., and Song, M. Root-aligned SMILES: A Tight Representation for Chemical Reaction Prediction. arXiv:2203.11444.

Questions for authors

Why are Chemformer [1] and R-SMILES [2] omitted from Table 3? Could the authors include their results or explain which differences in experimental conditions prevent a meaningful comparison?

Are the published baseline scores based on the same product-matching criteria used for NPF, particularly regarding additional predicted molecules? For the FlowER comparison, how do beam search and sampling differ in inference cost, and could these differences contribute to the reported gains?

Could the authors clarify the scope of the bond-based model’s validity guarantee and quantify how often its fixed budgets or decoding rules prevent reconstruction of the recorded product?

3 comments

(Part 1 of 2)

Thank you for the review and the very precise questions. In the revision, we now keep budget validity and molecular validity apart, report a Kekulé variant of the net, add Chemformer, R-SMILES, and the augmented transformer to Table 4 (with the parameters and the pretraining), name the scoring rule and the setting in Table 4, state what the budgets and the decoding rules exclude in Section 7 and Appendix B, and correct how Tables 5 and 21 describe the FlowER baselines. Several points overlap with Reviewer TgvJ. We refer to those answers for details. The table numbers are those of the revised PDF, where changes made only for your review are blue, and other changes magenta.

W1. Forward-prediction baselines

Indeed, the additional baselines belong in Table 4 (Table 3 before the revision). We did limit Table 3 to the baselines from Nguyen et al., 2026. We have now added both Chemformer and R-SMILES, plus the augmented transformer by Tetko et al, 2020, in a block of their own (as they’re pretrained or augment the SMILES at test time). In th mixed setting, R-SMILES reaches 91.0 % at top-1 and 96.8 % at top-5. It's pretrained on molecules of the USPTO-FULL data set, uses 5 root-aligned SMILES per reaction during training and at test time, and it scores an exact match of the product SMILES. Chemformer-Large reaches 91.3 and 93.7 % with significantly more parameters (230M compared to 45M and 25M for R-SMILES and augmented transformer, respectively. NPF has 9.8M.) and is pretrained on the approx. 100 million ZINC molecules (they don't state a matching rule). The augmented transformer reaches 90.6 and 96.1 % with a beam of 10 on a test set that was augmented 100 times. Compared to these numbers, with one beam of 5 and the major-product rule, the closest we come to an exact match (see Q2), NPF reaches 87.2 and 93.5 % over three seed (Table 4 and the answer to reviewer TgvJ, Q2). Note that we have not tried pretraining or augmentation with NPF, but this might be interesting future work.

W2. Valence budgets and molecular validity

Agreed, and the revision now keeps the two apart (Sections 1 and 6, as well as the caption of Table 4). We provide the details in the answer W1d to reviewer TgvJ. In short:

(a) Budget validity is a theorem. For every value of the weights, each marking visited by the token game keeps every atom within the valence budget plus half a token (Proposition 4). The half token only matters in a game that forms or changes an aromatic bond. We investigated this by analysing the reactions in USPTO-480K, and no test reaction forms or changes an aromatic bond. See more details in answer W1d to reviewer TgvJ.

(b) Molecular validity is measured. RDKit accepts every molecule of 99.9 % of the greedy predictions on USPTO-480K (this is all but 41, see Table 19 in the revision). The 41 are also the predictions with a dropped fragment, so the two cases you mention are the same. In the caption of Table 19, we now clearly say that dropping a fragment is a filter.

(c) Addition of Kekulé variant. We ran an experiment on the Kekulé variant where bond orders are whole, so no half token is required. We added a rule that a cationic atom takes no spare token after we saw failures on cationic atoms. With this rule, RDKit accepted every final marking we tested. On Schneider50k, this variant performs the same as the aromatic net that we retrained for comparison (93.4 vs 93.0 % top-1, see Table 20). We report this as a measurement rather than a theorem.

As you note, a valid molecule is returned by the electron net for every top-1 prediction on FlowER (see Table 21).

W3. Fixed valence budgets

We agree. The budget of an atom is fixed by its state in the precursors plus one capacity per element. So this is a limit that isn't removed by training. We ran an analysis with the following results. Of all 40,000 USPTO-480K test reactions, the budgets rule out 131 at most (see Appendix B), these count as errors. 73 form an exocyclic double bond at an aromatic carbon, 41 oxidise a sulfide to a sulfone (plus 1 to a sulfoximine), and 16 give new bonds to a metal (10 of them to magnesium). The decoding rules miss at most 1.1 % and at least 98.5 % are reproduced within the budgets (see Q3). For the excluded reactions, NPF returns a different product within the budgets. We now report these limits in Appendix B and the coverage in Section 7.

Looking at the sulfur case: Sulfur keeps two spare tokens, so in one reaction a sulfide can become a sulfoxide and a sulfoxide a sulfone. What is excluded are the two steps at once. Of the test reactions with a usable recorded map, all 175 that oxidise a sulfide to a sulfoxide and the 17 that oxidise a sulfoxide to a sulfone are within the budgets. As the capacities are one number per element, they could be widened but at the price of a looser guarantee.

(Part 2 of 2)

Q1. Chemformer and R-SMILES

See W1.

Q2a and Q2b. Product matching and FlowER inference

(a) Matching rule. No, the rules differ (also among the baselines). Graph2SMILES, MAELLE, R-SMILES, and the augmented transformer score an exact match of the product (the molecular transformer does the samea s far as we could assert, but it is not stated in the paper). Chemformer-Large and GTPN do not state a rule, MEGAN compares the longest SMILES, and NERF evaluates a subgraph match. We detail this in our answer to reviewer TgvJ (Q2). In Table 4 we now name the rules and the settings. NPF returns every molecule of the final marking, while the ground truth omits by-products. 90.4 % of NPF's predictions (the correct ones) also return by-products, so an exact match of the whole outputisn'tt possible. As NERF does, and we state in the rule of Appendix B, our rule ignores molecules that are beyond the ground truth product. Under the major-product rule (we added to Table 4) we now only compare the largest predicted and ground truth molecules. For the 96.6 % of test reactions that only have one product, this is an exact match of the product. Under this major rule, NPF reaches 87.17 vs 87.42 % at top-1 against the rule from Appendix B, while 87.10 % pass both. On FlowER, NPF is scored as in the evaluation script of Joung et al., so the rule is the same as for the baselines.

(b) Beam search against sampling. NPF costs approximately an order of magnitude less. With a beam of 5 it took 2.3 to 2.4 hours for all 162,002 test steps on a single GPU (a 4070 TI). Compared to that, FlowER integrates an ODE for each of 32 to 64 samples per step. The paper mentions about 28 hours for its test set on on an RTX A5000, but without the number of samples (Joung et al., 2025, in SI S4).

The decoding can contribute through the ranking to the gains. FlowER ranks samples by frequency and drops non-valid ones, while NPF, molecular transformer, and Graph2SMILES rank beams by likelihood. Beam vs beam, NPF leads against the sequence models at step top-1 by 1.4 to 3.1 points (see Table 5). Against FlowER-large, NPF is 0.7 points ahead (this is slightly ahead and we do not give the entire credit to the enabling rule). NPF rank each product by the summed probability of the beam paths that end in it (this is what the frequencies of FlowER estimate from samples). What NPF adds is a Lewis structure by construction. That is, every end marking of the electron net, at every rank, is a Lewis structure, and every top-1 prediction a valid molecule. FlowER drops around 5 % of its samples as invalid, and 21 to 30 % of the 30 best predictions of the sequence models are not valid SMILES.

Q3. Scope of the guarantee and coverage

The guarantee is about the markings visited by the token game and not about the decoded molecules. For every value of the weights, no atom exceeds its budget in the precursors by more than 1/2 token (Proposition 4). Whether the decoded product is a valid molecule is then measured (see W2). We ran an analysis to explore the scope. On the 40,000 test reactions of USPTO-480K, at least 39,420 have a firing vector that is within the budgets that decodes to the ground truth product. Of the remaining 580, the decoding rules miss at most 443, the budgets rule out at most 131, and RDKit does not parse 6. We now report these counts in Appendix B of the revision. The 96.6 % also in Appendix B is lower because 2.0 % of the test reactions do not have a usable ground truth map.

Thank you for the detailed response. The revisions substantially address my concerns. I maintain my score.

Reviewer_TgvJOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper models a single reaction as a Petri net, the "valence net":

  • Places: bond places hold bond orders, and slack places hold each atom's free valence.
  • Transitions: form or break bonds.
  • P-invariants: the per-atom valence budgets.
  • Enabling rule: the valence rule.

Only the rate law, or a classification readout, is learned. The firing form and the enabling rule are hard-wired.

Atom mapping, reaction/EC classification and forward prediction become three tasks on one firing vector, an unordered set of bond changes:

  1. Mapping finds the minimum firing vector by branch and bound, without learning.
  2. Classification reads a label from the firing vector, or from a difference-of-embeddings readout.
  3. Forward prediction plays a masked "token game" that fires one bond change at a time.

An electron-token variant handles mechanism steps.

Results are strong on several benchmarks:

  • Golden and EnzymeMap atom mapping
  • EC prediction on ECREACT (EC3: 90.2 vs 84.6) and CARE
  • FlowER steps and pathways

Results on USPTO-480K forward prediction are competitive but not state of the art.

Critical strengths
  1. Unifying formulation with strong results. Casting mapping, classification and forward prediction on one firing vector is elegant. It gives state-of-the-art EC prediction (ECREACT EC3: 90.2 vs 84.6) and mechanism prediction on FlowER, and strong low-data forward prediction (67.4% top-1 with 1% of USPTO).
  2. Practical training-free mapper. The minimum-firing-vector mapper is competitive with learned mappers. Its tied optima serve as forward-model targets that outperform the dataset's own atom maps.
Critical weaknesses
  1. The Petri formalism's necessity is not shown, and the theoretical claims are overstated.
    • Each component has an established counterpart. The valence net is the Dugundji–Ugi bond–electron matrix, the firing vector is a set of graph edits (a CGR), the enabling rule is valence masking, and the token game is autoregressive graph-edit decoding.
    • No Petri-net theory beyond linear algebra is used. The paper relies only on the state equation and P-invariants.
    • Proposition 2 does not cover the actual model. It forces the enabling rule only for local rate laws, while the chemistry model's rate law is non-local.
    • The validity guarantee is weaker than stated. It needs half a token of tolerance per atom and data-derived decoding rules. Some USPTO predictions still fail RDKit sanitisation (Table 17), and removing the mask leaves accuracy unchanged.
    • The gains come from elsewhere. The real gains appear to come from sequential decoding and order-free targets, not from the Petri structure.
  2. The net-to-task correspondence is not explained clearly.
    • Net variants are not tied to tasks. Several variants are used (an aromatic net with tolerance, a Kekulé net, and the mapper's net with C–H places) without stating which applies where.
    • A marking does not determine a set of molecules. Hydrogens, lone pairs and charges share the slack places, and decoding relies on heuristic rules.
    • The classification readout does not use the net at all.
    • Suggested fix. One table giving, for each task, the net variant, its places and transitions, and the decoding procedure would substantially improve clarity.
Questions for authors
  1. What does the Petri formulation provide that a bond–electron-matrix or graph-edit formulation does not?
  2. Are the USPTO-480K baselines scored with the same correctness rule as NPF? Can you report results over several seeds?
  3. What is the effect of count-equivalent beam merging, as an ablation?
6 comments

We have to split our answer into multiple parts as we try to address your comments and questions backed by data and new analyses.

(Part 1 of 4)

Thank you for your constructive comments and questions. We have addressed them in the following way: (1) We have narrowed the claims, (2) added the table you suggested, and (3) run additional checks, including three seeds on the full USPTO data (Q2). We uploaded a new version of the manuscript where changes for your review are highlighted in magenta. The table numbers in this text relate to the revised manuscript.

W1. Necessity of the Petri formalism

W1a. Counterparts

We also answer Q1 here, as it is related.

We agree that every part has a counterpart, and we do not claim the parts as new. Our contribution is the combination of the parts and their application. Furthermore, we also agree that the net is not required to build a forward model. A graph-edit decoder that masks by valence at every step and trains on an order that the mask allows plays the same game and shares the guarantee (see Section 1 of the revision). However, the net is how we derive this decoder and it gives us three things. (1) It gives us the mask and the order. A decoder which changes a single bond at a time stays within the valence budgets at every step, for every value of its weights, only if every step that it allows is enabled in the net (see the new Proposition 2b in the revision and W1c). Per Lemma 2, breaking first is enabled for every product within the budgets. For the forward model, both hold with half a token of tolerance (see W1d). A bond-electron matrix and a condensed graph of reaction give the overall change of a reaction, as the firing vector does, but no order. GTPN and MEGAN, which are the graph-edit decoders in Table 4, check no valence at the step level. MAELLE checks electrons at the source but not the room on the receiving atom. Our model loses 20 to 32 points under the mask without a fitting training order (see (a) below). (2) The minimum firing vector is the training target, and all its enabled orders reach the same marking. This means that NPF does not need external atom maps or edit order. On Schneider50k, the same vector feeds the classifier that leads with few labels (see W1e). (3) The same rate law and token game run on a net of electrons, where they compare well with the bond-electron-matrix model on the same task (see (b) below).

(a) Valence mask and order of changes must be chosen together. When generating molecules a valence mask is safe as bonds are only added. However, in a reaction, a new bond may need a break first. Therefore, a mask can block a step of the target sequence. In fact, GTPN states that it checks no valence at its step for exactly this reason and removes invalid products after decoding. This lifts the top-1 on USPTO from 73.3 to 83.2 % (see Do et al., 2019, Table 3). While a filter can only remove candidates, the mask keeps every candidate within the budgets (we do not claim a gain in accuracy because of this, see W1d). In short, Lemma 2 removes the obstacle that is named by Do et al. We ran two new analyses/experiments to investigate the behaviour of our model. (1) On the USPTO-480K test set. In 59.7 % of the correctly predicted reactions with two or more bond changes, the valence rule blocks every order that puts all new bonds first. In contrast, MEGAN gives forward targets such an order and does not mask actions by valence (see Sacha et al., 2021), this means that the intermediates would break a valence budget. This is of course fine for MEGAN, which only needs a valid product, but a valence mask would block its targets. (2) On Schneider50k (new Table 23). We trained NPF with teacher forcing on one fixed order of the changes without the mask, then switched the mask on at test time. Trained with every gain of bond order first, that means with all new bonds before any break, the performance decreases form 93.5 to 61.6 % at top-1. When trained on the order of atom indices, it decreases from 93.7 to 73.7 %. Of course, training with the mask fails too because a blocked target step gets zero-probability. When trained on breaks first, the order in Lemma 2, it keeps its 93.0 %.

(Part 2 of 4)

W1a. Counterparts (continued)

(b) The same game runs on other nets. With electrons as tokens, the octet rule is the enabling rule. FlowER is the BEM model for mechanism steps on the same data, which conserves electrons and learned the valence rules without enforcing theme. According to Joung et al., 2025, around 5 % of its samples are invalid. On the other hand, the electron net returns a valid molecule for every top-1 prediction (see Table 21). It also predicts 90.47 % of the steps first, compared to 89.74 % for FlowER-large with around 2x the parameters (see Table 5). Furthermore, FlowER ranks samples, while we rank a beam, this means that neither comparison is like for like, and we do not credit the lead to the enabling rule alone.

W1b. Only linear algebra

We agree that the theory is elementary. The state equation and the P-invariants are linear algebra, neither depending on the order in which transitions fire. However, the enabling rule does, because it has to hold at every step. So what goes beyond linear algebra is whether an enabled order exists. According to Murata, 1989 (their Theorem 16), an order exists in an acyclic net whenever the end marking is non-negative. In the reaction nets, the transitions of a firing vector that changes each bond at most once form such an acyclic subnet. Lemma 2 is this theorem extended with an explicit order, break firsts. The revision now credits Murata in Section 3 and after the lemma.

W1c. Proposition 2

This is correct and we say so in Section 4.1. However, we cited Proposition 3, which was a LaTeX user error on our side. In the initial submission, we left the argument for the token game at the end of Appendix A.2. Proposition 2 covers local rate laws, which the metabolic nets use. The token game does not need locality and fires one transition at a time on a pure net where no transition puts tokens back on a place it takes them from. So firing a transition that is not enabled makes an input place negative, and a rate law that keeps every visited marking non-negative has to give it rate zero. We now present this in Proposition 2b in Appendix A.2 and reference it in Section 4.1 where Proposition 2 doesn't apply. The abstract and the contribution now name both cases.

W1d. Validity guarantee

We agree. In the revision we now keep two claims apart (in Sections 1 and 6, and in the caption of Table 4):

(a) Budget validity is a theorem. For every value of the weights, each marking the token game visits keeps every atom within its valence budget plus half a token (Proposition 4). The claim is about markings and not about decoded molecules. The 1/2 token only matters in a game that creates or changes an aromatic bond. In the test set of USPTO-480K, no recorded reaction creates or changes one, and 2 out of 39,994 top-1 predictions fire on an aromatic bond. On Schneider50k, 35 of the 46,672 reactions require the 1/2 token. Decoding without it changes no scores of the three Schneider50k models that we trained. The revision now reports these in Appendices A.4 and D.5.

(b) Molecular validity is measured. RDKit parses all but 41 greedily predicted molecules of USPTO-480K (Table 19 in the revision). In those, the decoder leaves out a fragment that fails sanitisation. This is of course a filter and we now say so in the revision. We further investigated this and found that the top-1 predictions of the beam hold 42 of these (one of them uses the half token). The caption of Table 19 falsely blamed the half token, which we corrected in the revision. In short, the decoding rules are heuristics, and the revision now calls them that.

A note: In Appendix B, we described a Kekulé net variant without a result. Based on the above observations, we decided to run it. Its enabling rule needs no half token, but it caps the bonds of an atom at a maximum per element (read from the training data). On Schneider50k, it is level with the aromatic net retrained in the same way (93.4 vs 93.0 %). On USPTO-480K, 0.1 % of the markings failed at cationic atoms, so we added a rule that forbids a cationic atom to take a spare token. With this rule added, RDKit rejects none of the 288,000 markings from untrained weights and none of 1.4 million samples of three trained Schneider50k models on three data sets, including USPTO-480K (see Table 20). We report this as a measurement and not a theorem.

(Part 3 of 4)

W1d. Validity guarantee (continued)

Regarding the mask, it does not add accuracy, but it does prevent predictions that break a valence budget by more than 1/2 token (for every value of the weights, Proposition 4). By Lemma 2, it removes no product within the budgets (and we measure no significant loss). We ran the following analysis. Without the mask, 319 of 40,000 test reactions in USPTO-480K have a candidate among the top 5 that breaks a valence budget by more than 1/2 token, and the greedy prediction breaks one for 5 (Appendix D.5, Table 19). Schneider50k models trained without the mask on only 500 reactions have such a candidate in the top 5 for 14.6 % vs 2.3 % with all training reactions. This means that without the mask, budget validity is a learned regularity that weakens with very few data.

W1e. Source of the gains

For the forward accuracy, we agree. Appendix D.5 says the same. The gain comes from firing one change at a time, with 93.3 vs 79.7 % for a one-shot model of same size (see Table 18). The mask adds none (see Table 18). We also now say in Section 6.2 that the order-free loss does add to gain either. For models that we trained for this response on Schneider50k, each on one target vector per reaction, teacher forcing on a fixed order of the changes reaches 93.0 to 93.7 % top-1 vs 92.9 % for the order-free loss (Table 23). However, what the net gives the forward model is the guarantee at no significant cost in accuracy, and training without external atom maps or a choice of edit order.

For Schneider50k, the firing vector adds accuracy in classification where labels are scarce. With only 250 labelled reactions, the classifier on the firing vector reaches 81.2 %, while DRFP reaches 65.0 %. On the same maps, a generic head reaches 49.0 %, and the best CGR D-MPNN (balanced reaction mode, matched steps), reaches 40.5 % (see Table 13). The readout classifier is linked to the net by its form (see W2c). We also ran a short additional experiment on training set size and distribution. On ECREACT, this form adds 1.6 points at EC level 3. The lead is 3.1 points with a tenth of the training reactions and 3.8 points for EC numbers with at most seven training reactions, so it is more data efficient at EC level 3.

W2. Net-to-task correspondence

W2a. Net variants

We added the table you suggested as Table 1 to Section 5. It shows the net, places, transitions, enabling rule, learned part, output with its encoding rules, and the table where results of it are presented. In short, forward prediction uses the aromatic valence net (with 1/2 token of tolerance in the enabling rule). Mapping uses the same net with one extra bond place per hydrogen on carbon and the firing-vector classifier reads the firings of the mapper on the bonds between heavy atoms. For the mechanism, we use the electron net on Kekulé structures. The readout classifier uses no net. Note that the Kekulé variant we introduced in Appendix B previously had no results. We now report it in Table 20 (see W1d).

W2b. Markings and molecules

This is correct. Section 1 called a marking a set of molecules. It now correctly states that a marking fixes the bonds between heavy atoms and the free valence of every atom. Hydrogens and charges are read from the free valence by the five rules from Appendix B. On USPTO-480K, these rules miss 1.1 % of the test reactions max. The electron net differs in that it has no such step. Every atom (including hydrogens) has a place for its own electrons, and the end marking is a Lewis structure.

Comments on the paper

1 comment
AuthorsOpenReview ↗

A summary of our rebuttal and associated revision

  • Clarity (TgvJ). We added a new table (Table 1) listing the net variants, per task, with their places, transitions, enabling rules, learned parts, and decoding part.
  • Validity (TgvJ, kTyK). We now differentiate between budget and molecular validity, the former is a theorem (Proposition 4), while the latter is measured. Decoding rules are now correctly called heuristics.
  • Source of gains (TgvJ, U23i). Forward accuracy comes from firing one change at a time (93.3 vs 79.7 % one-shot), and not form the order-free loss or the mask. The net gives a guarantee at no significant accuracy cost. We added controls that show that the firing vector helps classification with few labels (81.2 vs 65.0% baseline with 250 labelled reactions). Furthermore, we showed that the form of the readout adds significantly to the accuracy.
  • Baselines and scoring (kTyK, TgvJ). We now include Chemformer, R-SMILES, and the augmented transformer in Table 4, and name the relevant scoring rule and setting for the methods.
  • Narrowed claims (TgvJ). We now state clearly that every component has a counterpart and that their combination is novel. Furthermore, we showed that the net is required to build a forward model, enabling us to derive the valence mask and the training order together (Lemma 2, added Proposition 2b). In addition, this gives us training targets without a requirement for external atom maps.
  • Limits (kTyK). We measured that at least 98.5 % of the USPTO-480K test reactions can be reproduced within the budgets (Section 7, Appendix B).
  • New ablations (general). We showed that mask and edit order must be chosen together (Table 23), and that beam merging adds almost no accuracy (Appendix D.5).

According to reveiwer TgvJ, we addressed most of their concerns, and reviewer U23i comments that the new controls clarify where NPF is beneficial.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.