{"id":"5d424294-bbb8-4ed7-8453-14b4776019c6","arxiv_id":"2411.17629","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A graph encoder that aligns reactant and product atoms, plus a reaction-center-aware decoder, improves prediction of reaction conditions, yield, and selectivity across several benchmarks.","lead":"This paper presents RAlign, a machine learning model that represents chemical reactions by matching atoms between reactants and products and focusing on the reaction center. Tests on reaction condition, yield, and selectivity benchmarks show it outperforms most existing reaction representation methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Privileged reaction-center input at inference may explain part of the reported gains; the paper's ablations do not test this on the headline task.","rationale":"The reader's weakest_assumption identified exactly the same load-bearing concern: the RC-aware decoder is given reaction centers at inference while baselines are not, and the robustness analysis does not isolate this effect. My stress-test confirms this is the most consequential issue. The paper discloses the atom-mapping requirement in its Limitations section, which is good, but the central empirical claim is still conditional on showing that the gains do not come primarily from this extra input. The existing ablation (Table 8) removes the RC-aware decoder on two tasks and shows only a modest drop, but the headline top-1 result is on USPTO CONDITION, where no such ablation is reported. A dedicated ablation on that task, plus a baseline given equivalent reaction-center information, would settle the question. Because the reader already conditioned the verdict on addressable issues and the concern does not by itself invalidate the method, I do not move the verdict; it remains CONDITIONAL in substance, which in the schema corresponds to UNCHANGED from the reader's CONDITIONAL verdict.","tokens_in":17,"tokens_out":4143,"duration_ms":126695,"concrete_test":"Run a version of RAlign with all RC-aware heads replaced by vanilla cross-attention on USPTO CONDITION, matching Table 8's ablation but on the task where the headline gain is claimed. In parallel, give the strongest baseline (Parrot-LM-E) reaction-center tokens (e.g., append atom-mapped center annotation to the SMILES input) and retrain. If RAlign's top-1 accuracy drops substantially or Parrot-LM-E with centers closes the gap, the reported improvement is largely due to the extra reaction-center input rather than to the learned representation; if neither happens, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that RAlign's performance gains come from a learned reaction representation rather than from privileged reaction-center input. In Section 4.3, the decoder receives Vrc, the reaction-center atom set, at inference; Vrc is computed from atom-mapping (Section 3.1) and identifies exactly which atoms change bonds, change hydrogen counts, or belong to the leaving group. No baseline receives this information: Parrot-LM-E, T5Chem, GCNN, and fingerprint methods see only SMILES or molecular graphs. The RC-aware heads are therefore an extra input that is strongly correlated with reaction type. The robustness check in Appendix A.3 correlates accuracy with RXNMapper confidence but does not remove or perturb Vrc, and the ablation in Table 8 removes the RC-aware decoder only on USPTO 500MT and Buchwald-Hartwig, not on the headline USPTO CONDITION task. If the top-1 gain of 34.30% versus 27.42% partly reflects the center mask, the central claim that the architecture learns better representations is overstated. This is not a claim of label leakage: atom-mapping is a legitimate input, but the comparison to baselines that lack it is not apples-to-apples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces RAlign, a supervised graph-based model for learning chemical reaction representations from molecular graphs together with atom-mapping-derived atomic correspondences between reactants and products. The architecture has three components: an atom-aligned encoder (Sec. 4.1) that fuses reactant and product node features for corresponding atoms through an information fusion layer; a condition-incorporating adapter (Sec. 4.2) that lets reaction conditions enter the encoder via multi-head attention; and reaction-center-aware (RC-aware) decoders (Sec. 4.3) in which half of the cross-attention heads attend only to reaction-center atoms Vrc, defined in Sec. 3.1 from bond changes, hydrogen-count changes, one-hop neighbors, and leaving groups. RAlign is evaluated on reaction condition prediction (USPTO CONDITION, Tables 1-2), condition generation (USPTO 500MT, Table 3), yield prediction (Buchwald-Hartwig, Tables 4-5), and selectivity prediction (C-H functionalization and thiol addition, Tables 6-7). Headline results include top-1 condition-prediction accuracy of 34.30% versus 27.42% for the strongest baseline Parrot-LM E, and best or near-best results on most other tasks, with third-place finishes on the Buchwald-Hartwig random splits and on thiol addition.","tokens_in":23068,"tokens_out":17535,"duration_ms":145410,"significance":"This paper contributes a clean, well-motivated architectural inductive bias: explicit reactant-product atomic alignment (Eq. 3) and reaction-center-constrained attention (Eq. 6), packaged as a modular encoder-decoder with an interchangeable condition adapter. The empirical validation is broad, spanning four downstream tasks including out-of-sample splits on Buchwald-Hartwig, and the authors disclose the regimes where the method does not win (third on Buchwald-Hartwig random splits and on thiol addition). Credit should also go to the ablations in Table 8, the robustness analysis in Appendix A.3, the detailed implementation appendices, and the plain-spoken limitation statement about atom-mapping dependence. If the reported numbers hold, RAlign is a competitive, pretraining-free alternative to large pretrained reaction models on several tasks. However, the two reservations below, namely the unablated reaction-center input on the headline condition-prediction task and the absence of variance or significance statistics in Tables 1-3, mean that the strength of the central claim is not yet fully established.","major_comments":[{"comment":"The fairness of the headline condition-prediction comparison is not established because the RC-aware decoder receives the ground-truth reaction-center atom set Vrc at inference (computed from atom mapping, Sec. 3.1), an input that none of the baselines (Parrot-LM E, T5Chem, GCNN, FPRCR) receive. This is not label leakage, but it is an extra, task-correlated input channel: the case study in Sec. 5.5 states that the RC-aware heads 'provid[e] the model with information about the reaction type,' and reaction type correlates with catalyst, solvent, and reagent choices in exactly the USPTO CONDITION task. The ablation in Table 8 removes the RC-aware decoder only on USPTO 500MT and Buchwald-Hartwig, not on USPTO CONDITION, so the 6.88-point top-1 margin over Parrot-LM E has no control. The robustness analysis in Appendix A.3 correlates accuracy with RXNMapper confidence but never removes or perturbs Vrc. Please (a) report the '- Reaction-Center-Aware Decoders' variant on USPTO CONDITION; (b) add a control where Vrc is replaced by a same-size random atom set, to separate the effect of the center information from decoder capacity; and (c) add a mapping-perturbation study (e.g., shuffled atom alignments) to Appendix A.3. Until these are provided, part of the reported gain may be attributable to the privileged center input rather than to the learned representation.","section":"Sec. 4.3, Tables 1-2, Table 8, Appendix A.3"},{"comment":"The statistical support for the 'markedly outperforms' claim is incomplete. Tables 1-3 report single-run accuracies with no variance or significance information, so the 6.88-point and 1.55-point top-1 gaps on USPTO CONDITION and USPTO 500MT cannot be assessed for stability. In contrast, Tables 5-7 report mean plus or minus standard deviation over ten splits, and on the Buchwald-Hartwig random splits RAlign is third (MAE 3.6331 +/- 0.1259 versus T5Chem 3.5059 +/- 0.1562; R2 0.9581 +/- 0.0049 versus 0.9662 +/- 0.0034). The text's assertion that the model is 'not significantly outperformed' by the pretrained baselines is based on overlapping standard deviations rather than a paired test, and the claim that the MAE gap is 'less than 0.1' (Sec. 5.2) is inaccurate for the strongest baseline, since the gap to T5Chem is 0.127. Please report multiple seeds with standard deviation or a significance test for the headline Tables 1-3, and run a paired test or effect-size analysis for the Buchwald-Hartwig random-split and thiol-addition comparisons.","section":"Tables 1-3 and 5-7"}],"minor_comments":[{"comment":"The left-hand side of the reactant MPNN equation lists the intermediate node features as {hR(k)_i, hR(k)_i, ..., hR(k)_i} with a repeated index; this should range over vR_1 through vR_n as in Eq. (2).","section":"Sec. 4.1, Eq. (3)"},{"comment":"The square-root-of-d normalization appears to be applied outside already-normalized softmax weights alpha1_l and alpha2_l; if scaled dot-product attention is intended, the factor should be inside the exponent. Please correct the formula for reproducibility.","section":"Sec. 4.3, Eq. (6)"},{"comment":"The statement that the MAE gap to the strongest baseline is 'less than 0.1' holds only versus YieldBert; the gap to T5Chem is 3.6331 minus 3.5059, which is 0.1272, so the sentence should be corrected.","section":"Sec. 5.2"},{"comment":"The margins '9.03% and 11.93%' reported against 'the strongest baseline that did not utilize pretraining' do not tie out to a single model in Table 3: 9.03 is the top-1 margin over Reagent Transformer, which is itself initialized from a pretrained Molecular Transformer checkpoint, while the top-10 margin over T5Chem-from-scratch is 11.97. Please clarify which baseline is intended and reconcile the numbers.","section":"Sec. 5.1"},{"comment":"The contribution bullet claiming 'the first graph-based architecture specifically tailored for chemical reactions' and the first modeling of atomic correspondence overstates novelty, given GCNN (Maser et al., 2021), the condensed-graph-of-reaction encoding used by Chemprop, and prior atom-mapping-aware architectures such as Retroformer (Wan et al., 2022). Please soften these claims and position the contribution as a general reaction-representation backbone.","section":"Sec. 1 and Sec. 2"},{"comment":"For USPTO CONDITION, please state explicitly whether the train, validation, and test split is identical to the one used by Wang et al. (2023) for Parrot-LM E, since Table 2 compares directly against the reported Parrot-LM E numbers; the RXNMapper re-annotation should be described as performed within each split to avoid any ambiguity.","section":"Sec. 5.1 and Appendix B"},{"comment":"For USPTO 500MT the Pearson coefficients for top-5 and top-10 accuracy are 0.367 and 0.415, which are moderate rather than negligible, so the conclusion that 'there is no strong correlation' should be tempered or the analysis extended with a perturbation test.","section":"Appendix A.3"},{"comment":"Because the USPTO 500MT generation labels are restructured with the Type I/II/III ordering and GCNN is reformulated as multilabel classification over a predefined library, the USPTO 500MT comparison mixes task definitions across baselines; please state explicitly that the structured output ordering is an implementation choice and discuss whether it favors RAlign relative to the SMILES-generation baselines.","section":"Sec. 5.1 and Appendix D.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript compares directly against Parrot-LM E's published number; given that the paper re-processes the dataset with RXNMapper, an audit of split identity would be prudent before publication. The code is promised only upon acceptance; for an empirical architecture paper, releasing the code and data-processing scripts would materially raise confidence in the comparisons. The novelty framing in the contributions section is stronger than the related-work section supports, since the RC-aware decoder is adapted from Retroformer's local-global decoder and the authors' own UAlign already uses alignment ideas for retrosynthesis; this is a presentation issue rather than a correctness issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RAlign is a serious attempt at reaction representation learning that actually uses the structure of a reaction. Atom-mapped reactant/product pairs are fused in every encoder layer, and half the attention heads are restricted to reaction centers. That combination is new, and the evaluation is broad: condition prediction and generation, yield, and selectivity. On C-H functionalization and the out-of-sample Buchwald-Hartwig splits, RAlign is consistently strong and beats pretrained baselines like T5Chem and RXNFP. The authors disclose the main dependence (atom mapping) and include a robustness check against RXNMapper confidence. This is not hype.\n\nThe soft spots are real but not fatal. The biggest one is the reaction-center input. In the decoder, Vrc—the set of atoms that change bonds or hydrogen counts, plus leaving groups—is computed from atom mapping and supplied to the model at inference. No baseline gets this. On USPTO CONDITION, the headline task, this is an extra input that correlates with reaction type, so the 34.3% vs 27.4% top-1 gain is not a clean apples-to-apples comparison. The ablation of the RC-aware decoder is only run on USPTO 500MT and Buchwald-Hartwig, where the effect is small (roughly 0.3 top-1 and 0.003 R2). That suggests the reaction-center mask is not the whole story, but the task where the gain is largest is exactly where the control is missing. The robustness analysis correlates accuracy with mapping confidence; it does not perturb or remove Vrc. A simple control on USPTO CONDITION without the RC-aware heads would settle this.\n\nOther issues are smaller. Tables 1–3 have no variance or significance estimates, and there are random splits available. On the Buchwald-Hartwig random splits RAlign is third (MAE 3.63 vs T5Chem 3.51), and on Test4 its R2 (0.532) is below T5Chem (0.605). The abstract says 'markedly outperforms,' which is stronger than the full set of tables. The citation of prior work is fair, though the introduction's 'first graph-based architecture tailored for chemical reactions' is overstated given Retroformer's local-global decoder.\n\nThe math is not circular: atom mapping comes from input molecules, not from labels, so the gains are not label leakage. The architecture is specified in enough detail to reproduce from the appendix, but code is not released. This is an important detail for the field.\n\nVerdict: deserves referee time. I would send it to review, with a major-revision request for the missing ablation on the headline task, uncertainty estimates, and code or a strong reason for its absence. If the RC-aware decoder accounts for most of the USPTO CONDITION gain, the paper still has value—it becomes an input-augmentation method, and the claim needs to be scaled back accordingly.","headline":"Solid applied paper with a real architectural idea; the headline gain is plausible but the reaction-center mask is an extra input baselines lack, and the missing ablation on the main task is the one thing a referee should push on.","tokens_in":23555,"tokens_out":3120,"would_cite":true,"duration_ms":27197,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RAlign learns a chemical reaction representation by aligning each product atom with its reactant counterpart, then reports accuracy gains over existing reaction encoders on condition, yield, and selectivity tasks.","keywords":["chemical reaction representation learning","atom mapping","reactant-product alignment","reaction center","reaction condition prediction","reaction yield prediction","reaction selectivity prediction","graph neural network"],"falsifier":"Perturb the atom mappings on the USPTO CONDITION test set by randomly swapping which product atoms are paired with reactant atoms, or replace the mapping tool with another one, and retrain; if top-1 condition accuracy stays roughly the same, the alignment and reaction-center inputs are not what drives the reported gains.","tokens_in":22606,"feed_emoji":"🧪","tokens_out":5841,"duration_ms":50074,"temperature":0.7,"pith_summary":"RAlign is a graph-based encoder-decoder for chemical reactions that uses atom-level correspondence between reactants and products as its core inductive bias, rather than treating a reaction as a bag of molecules or a SMILES string. The paper argues that fusing features of corresponding atoms lets the model expose which bonds break and form, and that a reaction-center-aware decoder focuses attention on the atoms that actually change. On the USPTO CONDITION benchmark, RAlign reaches 34.30% top-1 condition-combination accuracy versus 27.42% for the strongest baseline, and it posts the best or near-best results on most yield and selectivity splits without large-scale reaction pretraining. The intended payoff is a reusable reaction representation backbone that can absorb reaction conditions in different formats and transfer across synthesis-related tasks.","feed_headline":"Aligned atoms lift reaction-condition prediction to 34.3%","feed_subtitle":"A single graph encoder beats pretrained sequence baselines on condition, yield, and selectivity tasks.","key_machinery":"The Atom Aligned Encoder is the load-bearing component: it runs two message-passing networks for reactants and products and, after each layer, concatenates the intermediate features of every atom pair matched by atom mapping and passes them through a feedforward network, so each side's representation is informed by its counterpart's local context. A condition adapter uses multi-head attention to inject encoded reaction conditions into node features at every layer, making the encoder modular across datasets whose conditions are reagents, temperatures, or experimental protocols. The RC-aware decoder splits its cross-attention heads: half attend normally and half attend only to reaction-center atoms, defined as atoms whose bonds change or hydrogen counts change, their one-hop neighbors, and leaving-group atoms. The reaction-center heads give the model a direct focus on the functional groups that determine the reaction outcome.","core_discovery":"The paper's central claim is that a chemical reaction representation is substantially improved when the encoder is built around the known atomic correspondence between reactants and products. RAlign maps each product atom to its reactant counterpart, runs separate message-passing layers on the two molecular graphs, fuses the paired atom features at every layer, and keeps leaving-group atoms as special nodes. The decoder then mixes standard cross-attention heads with heads that can only attend to reaction centers, defined as atoms whose bonds or hydrogen counts change plus their one-hop neighbors and leaving groups. On the USPTO CONDITION task this yields 34.30% top-1 accuracy, an absolute gain of 6.88 points over the Parrot-LM E baseline, with similar gains in components such as solvents; on Buchwald-Hartwig yield and C-H functionalization selectivity, RAlign achieves the best R2 among compared methods on most splits. The authors interpret these results as evidence that modeling bond changes and focusing on reaction centers are stronger inductive biases than fingerprints, pretrained SMILES encoders, or simple aggregation of per-molecule features.","pith_inferences":["An unstated but direct consequence is that atom alignment could become a standard preprocessing step for graph-based reaction encoders, replacing implicit feature aggregation with explicit correspondence between the two molecular sides.","Because the reaction centers are computed from atom mapping, a fair comparison with baselines would require giving those baselines the same reaction-center information; without that ablation, part of the reported gap may reflect an extra input rather than a purely better learned representation.","Since the encoder is agnostic to how node and edge features are computed, it could be combined with 3D conformer or quantum-chemical descriptors for selectivity tasks where current small-data performance lags fingerprint methods.","The condition adapter points toward a natural extension: encoding experimental procedures written in natural language and injecting them through the same attention-based adapter, which would broaden the model beyond reagent and temperature inputs."],"forward_implications":["Reaction condition combination prediction on USPTO CONDITION improves to 34.30% top-1 overall accuracy, with per-component gains in catalyst, solvent, and reagent prediction; the same architecture also improves de novo condition generation on USPTO 500MT over non-pretrained baselines.","On Buchwald-Hartwig yield prediction, RAlign achieves the best R2 on all four out-of-sample ligand splits and competitive random-split results without large-scale reaction pretraining.","On C-H functionalization selectivity, RAlign attains the best MAE, RMSE, and R2 among compared methods; on the small thiol addition selectivity dataset it outperforms other deep learning models but trails hand-crafted fingerprints.","Ablations removing the atom alignment fusion or replacing RC-aware attention with vanilla cross-attention reduce performance on both tested tasks, indicating that both components contribute to the reported gains.","The condition adapter allows pretrained molecular condition encoders to be plugged in, so the architecture can use existing molecular models rather than requiring reaction-specific pretraining."],"supporting_citations":[{"why":"Supplies the atom-mapping tool used to create the reactant-product correspondence that the Atom Aligned Encoder is built on.","marker":"Schwaller et al. 2021a"},{"why":"Provides the USPTO CONDITION dataset and the Parrot-LM E baseline against which the headline condition-prediction gain is measured.","marker":"Wang et al. 2023"},{"why":"T5Chem is the principal pretrained sequence baseline across condition generation, yield prediction, and selectivity prediction.","marker":"Lu & Zhang 2022"},{"why":"DRFP is the differential-fingerprint baseline on yield and selectivity, and it supplies the Buchwald-Hartwig data splits used by RAlign.","marker":"Probst et al. 2022"},{"why":"Provides the Buchwald-Hartwig cross-coupling dataset with its out-of-sample ligand splits used for yield prediction.","marker":"Ahneman et al. 2018"},{"why":"GCNN is the graph-based baseline for reaction condition prediction and generation.","marker":"Maser et al. 2021"},{"why":"Retroformer supplies the local-global decoder structure and the reaction-center definitions that the RC-aware cross-attention mechanism adapts.","marker":"Wan et al. 2022"},{"why":"FPRCR is the fingerprint-based baseline for fixed-composition reaction condition prediction.","marker":"Gao et al. 2018"}],"fun_headline_variants":["Aligned atoms lift reaction-condition prediction to 34.3%","RAlign graph encoder beats pretrained baselines on reaction tasks","Aligned atoms and reaction centers achieve 34.3% on condition","Reactant-product alignment boosts top-1 accuracy to 34.3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model is given reaction centers computed from atom mappings at every inference, so its reported advantage assumes reliable atom mapping is available for every reaction and that supplying this reaction-center information is a fair, transferable input.","fun_headline_variants_meta":{"raw":{"variants":["Aligned atoms lift reaction-condition prediction to 34.3%","RAlign graph encoder beats pretrained baselines on reaction tasks","Aligned atoms and reaction centers achieve 34.3% on condition","Reactant-product alignment boosts top-1 accuracy to 34.3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001186,"raw_usage":{"total_tokens":4904,"prompt_tokens":959,"completion_tokens":3945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":3868}},"tokens_in":575,"tokens_out":3945,"duration_ms":55706,"temperature":1.0,"reasoning_tokens":3868,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:54:14.811897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Perturb the atom mappings on the USPTO CONDITION test set by randomly swapping which product atoms are paired with reactant atoms, or replace the mapping tool with another one, and retrain; if top-1 condition accuracy stays roughly the same, the alignment and reaction-center inputs are not what drives the reported gains.","supporting_citations":[{"cited_title":"Generic interpretable reaction condition predictions with open reaction condition datasets and unsupervised learning of reaction center","cited_arxiv_id":null,"evidence_quote":"Provides the USPTO CONDITION dataset and the Parrot-LM E baseline against which the headline condition-prediction gain is measured."},{"cited_title":"Reaction classification and yield prediction using the differential reaction fingerprint drfp","cited_arxiv_id":null,"evidence_quote":"DRFP is the differential-fingerprint baseline on yield and selectivity, and it supplies the Buchwald-Hartwig data splits used by RAlign."},{"cited_title":"T., Estrada, J","cited_arxiv_id":null,"evidence_quote":"Provides the Buchwald-Hartwig cross-coupling dataset with its out-of-sample ligand splits used for yield prediction."},{"cited_title":"R., Cui, A","cited_arxiv_id":null,"evidence_quote":"GCNN is the graph-based baseline for reaction condition prediction and generation."},{"cited_title":"Retroformer: Pushing the limits of end-to-end retrosynthesis transformer","cited_arxiv_id":null,"evidence_quote":"Retroformer supplies the local-global decoder structure and the reaction-center definitions that the RC-aware cross-attention mechanism adapts."}],"review_version":1}