What Does Overall Column Have to Do with the LvG Identification Task?

In a chemical reaction, a leaving group (LvG) is an atomic group. It is present in the reactants but absent in the products. Leaving groups are classified into different types based on their structures.

Leaving group identification is an atom-level multi-classification task. It determines if an atom belongs to a leaving group. It also identifies the type of leaving group.

The Overall column records the classification accuracy for all atoms, while the LvG column excludes non-leaving group atoms when calculating accuracy. The Overall column indicates the model's ability to locate leaving groups, whereas the LvG column assesses its ability to distinguish between different types of leaving groups.

For example:

Table 3: Leaving group (LvG) identification results of Molecular Graph and Reaction Graph representations, with Overall and LvG atom-specific evaluation.

As shown in the table above, in the molecular graph, the Overall column shows relatively low accuracy, suggesting that the model mislocates leaving groups. In contrast, the Reaction Graph not only locates leaving groups correctly but also classifies them accurately.

What Was The Purpose and Procedure of Filtering of Data

Purpose

The purpose of filtering is to enhance the quality of the dataset by eliminating erroneous or incomplete records from the reaction database. We extract reaction task datasets from real-world reaction databases. These databases often contain errors and missing labels. Without filtering, these issues can negatively impact model performance.

Procedure

The cleaning steps starts with parsing each reaction with chemical toolkit (e.g. RDKit) and discarding unrecognized reactions. Then, reactions which lack task-related labels (e.g. reaction conditions) are removed. More details on the data processing workflow and results can be found in Sec.F.

Thank you for acknowledging our novelty and the structure of the paper cohesive. We also appreciate your constructive suggestions.

Why did we include both the schematics of other existing methods and the schematics of our method in Fig.1?

The main purpose of Fig. 1 is to guide readers through the transition from existing approaches to our novel contribution. This helps readers efficiently compare methods and grasp the unique aspects of our approach.

We also include a revised schematic of our Reaction Graph Pipeline, which can be found in Sec. H.1.

How to look for the "Attention weight" and what does "Mismatch" imply in Fig.3?

In Figure 3, the node colors in the schematics below "Set2Set Attention" represent Attention Weights, corresponding to the colors in the Colorbar.

"Mismatch" indicates that the atom with the highest Attention Weight is not the same as the reaction center atom (i.e. the atom where the reaction takes place). This situation occurs in the molecular graph, indicating that the model using the molecular graph has incorrectly understood the mechanism of the chemical reaction.

We also added the revised figure to avoid the color conflict between atomic types and the Attention Weights in Sec.4.1 and Sec.H.1.

Sentences and Formulations

We have addressed these issues in the revision version.

Why not use XYZ Coordinate

Directly using XYZ coordinates is not rotational and transcational invariant. The property of the chemical reaction is independent of the pose of the molecule. If we directly use the XYZ coordinates, the chemical reaction representation will change with their pose, increasing the chanllenge for neural networks to recognizing reactions.

To verify the above statements, we conduct the following experiment: we directly use XYZ coordinates as an additional node feature, and compare its performance with that of our Reaction Graph on USPTO-Condition dataset. The result is as shown below:

Table 1: Influence of XYZ coordinates on Reaction Graph.

The results show that our method effectively provides 3D prior, while the XYZ coordinates is difficult for the model to understand and utilize, resulting in drop in performance. We provide specific implementations and result analyses in the Sec.G.5.1.

Difference between our Reaction Graph’s Angular Edge and the supernode in Functional Groups or Pseudoatoms/Neural Atoms.

Motivation

Functional Groups[1] or Pseudoatoms/Neural Atoms[2] aims to facilitate long-range message passing. Our Angular Edge is designed to incorporate 3D information for structure modeling purposes.

Methodology

Functional Groups or Pseudoatoms/Neural Atoms shorten message passing path by reducing the size of the molecular graph. Our Angular Edge constructs triangles in molecular graph to convey bond angle feature.

Experiment

To verify these distinctions, we set the lengths of the Angular Edges to zero to eliminate the 3D information while still connecting related atoms. The results are as follows:

Table 2: Impact of connection and length (for 3D modeling) on Angular Edge. Angular Edge is mainly used for providing 3D infromation, instead of long-range message passing.

The result shows that the Angular Edge without 3D prior did not have a significant impact on performance, while Angular Edge with edge length significantly improved performance. This indicates that the role of the Angular Edge in providing messages is greater than its role in passing messages.

[1] Hajiabolhassan, Hossein, et al. "FunQG: Molecular representation learning via quotient graphs." Journal of chemical information and modeling 63.11 (2023): 3275-3287.

[2] Li, Xuan, et al. "Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel." The Twelfth International Conference on Learning Representations. 2024.

Dear Reviewer L6NH,

Thank you very much for your time and effort in reviewing our submission. We have tried our best to address the concerns and problems mentioned. We look forward to your feedback and further discussion. Feel free to let us know if there is anything unclear. We are happy to provide additional information. Thank you once again.

Best wishes,

Paper 4277 Authors

Dear Reviewer L6NH,

Thank you very much for your efforts in reviewing our paper. The discussion phase is nearing its end, and we are looking forward to receiving your feedback on our responses. Should you have any further concerns, we are glad to provide additional information. Thank you once again.

Best regards,

Paper 4277 Authors

Thank you for acknowledging our paper well-written and our approach clear and reasonable. We also appreciate your constructive suggestions.

Innovative Modeling of the Interaction between Reactant and Product

When employing AI techniques to solve chemistry problems, it involves two key aspects: molecular structure modeling and molecule interaction modeling. Existing methods typically focus on individual molecule modeling, which can address a few challenges such as molecule property, function, and attribute prediction. However, when it comes to chemical reactions, molecule interaction modeling at the atom level is required. Due to the lack of interaction modeling, existing methods have to first employ individual molecule modeling for reactants and products and then concatenate the global features for reaction reasoning, where the relation between reactants and products are not well processed. In contrast, our Reaction Graph fills the gap and can capture the interactions between molecules at the atomic level.

Incorporating 3D Information in Reactions vs. Incorporating 3D Information Molecules

3D information is crucial in modeling molecular structures, yet its application in molecular interactions (reactions) remains underexplored. Our Reaction Graph focuses on modeling the interactions between reactants and products. It innovatively integrates 3D information with reaction edges to model structural transformations at the atomic level. Additionally, it incorporates angular edges to represent bond angle features, which address the rotational degrees of freedom. These aspects are treated distinctly in structural and interaction modeling. This innovative approach positions Reaction Graph as a pioneering 3D method specifically tailored for the molecular interaction modeling of chemical reactions.

Computational Time Analysis for Reaction Graph and Other Models

Our method

1. Theoretical Analysis

   We define all graph-based representations of chemical reactions as $G(V, E)$, where $V$ represents the set of atoms and $E$ represents the set of bonds. The computational complexity of the Reaction Graph depends on the number of atoms $|V|$, the number of bonds $|E|$, the maximum degree $D$ of the atoms, and the specific implementation of the graph.

   In Reaction Graph, a reactant atom can match with at most one product atom, so the number of reaction edges does not exceed half the total number of atoms, $|V|/2$. Furthermore, as defined for angular edges in Fig. 1, the maximum number of angular edges is $|V| D^2$.

   Thus, the computational complexity contributed by bond edges is $|E|$, by reaction edges is approximately $|V|/2$, and by angular edges is around $|V| D^2$. Consequently, the total computational complexity of Reaction Graph can be approximated as $|E| + |V|/2 + |V| D^2$.

2. Experimental Results

   We assess the impact of reaction edges and angular edges in the Reaction Graph on model inference speed as the number of atoms $|V|$ increases. The findings are presented in rows 1-3 of the table below. The time cost associated with reaction edges is lower than that caused by angular edges.

Compared with Other Methods

1. Theoretical Analysis

   • Graph-based Method

     When the molecular graph is represented as $G(V,E)$, $\vert V \vert$ indicates the total number of atoms in reactants and products, and $\vert E \vert$ represents the total count of bonds. The computation cost for the molecular graph is $\vert E \vert$. The computation cost of Rxn Hypergraph is at least $\vert E \vert + \vert V \vert$. The edge count in Condensed Graph of Reactions (CGR) is roughly $\vert E \vert /2$. However, when handling the same amount of information, the dimensions of the fully connected layers in CGR far exceed those in the Reaction Graph and Rxn Hypergraph. This leads to greater computational overhead.

   • SMILES-based Method

     SMILES-based methods are generally based on Transformer. The complexity of Transformer is $\vert L \vert ^2$, where $\vert L \vert$ is the length of the input sequence. As a result, the computation cost of SMILES-based methods are roughly $\vert L \vert ^2$. Since there is no direct conversion between the length of SMILES strings $\vert L \vert$ and the number of nodes $\vert V \vert$ and edges $\vert E \vert$ in molecular graphs, we directly measured their computational efficiency through experiments.

2. Experimental Results

   • Compared with graph-based methods

     We compare the inference time of Reaction Graph and other graph-based methods. The results are shown in rows 4-6 of the table below. Due to incorporating additional reaction information and 3D structures, our method requires slightly more time than molecular graphs and Rxn Hypergraph representations. However, it delivers significant performance improvement, justifying the additional time cost.

     Table 2: Model inference time for different graph representations under varying numbers of graph nodes. Time units are milliseconds (ms).

     Representation	Reaction Information	3D Structure	2500	3000	3500	4000	4500
     Reaction Graph	√	√	37.28	43.77	50.57	58.03	65.45
     w/o Reaction Edge	×	√	35.36	40.92	48.68	53.30	58.79
     w/o Angular Edge	√	×	22.51	25.96	29.82	34.21	38.77
     Molecular Graph	×	×	18.59	21.93	25.39	28.30	31.48
     Rxn Hypergraph	×	×	31.18	37.28	43.31	48.53	54.06
     Condensed Graph of Reactions	√	×	44.55	52.48	60.75	66.32	75.34

   • Compared with SMILES-based methods

     We compare the inference time of Reaction Graph and SMILES-based methods. The results are shown in the table below, where Reaction Graph, as a graph-based model, exhibits the shortest inference time.

     Table 3: Inference time comparison for Reaction Graph and other methods.

     Methods	Representation	Time
     T5Chem	SMILES	3min 19s
     RXNFP	SMILES	23min 43s
     Reaction Graph (ours)	Reaction Graph	2min 36s

The details of the cost analysis can be found in Sec.E.

Is it possible to model the interactions between reactants and products by adding another hypernode in the Rxn Hypergraph

We can model the interactions between reactants and products by adding another hypernode in the Rxn Hypergraph. Specifically, we can connect the hypernodes of reactants and products with an additional hypernode.

However, this method still exhibits limitations compared to our Reaction Graph.

1. Lack of Atomic Mapping. To effectively track the transformation of substructures in a reaction, GNNs need to establish a mapping that identifies the correspondence between the substructures in reactants and products. However, the hypernode method falls short in directly representing this mapping, posing challenges in accurately expressing the interactions between reactants and products.

2. Limited Capability for Reaction Modeling . When a node has too many neighbors, it may lead to an information loss during message aggregation in GNNs (referred to as the Over-squashing Issue). This additional hypernode also faces the similar Over-squashing problem. Because the additional Hypernodes serve as a bridge for interactions between reactants and products, it usually has massive numbers of neighbors. Since it is difficult for a single node to carry so much information, the additional hypernode method may have a limited capability for reaction modeling, especially when many large reactants and products involved.

We conduct experiments on various tasks using the hypernode method. The results are presented as below:

• Reaction Condition Prediction:

  Table 4: Influence of the additional hypernode on the performance of Rxn Hypergraph in condition prediction.

  Method	USPTO-Condition	Pistachio-Condition
  Rxn Hypergraph w/o additional Hypernode	0.213	0.288
  Rxn Hypergraph w/ additional Hypernode	0.211	0.289
  Reaction Graph (ours)	0.324	0.392

• Reaction Yield Prediction

  Table 5: Influence of the additional hypernode on the performance of Rxn Hypergraph in yield prediction.

  Method	BH	BH1	BH2	BH3	BH4	SM	Gram	Subgram
  Rxn Hypergraph w/o additional Hypernode	0.96	0.81	0.83	0.71	0.56	0.85	0.118	0.196
  Rxn Hypergraph w/ additional Hypernode	0.96	0.82	0.81	0.75	0.57	0.86	0.112	0.187
  Reaction Graph (ours)	0.97	0.80	0.88	0.76	0.68	0.89	0.129	0.216

• Reaction Classification

  Table 6: Influence of the additional hypernode on the performance of Rxn Hypergraph in reaction classification.

  Method	USPTO-TPL	Pistachio-Type
  Rxn Hypergraph w/o additional Hypernode	0.954	0.911
  Rxn Hypergraph w/ additional Hypernode	0.984	0.936
  Reaction Graph (ours)	0.999	0.987

The hypernode method shows improvement in reaction classification, while it has a slight impact on condition and yield prediction. This may be because reaction classification is relatively intuitive, while condition and yield prediction are complex. To improve the performance of condition and yield prediction, more accurate interaction modeling (e.g., Reaction Graph) is needed. Reaction Graph surpasses the performance of the hypernode method, demonstrating its efficiency in interaction modeling. Details can be found in Sec.C.1.2 and G.5.2.

Set2Set vs. SetTransformer

We use Set2Set to capture the global representation of a Reaction Graph by aggregating node features.
We compare the ability of Set2Set and Set Transformer for capturing the global representation of a Reaction Graph.

Table 7: Influence of Set Transformer on the model's performance in USPTO-Condition dataset.

The results indicate that, when used as an aggregation module, Set2Set surpasses the Set Transformer by 5% to 10% in performance and is also more computationally efficient. Detailed implementation methods and analysis can be found in the Sec.C.2.2 and Sec.G.4.3.

Analyzing the Additional Computational Cost in Preprocessing

The proposed Reaction Graph includes angular edges, bond edge lengths, and reaction edges.

To obtain angular edges and bond edge lengths, it is necessary to calculate the 3D conformers of reactants and products based on their SMILES representations. To acquire reaction edges, the RXNMapper tool is used to calculate the mapping of atoms involved in reactions. Therefore, the preprocessing time primarily involves the calculation of reaction edges and the generation of 3D conformers.

The table below details the time required to acquire this information during data preprocessing, in relation to the SMILES length (number of atoms). During data preprocessing, the primary time expenditure is attributed to the calculation of 3D conformers. However, this is an unavoidable and inevitable cost for any method that utilizes 3D structural information.

Table 1: Data pre-processing time for different representations under varying numbers of graph nodes. Time units are milliseconds (ms).

More details of the preprocessing cost analysis can be found in Sec.E.

Formulations and Analysis

Set2Set

• Formulation.

  As detailed in Sec.2 and Sec.C.2.1, the process of Set2Set can be formulize as:

  $$\\boldsymbol{q}^{t+1} = \\sum\_{i=1}^{N} w(\\boldsymbol{n}\_i \\cdot \\boldsymbol{h}^{t}) \\times \\boldsymbol{n}\_i,~~~~~~\\boldsymbol{h}^{t+1}, \\boldsymbol{c}^{t+1} = \\text{LSTM}([\\boldsymbol{q}^{t+1}; \\boldsymbol{h}^t];~\\boldsymbol{h}^t, \\boldsymbol{c}^t),$$ $$\\boldsymbol{r} = \\boldsymbol{W}\\cdot[\\boldsymbol{q}^{T}; \\boldsymbol{h}^{T-1}] + \\boldsymbol{b},$$

  where $w$ is an activation function, $\\boldsymbol{h}$ and $\\boldsymbol{c}$ are the hidden state and cell state of LSTM. $\\boldsymbol{W}$ and $\\boldsymbol{b}$ are learnable parameters of a fully connected layer. After $T$ iterations, we concatenate LSTM's hidden state $\\boldsymbol{h}^{T-1}$ with the aggregated node features $\\boldsymbol{q}^{T}$, and acquire the final reaction representation $\\boldsymbol{r}$ by linear projection.

• Analysis.

  1. Without Unnecessary Message Passing. Set2Set does not introduce unnecessary set message passing operations, preserving the topological features of the graph.

  2. Adaptive Aggregation. The role of $\\boldsymbol{h}\_t$ is similar to the seed vector $s$ in Set Transformer. However, $\\boldsymbol{h}\_t$ is specifically extracted for the current graph, whereas $\\boldsymbol{s}$ remains fixed across all graphs. This adaptive seed vector $\\boldsymbol{h}\_t$ enhances the adaptability of the aggregation process to the graph.

Experiment

We design the following experiment to demonstrate our point:

1. We gradually reduce the number of layers in the Set Transformer Encoder to observe whether the set message passing has a negative effect on graph information extraction.

2. We compare the performance of the Set Transformer Decoder (without adaptive seed vector) and the Set2Set(with adaptive seed cector) to observe whether the adaptive seed vector contributes to the model's performance.

The result is shown in Table 1. Details can be found in Set.G.4.3.

Table 1: The influence of set message passing and adaptive seed vector on model's performance. Using a subset of USPTO-Condition for training efficiency.

According to the results, with the set message passing reduces, the performance gradually improves. The results indicate that set message passing exhibits a negative effect when extracting features from the Reaction Graph. Meanwhile, adaptive seed vector achieves better performance than directly using a fixed seed vector, validating its effectiveness.

Thank you once again for your valuable contributions to improving this paper. If you feel that our efforts have addressed your concerns, we would greatly appreciate it if you could consider raising your rating. Your support is crucial to our work. Should you require any further information, please feel free to contact us. We look forward to your response.

Thank you for acknowledging that we propose an "innovative unified graph representation combining reactants and products" and that our method demonstrates "promising results". We also appreciate your suggestions. Part Ⅰ- Part Ⅵ are responses to weaknesses, and ReplyⅠ- Repy Ⅶ are responses to questions.

Reproducibility

We provide reproducibility support in the following ways:

• Code: We have submitted the code in initial supplementary materials. It includes dataset processing, Reaction Graph construction, network architecture and training framework. It also contains implementations of the compared baselines.

• Guidelines: We provide detailed guidelines in Sec.C and Sec.F. The guidelines include the operating environment, data preprocessing, hyperparameter settings, training methods, and other related aspects.

• Parameters: In this revision, we provide the training parameter checkpoints in the link below. https://huggingface.co/reactiongraph/ReactionGraph

1. Multiple Trials with Different Random Seeds

We train and test the model on all tasks using different random seeds. The results are summarized in the Standard Deviations section. More details on the standard deviations of baseline methods have been added in Sec. G.3.2 and Sec. G.6.1.

2. Standard Deviations

We summarize the standard deviations of all metrics. More detailed information, including the baselines and implementation details, is provided in the Sec. G.3.2 and Sec. G.6.1.

Table 1: Results for leaving group identification on USPTO. The metrics include mean and standard deviations from multiple trials with different random seeds.

Table 2: Results for reaction condition prediction on USPTO-Condition. The metrics includes mean and standard deviations from multiple trials with different random seeds.

Table 3: Results for reaction condition prediction on Pistachio-Condition. The metrics includes mean and standard deviations from multiple trials with different random seeds.

Table 4: Results for reaction yield prediction on Buchwald-Hartwig dataset. The metrics includes mean and standard deviations from multiple trials with different random seeds.

Table 5: Results for reaction yield prediction on Suzuki-Miyaura and USPTO-Yield dataset. The metrics includes mean and standard deviations from multiple trials with different random seeds.

Table 6: Results for reaction classification on USPTO-TPL dataset. The metrics includes mean and standard deviations from multiple trials with different random seeds.

Table 7: Results for reaction classification on Pistachio-Type dataset. The metrics includes mean and standard deviations from multiple trials with different random seeds.

Based on the results, we observe that the standard deviations of the training results is smaller ($≤0.01$) on larger datasets (USPTO, Pistachio). In contrast, we observe a larger training standard deviations ($>0.01$) on some smaller test set.

3. Statistical Significance Tests

Statistical significance testing helps us determine whether Reaction Graph have improvement compared to the baseline method. We select the optimal baseline for each task for comparison. Specifically, $t$-test method is adopted to calculate the statistical significance. $t$ is an indicator of the difference between two experimental groups. The formula is $t = \sqrt{n} \cdot \frac{x - b}{s}$ , where $b$ is the baseline performance and $n = 4$ .

The significance level $P$ corresponding to the $t$-value can be found in the lookup table (https://en.wikipedia.org/wiki/Student%27s_t-distribution). We use $-\log P$ to make the results easier to observe. The larger the value, the higher the probability that Reaction Graph is better. The final calculated significance levels are as follows. Details can be found in Sec.G.6.2.

Table 8: Results for statistical significance analysis on leaving group identification in USPTO. The significance level is calculated by $t$-test method. The data is from the mean and standard deviations of multiple trials with different random seeds.

Table 9: Results for statistical significance analysis on reaction condition prediction in USPTO-Condition. The significance level is calculated by $t$-test method. The data is from the mean and standard deviations of multiple trials with different random seeds.

Table 10: Results for statistical significance analysis on reaction condition prediction in Pistachio-Condition. The significance level is calculated by $t$-test method. The data is from the mean and standard deviations of multiple trials with different random seeds.

Table 11: Results for statistical significance analysis on reaction yield prediction in Buchwald-Hartwig. The significance level is calculated by $t$-test method. The data is from the mean and standard deviations of multiple trials with different random seeds.

Table 12: Results for statistical significance analysis on reaction yield prediction in Suzuki-Miyaura and USPTO-Yield. The significance level is calculated by $t$-test method. The data is from the mean and standard deviations of multiple trials with different random seeds.

Table 13: Results for statistical significance analysis on reaction classification in USPTO-TPL and Pistachio Type. The significance level is calculated by $t$-test method. The data is from the mean and standard deviations of multiple trials with different random seeds.

Based on the results of the $t$-test statistical significance analysis, the advantages achieved by our model have a high level of confidence ($-logP > 3$ indicates $confidence > 95\%$), across most of the datasets and metrics.

4. Cross-validation Results

Cross-validation can reduce the bias introduced by the unevenly distributed reaction dataset. This ensures the model's generalization. We evenly split the dataset used in each task. And we test the model's performance using K-fold cross-validation. The results are as follows. Details of can be found in Sec.G.6.3.

Table 14: Cross-validation results for leaving group identification on USPTO. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

Table 15: Cross-validation results for reaction condition prediction on USPTO-Condition dataset. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

Table 16: Cross-validation results for reaction condition prediction on Pistachio-Condition dataset. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

Table 17: Cross-validation results for reaction yield prediction on Buchwald-Hartwig and Suzuki-Miyaura dataset. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

Table 18: Cross-validation results for reaction classification on USPTO-TPL dataset. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

Table 19: Cross-validation results for reaction classification on Pistachio-Type dataset. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

Based on the results of cross-validation, for large reaction datasets (USPTO, Pistachio), the data distribution differences between different splits are minimal due to the sufficient amount of data, leading to small standard deviations ($≤0.01$) in the training results. For the high-quality Buchwald-Hartwig and Suzuki-Miyaura datasets, the standard deviations introduced by different training/testing splits is also small ($≤0.01$).

5. Confidence Intervals

We use the normal distribution method to calculate the model's confidence interval. Specifically, $CI = x±z · s / \sqrt n$. We take a 95% confidence level, where the corresponding $z = 1.96$. The result is as shown below, with more details in Sec.G.6.4.

Table 19: Confidence interval results ([Min,Max]) for leaving group identification on USPTO. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

Table 20: Confidence interval results ([Min,Max]) for reaction condition prediction on USPTO-Condition dataset. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

Table 21: Confidence interval results ([Min,Max]) for reaction condition prediction on Pistachio-Condition dataset. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

Table 22: Confidence interval results ([Min,Max]) for reaction yield prediction on Buchwald-Hartwig dataset. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

Table 23: Confidence interval results ([Min,Max]) for reaction yield prediction on Suzuki-Miyaura and USPTO-Yield dataset. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

Table 24: Confidence interval results ([Min,Max]) for reaction classification on USPTO-TPL dataset. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

Table 25: Confidence interval results ([Min,Max]) for reaction classification on Pistachio-Type dataset. The interval is calculated using normal distribution method. The data is from mean and standard deviations of multiple trials with different random seeds.

According to the results, in larger datasets (USPTO, Pistachio), the 95% confidence interval for model performance is small ($≤0.01$). This result indicates that the model can achieve stable performance. However, in smaller datasets (Buchwald-Hartwig and Suzuki-Miyaura), some test sets have wider confidence intervals ($>0.01$). It suggests that the model's training variance cannot be ignored.

1. Comprehensive Testing Across Many Reaction Types

In Sec.G.1.1, we provide examples of Attention Weights for various reaction types. The types include amidation, cyclization, and nucleophilic substitution reactions. We also include experiments on dataset scale as in Sec.G.1.2, and analyze the reaction type distribution of the selected dataset in Sec.F.1. It validates that the Reaction Graph effectively helps the model identify reaction centers across different reaction types.

In addition, we conduct a detailed analysis of the results from the dataset scale tests. We calculate two metrics based on the type of response. The first is the proportion of attention weight for the reacting atoms, and the second is the ratio of attention weight between reacting and non-reacting atoms. The results are as follows:

Table 26: The proportion of attention weight for the reacting atoms, across various reaction types.

Table 27: The ratio of attention weight between reacting and non-reacting atoms, across various reaction types.

According to the results, the attention weight shows significant variation across different reaction types. Reaction Graph consistently maintains a clear advantage in all types. For details of these reaction types, please refer to Table 37 in the appendix in Sec.G.3.1.

2. Analysis of Edge Cases and Failure Modes

We add visualization and analysis of boundary cases in Sec.G.2. The Reaction Graph may occasionally misidentify reaction centers, typically due to:

• errors in the reaction itself
• overly complex molecular structures
• rare reactions that do not occur at functional groups

3. Statistical Validation of the Hypothesis

We conduct additional experiments on dataset scale, and quantitatively validate that the Reaction Graph aids in identifying reaction centers. We use USPTO dataset and Condition Prediction Model as in the main text. Attention weights are extracted for Reacting Atoms and non-Reacting Atoms from the Set2Set layer.

We calculate the proportion of attention weights for Reacting Atoms relative to the total, and the results are as follows:

Table 28: The proportion of attention weights for Reacting Atoms to the total attention weights of all atoms. Compare between Molecular Graph and Reaction Graph.

We also calculate the average attention weights for Reacting Atoms and non-Reacting Atoms, denoted as $a$ and $b$. And we compute their ratio $a/b$. The results are as follows:

Table 29: The ratio $a/b$ of the attention weights $a$ of Reacting Atoms to the attention weights $b$ of non-Reacting Atoms. Compare between Molecular Graph and Reaction Graph.

According to the results, the model using the Reaction Graph shows significantly stronger attention to reaction centers. Reaction Graph achieves a performance improvement of 40%. Considering the sparsity of leaving groups, this is a substantial enhancement.

4. Discussion of Potential Selection Bias in Examples

We use the following methods to reduce selection bias. These methods enhance the reliability of the conclusions from attention weight visualization.

• Automated Algorithms. We use the algorithms described in Sec. D.3 to automatically extract visual samples from the dataset. This automated procedure reduces the selection bias introduced by subjective factors.

• Dataset Size Testing. In addition to selecting visual samples, we also conducted attention weight testing on the dataset size. This method can avoid the selection bias introduced by sample selection. The results can be found in Section 3. Statistical Validation of the Hypothesis.

Thank you for your comments. We provide further explanations step-by-step in the hope of addressing your concerns.

• Citation. Thank you for pointing out this issue. We correct the citation of RXNMapper in this revision.

• Multiple Possible Conformations Handling. Theoretically, for multiple possible conformations converged from MMFF, regardless of which one is chosen, the 3D features of Reaction Graph unchanged. Details will later be provided in Reply Ⅲ: Handling Multiple Conformation by Invariant 3D Feature Modeling.

• Elaboration of Edge Length. We do not use pairwise distances. Instead, we construct 3D information based on molecular graphs as in Sec.3.1. We only calculate the distances for nodes $i$ and $j$ that are connected by edges in the graph. This approach can address the issue of flexibility. For details, please refer to Reply Ⅲ: Handling Multiple Conformation by Invariant 3D Feature Modeling.

• Basis of Reaction Graph Details will later be provided in Reply Ⅳ: Reaction Graph uses Reaction Itself as Basis.

• Molecular Geometry Handling.

  • We do not explicitly calculate bond angle. Instead, we propose angular edge to implicitly represent bond angle. Details can be found in Sec.3.2 and Sec.C.1.
  • Torsion angle is considered in our analysis in Sec.A.2. Detailed discussion can later be found in Reply Ⅲ: Handling Multiple Conformation by Invariant 3D Feature Modeling.
  • Details of conformations selection can be found in Reply Ⅴ: Criteria for Conformer Selection
  • The statistical analysis and cross validation provides assessment of the impact of conformational change. Please refer to Sec.G.6 and Part Ⅱ: Statistical Analysis. And Reply Ⅲ also provides methods to deal with conformational change.

• Terminology. A terminology list will be provided to maintain consistency at Sec.J.

• Technical Detail. Sec.C, Sec.D, and Sec.F provides comprehensive details of our method.

1. Error Analysis of Misclassified Cases

Additional error analysis is provided in Sec.G.2.2. In summary, molecular graph often misidentify leaving group positions, thereby hindering accurate classification. In contrast, Reaction Graph position leaving groups correctly at most time. But it still have minor errors in specific type classification. This shows that Reaction Graph effectively helps the model to focus on reaction related features.

2. Comparison with Additional Baseline Methods

We conduct extra experiments with D-MPNN and Rxn Hypergraph. The results are shown in the table below (Table 30). According to the results, Reaction Graph achieves the best performance in leaving group identification compared to other baselines.

Table 30: Performance of different methods on the Leaving Group Identification task.

3. Cross-validation Results

We provide cross-validation results of Leaving Group Identification on USPTO (Table 31). According to the results of cross-validation, Reaction Graph's performance in leaving group identification still far exceeds that of the molecular graph. This demonstrates that its design for reaction modeling effectively addresses the shortcomings of the molecular graph.

Table 31: Cross-validation results for leaving group identification on USPTO. The metrics includes mean and standard deviations from multiple trials with different train/test splits.

4. Discussion of Systematic Errors or Biases

In leaving group identification, systematic errors may be introduced through:

(1) Dataset biases

(2) Randomness during the training process

(3) Differences in the computing environment

To address these issues, we follow your advise:

(1) We use cross-validation across various dataset split to mitigate dataset biases

(2) We adopt multiple random seeds to reduce systematic errors caused by randomness. According to the results in Table 32, the standard deviation for the leaving group identification across different trials is small. The small standard deviation indicates that the results reliably reflect the model's performance.

(3) We provide detailed descriptions of the machines and environment in Sec. C.3.

Table 32: Results for leaving group identification on USPTO. The metrics include mean and standard deviations from multiple trials with different random seeds.

5. Analysis of Chemical Pattern Dependencies

In chemical reactions, the most important chemical pattern is the structure of the molecules, as well as the changes during the reaction. Through comparisons with more baselines in Table 30, we found that the patterns of reaction changes play a crucial role in the positioning of leaving groups. At the same time, the molecular topology patterns also support the classification of leaving groups.

Reaction Graph conveys 3D structure while integrating patterns of reaction changes, thereby achieving the best performance in the leaving group identification task.

1. System Specifications

• CPU: Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz, 64 Cores, 1200-3900 MHz

• RAM: 472GB (32GB * 16), DDR4, 2666 MT/s

• Disk: 72.4TB, HDD

• GPU: NVIDIA GeForce RTX 4090, 24GB * 8

2. Benchmark Conditions

All the test is conducted under Ubuntu 22.04.5 LTS, with cuda 11.0, pytorch 1.12.1+cu113, rdkit 2023.9.4, dgl 0.9.1.post1. Other toolkit versions are provided in the attachments.

We use the time module to test the runtime.

3. Variability in Performance

We use the number of nodes as an indicator of data complexity, and measure the variability in performance at different data scales.

Specifically, we test the pre-processing efficiency (Table 33) and computational efficiency (Table 34) of different baselines and our method. According to the results in Table 33, the pre-processing time increases non-linearly with the number of nodes in the molecular graph. This is due to the complexity of the MMFF algorithm. We also observe certain fluctuations in computation time. This phenomenon is caused by the uncertainty in the iterative convergence of the MMFF algorithm. As shown in Table 34, the computation time for each type of graph increases linearly with the number of atoms. This trend is due to the relative sparsity of the molecular graph.

Table 33: Data pre-processing time for different representations under varying numbers of graph nodes. Time units are milliseconds (ms).

Table 34: Model inference time for different graph representations under varying numbers of graph nodes. Time units are milliseconds (ms).

4. Hardware Dependencies

• Training (same batch size and code as ours)

  • GPU: At least 24GB VRAM
  • RAM: At least 16GB

• Inference

  • GPU: 2GB VRAM

Thank you for your questions. We respond to the first two questions in Part Ⅰ: Ensuring Reproducibility: Code, Guidelines, and Parameters and Part Ⅱ-1 ~ Part Ⅱ- 4: Statistical Analysis.

For System Specifications and Benchmark Conditions, please refer to Part Ⅵ: Runtime Analysis Details. The runtime analysis is as follows.

Our method

1. Theoretical Analysis

   We define all graph-based representations of chemical reactions as $G(V, E)$, where $V$ represents the set of atoms and $E$ represents the set of bonds. The computation complexity of the Reaction Graph depends on the number of atoms $|V|$, the number of bonds $|E|$, the maximum degree $D$ of the atoms, and the specific implementation of the graph.

   In Reaction Graph, a reactant atom can match with at most one product atom, so the number of reaction edges does not exceed half the total number of atoms, $|V|/2$. Furthermore, as defined for angular edges in Fig. 1, the maximum number of angular edges is $|V| D^2$.

   Thus, the computational complexity contributed by bond edges is $|E|$, by reaction edges is approximately $|V|/2$, and by angular edges is around $|V| D^2$. Consequently, the total computational complexity of Reaction Graph can be approximated as $|E| + |V|/2 + |V| D^2$.

2. Experimental Results

   We assess the impact of reaction edges and angular edges in the Reaction Graph on model inference speed as the number of atoms $|V|$ increases. The findings are presented in rows 1-3 of Table 36. The time cost associated with reaction edges is lower than that caused by angular edges.

Compared with Other Methods

1. Theoretical Analysis

   • Graph-based methods

     When the molecular graph is represented as $G(V,E)$, $\vert V \vert$ indicates the total number of atoms in reactants and products, and $\vert E \vert$ represents the total count of bonds. The computation cost for the molecular graph is $\vert E \vert$. The computation cost of Rxn Hypergraph is at least $\vert E \vert + \vert V \vert$. The edge count in Condensed Graph of Reactions (CGR) is roughly $\vert E \vert /2$. However, when handling the same amount of information, the dimensions of the fully connected layers in CGR far exceed those in the Reaction Graph and Rxn Hypergraph. This leads to greater computational overhead.

   • SMILES-based methods

     SMILES-based methods are generally based on Transformer. When the length of the input SMILES is $\vert L \vert$, the computation complexity is roughly $\vert L \vert ^2$. Since there is no direct conversion between the length of SMILES strings $\vert L \vert$ and the number of nodes $\vert V \vert$ and edges $\vert E \vert$ in molecular graphs, we directly measured their computational efficiency through experiments.

2. Experimental Results

   • Compared with graph-based methods

     We compare the inference time of Reaction Graph and other graph-based methods. The results are shown in rows 4-6 of Table 36. Due to incorporating additional reaction information and 3D structures, our method requires slightly more time than molecular graphs and Rxn Hypergraph representations. However, it delivers significant performance improvement, justifying the additional time cost.

     Table36: Model inference time for different graph representations under varying numbers of graph nodes. Time units are milliseconds (ms).

     Representation	Reaction Information	3D Structure	2500	3000	3500	4000	4500
     Reaction Graph	√	√	37.28	43.77	50.57	58.03	65.45
     w/o Reaction Edge	×	√	35.36	40.92	48.68	53.30	58.79
     w/o Angular Edge	√	×	22.51	25.96	29.82	34.21	38.77
     Molecular Graph	×	×	18.59	21.93	25.39	28.30	31.48
     Rxn Hypergraph	×	×	31.18	37.28	43.31	48.53	54.06
     Condensed Graph of Reactions	√	×	44.55	52.48	60.75	66.32	75.34

   • Compared with SMILES-based methods

     We compare the inference time of Reaction Graph and SMILES-based methods. The results are shown in the table below, where Reaction Graph, as a graph-based model, exhibits the shortest inference time.

     Table37: Inference time comparison for Reaction Graph and other methods.

     Methods	Representation	Time
     T5Chem	SMILES	3min 19s
     RXNFP	SMILES	23min 43s
     Reaction Graph (ours)	Reaction Graph	2min 36s

The details of the cost analysis can be found in Sec.E.

Thank you for your comment. We respond to this question in Part Ⅴ: Leaving Group Identification Analysis.

The algorithm for constructing the reaction graph is described in Sec.3.2, Sec.C.1.1, and Sec.D.1. Here, we provide a brief summary:

Reaction Graph presents a unified modeling of the entire reaction, rather than being based solely on part of the reaction. During the construction of Reaction Graph, the reactants and products are built simultaneously (including their conformations) in the first step. In this phase, they have no interdependencies or specific order. After each molecule is generated, the construction algorithm connects the reactants and products into a unified Reaction Graph. This stage is still based on both reactants and products.

In Sec. D.1 and Sec. F.1, we detail our handling and selection of conformations.

The selection of conformations theoretically does not affect the 3D information of Reaction Graph, as metioned in Reply Ⅲ. However, we still generate multiple conformations for each molecule to account for errors from the MMFF algorithm. Unlike molecular datasets, we do not generate multiple conformations for each reaction in reaction dataset. This is because the same molecule can appear in different reactions. We only need to calculate one conformation for each reaction, then the molecules in the dataset can have multiple conformations.

For conformation for each molecule, we try to select the conformation calculated by the most accurate algorithm. Specifically, we use RDKIT, starting with the ETKDG algorithm for initialization. The MMFF algorithm is then used for optimization. If it fails to converge, we switch to the UFF algorithm. If that also fails, we keep the ETKDG results. If ETKDG fails initially, we use a simpler 2D conformation.

Our basis is:

• Large dataset allows the model to learn from conformation calculation errors.
• We retain molecules whose conformation cannot be computed, because their structure and labels still provide valuable prior information for the model.

1. The 3D features of RG are invariant to different conformations.

A molecule has different conformations due to the rotation of single bonds. Reaction Graph (RG) have already considered multiple conformations of a molecule. Specifically, the 3D features used by RG are bond lengths and bond angles. In different conformations of the same molecule, bond lengths and bond angles generally change only slightly. Therefore, the 3D features of RG are invariant to different conformations. The conformational invariance of the RG ensures consistency of 3D features across training samples, making it easier for the model to learn.

2. Properties of chemical reactions are invariant to different conformations.

There is a chemical reaction $A + B \xrightarrow{\text{D}} C$. The catalyst $D$ won't change with the conformations of $A$, $B$, or $C$. Thus, RG is consistent with the task objectives.

3. Compared to other 3D features that vary with conformations.

We also explore other 3D features that vary with conformations, including torsion angles, bond vectors, and atom coordinates. The specific implementation of those methods can be referenced in Sec.C.1.1 and Sec.C.2.2. The results are shown in Table 35. It can be seen that RG, which includes only bond lengths and bond angles as 3D information, achieves the best results.

Table 35: Influence of methods to handling multiple conformation on models performance. Bond Vector, Atom Coordinate and Torsion Angle will change with conformation. In contrast, Bond Length and Bond Angle are invariant to specific conformation.

Dear Reviewer uB2j,

We are glad to have addressed your concerns. Thank you for raising the score, and we especially appreciate your efforts in improving this work.

Best regards,

Paper 4277 Authors

Thank you for acknowledging our method "achieves better prediction accuracy". We also appreciate your constructive suggestions.

Novelty of Reaction Graph

• First graph to understand chemical reactions at the atom level. When using deep neural networks for chemical reaction modeling, existing methods lack a mechanism to account for the interactions between atoms during a reaction. Typically, these methods use molecule-level global representations of reactants and products for chemical reaction reasoning, which fails to effectively capture the underlying reaction mechanisms. In contrast, our Reaction Graph is equipped with the ability to model relationships at the atom level, thus enabling effective reasoning of reactions.

• First graph with 3D chemical reaction modeling. Existing graph-based methods for representing chemical reaction only consider 2D molecular structure, while overlooking 3D structural information. In constrast, our Reaction Graph is capable of modeling 3D structural transformation during chemical reaction, which better captures the dynamics and changes occurring throughout the reaction.

• Data-and-Knowledge-Driven Chemical Reaction Reasoning. Existing methods for understanding chemical reactions typically employ either rule-based or data-driven approaches. In our work, we successfully integrate chemical knowledge and rules into deep learning to enhance reaction reasoning.

Explanation of Vertex-Edge Integration

Experiment

We also conduct experiment to tested different vertex-edge integration methods and demonstrated that our chosen approach has the best performance.

Table 2: Impact of different vertex-edge integration methods on model's performace.

The result shows that our method is superior to other methods, demonstrating its efficiency. Details could be found in Sec.C.2.2 and Sec.G.4.2.

Why an attention-based aggregation method with an LSTM is employed?

Effectively identifying and leveraging the chemical reaction mechanism to understand and reason about reactions is challenging. In our work, we use an attention mechanism to adaptively capture the most important cues for reaction modeling. However, since these cues are not always easy to identify in one time, we employ an LSTM to progressively and interactively discover them. As shown in Fig.3 in the main paper, the attention-based aggregation module with LSTM accurately locates the reaction center on Reaction Graph.

We conducted experiments to demonstrate the roles of Attention and LSTM. Specifically, we compared the performance of the aggregation module without using Attention and LSTM on USPTO-Condition dataset.

Table 3: Impact of Attention and LSTM in aggregation module.

The result in the table above shows that the Attention mechanism and LSTM contributes to the performance. Detail implementations and analysis can be found in Sec.C.2.2 and Sec.G.4.3.

Explanation of Vertex-Edge Integration

Design

Reaction Graph involves multiple types of edges. Effectively utilizing these edges to improve the modeling of chemical reaction structures and interactions with vertices (atoms) is crucial for accurate reaction reasoning. In our method, inspired by Dynamic Convolution[1] and Hypernetworks[2], we first use these edges to generate kernel weights

$\mathcal{M_{ij}} = \text{Reshape}(\boldsymbol{W} \cdot [\boldsymbol{l_{ij}};\boldsymbol{e_{ij}}]),$

which are then applied to vertex features.

$\\boldsymbol{v}^{t+1}\_i = \\boldsymbol{v}^{t}\_i + \\sum\_{j \\in \\mathbb{N}\_i} \\mathcal{M}\_{ij} \\cdot \\boldsymbol{v}^{t}\_j,$

In this way, our method can effectively capture the reaction information.

Our vertex-edge integration module handles different types of edges in the Reaction Graph and reflects the complementary role of 3D information to the 2D topological structure.

Specificly, we start from the basic paradigm of MPNN:

$\\boldsymbol{v}^{t+1}\_i = \\boldsymbol{v}^{t}\_i + \\sum\_{j \\in \\mathbb{N}\_i} f(\\boldsymbol{v}\_i,\\boldsymbol{e}\_{ij},\\boldsymbol{v}\_j),~~~~~\\boldsymbol{v}'\_i = \\boldsymbol{v}^{T\_1}\_i,$

where $f$ is the message function to integrate vertex and edge information, $\\boldsymbol{v}\_i$ is the node feature, and $\\boldsymbol{e}\_{ij}$ is the edge feature. Reaction Graph contains three distinct types of edges, including bond egde for modeling interatomic force ($e\_{ij} = 1$), reaction edge for modeling atomic mapping ($e\_{ij} = 2$), and angular edge for modeling 3D multi-body relation ($e\_{ij} = 3$). Based on the principle of MPNNs, we believe the message function $f(\\boldsymbol{v}\_i,\\boldsymbol{e}\_{ij},\\boldsymbol{v}\_j)$ for these edges should be fundamentally different:

$f = \begin{cases} f_{bond} & \text{if } e_{ij} = 1, \\\\ f_{reaction} & \text{if } e_{ij} = 2, \\\\ f_{angular} & \text{if } e_{ij} = 3. \end{cases}$

We use linear layers to implement these different message functions as $f\_t(\\boldsymbol{v}\_i,\\boldsymbol{v}\_j) = \\mathcal{M}\_t \\boldsymbol{v}\_j$, where $t \\in\\{bond,reaction,angular\\}$. Based on these, the hypernetwork-like design can generate different linear layer weight matrices based on the different edge types to process node features (assume $\\boldsymbol{e}\_{ij}$ is the one-hot representation of $e\_{ij}$):

$\\mathcal{M}\_t = \\text{Reshape}(\\boldsymbol{W} \\cdot \\boldsymbol{e}\_{ij})$

Meanwhile, the weight matrices generated from edge length embeddings serves as the adaptation of edge type weight matrices as (omit $ij$ for convenience):

$\\mathcal{M} = \\mathcal{M}\_t + \\Delta\mathcal{M}\_{3D},~~~~~~ \\Delta\\mathcal{M}\_{3D} = l\_1\\mathcal{M}\_1 + l\_2\\mathcal{M}\_2 + \\dots + l\_k\\mathcal{M}\_k,$

where $\\boldsymbol{l}$ is the RBF embedding, $\\mathcal{M}\_t \\in \\{\\mathcal{M}\_{bond}, \\mathcal{M}\_{reaction}, \\mathcal{M}\_{angular}\\}$, and $\\mathcal{M}\_1$ to $\\mathcal{M}\_k$ is the learningable 3D bias matrices. This approach is liken to using ControlNet to add prior information such as wireframes based on the original parameters of Stable Diffusion, and is consistent with the complementary role of 3D information to 2D topology and atomic information. The final formulation is as:

$\\mathcal{M}\_{ij} = \\text{Reshape}(\\boldsymbol{W} \\cdot [\\boldsymbol{l}\_{ij};\\boldsymbol{e}\_{ij}]),~~~~~~\\boldsymbol{v}^{t+1}\_i = \\boldsymbol{v}^{t}\_i + \\sum\_{j\\in \\mathbb{N}\_i} \\mathcal{M}\_{ij} \\cdot \\boldsymbol{v}^{t}\_j,~~~~~\\boldsymbol{v}'\_i = \\boldsymbol{v}^{T\_1}\_i,$

In the implementation, we also introduce GRU and residual connections. Since the MPNN increases the receptive field of node features in each iteration, the gating mechanism of GRU helps retain the sturctural information of appropriate neighbor size (some neighbor structure with certain size might act as a functional group in molecule), while residual connection combines neighbor information with atomic features.

For more details, please refer to Sec.C.2.

[1]Chen Y, Dai X, Liu M, et al. Dynamic convolution: Attention over convolution kernels[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 11030-11039.

[2] Ha D, Dai A M, Le Q V. HyperNetworks[C]//International Conference on Learning Representations. 2017. https://openreview.net/forum?id=rkpACe1lx.

Why RBF is used to embed the edge length?

RBF can effectively capture the non-linear relationships between distance and molecular property. This method lifts the raw edge lengths into a high-dimensional vector that can be more easily utilized by machine learning models, focus more on local structural pattern, and produces smooth mappings which help in capturing variations in continuous data. These advantages make RBF kernel suitable for tasks involving local continuous spatial relationships, such as in molecular structures where edge lengths indicate bond distances.

To demonstrate the efficiency of RBF kernel embedding, we conducted experiments to compare it with different embedding methods, using USPTO-Condition dataset.

Table 1: Impact of different edge embedding methods on model's performace.

As shown in the table above, RBF kernel outperforms other embedding methods, demonstrating its efficiency. Details could be found in Sec.C.2.2 and Sec.G.4.1.

Dear Reviewer N9iv,

Thank you very much for your time and effort in reviewing our submission. We have tried our best to address the concerns and problems mentioned. In particular, we explain the novelty of this work and its significance to the development of interdisciplinary computer science. We hope that our responses have addressed your concerns. Your support is significantly important to our paper. If you need further information, please let us know. We look forward to hearing from you. Thank you once again.

Best wishes,

Paper 4277 Authors

1. Contribution to Computer Science

• Promoting interdisciplinary development. Interdisciplinary research is an important direction in computer science. We are the first to integrate interaction modeling from computer science into reaction tasks, thereby promoting the development of computer science methods in the field of chemistry.

• Application of computer science concepts. The design of 3D information in Reaction Graph incorporates design principles from computer science, including modal consistency and algorithm parallelism. Our work demonstrates that the integration of computer science concepts is effective, and thus encourage their application in interdisciplinary research.

• Provides insights for computer science tasks. Our approach handles complex individual interactions and geometric information. This also provides insights for graph representation learning and 3D geometric data modeling tasks in computer science.

2. Contribution to Chemistry

• State-of-the-art performance. Our proposed Reaction Graph achieves state-of-the-art performance in multiple chemical reaction tasks. This provides better support for chemical tasks such as synthetic design, thereby driving chemical discoveries.
• Chemical knowledge integration. Reaction Graph integrates domain knowledge such as atomic mapping. This enhances the model's adaptability to chemical tasks, and our design philosophy promotes the development of domain models.
• Foundational for reaction modeling. Reaction Graph is an improvement at the graph representation level. It can be effectively utilized in any future method based on GNNs and enhance performance in chemical reaction tasks. This innovation promises a lasting impact on the field of chemical reaction property prediction.

Dear Reviewer N9iv,

Thank you very much for your efforts in reviewing our paper. The discussion phase is nearing its end, and we are looking forward to receiving your feedback on our responses. Should you have any further concerns, we are glad to provide additional information. Thank you once again.

Best regards,

Paper 4277 Authors