Thanks for your detailed and constructive review. We sincerely appreciate the time and effort you invested in evaluating our manuscript. We are encouraged by your recognition of the motivation, clarity, and potential impact of our proposed chemical core subset selection strategy and asymmetric molecular interaction network. At the same time, we acknowledge the areas that require further clarification and improvement, particularly in the discussion of GNN-related components, the differentiation from existing information-theoretic methods, and the need for more concrete efficiency metrics. Below, we provide a point-by-point response to your thoughtful comments and suggestions.

W1. Discussion on Graph Neural Network

Thank you for pointing out this issue. We will add the following discussion on GNN in molecular relationship learning:

• Graph neural networks (GNNs) [1] have shown strong expressive power in molecular relationship modeling and have been widely used to predict molecular properties, reaction behavior, and solubility free energy. Traditional GNNs model the interactions between atoms and bonds through message passing between nodes to extract the structural and chemical characteristics of molecules. Typical models such as D-MPNN [2] and GAT [3] use this method to complete the representation learning of molecular graphs. Models such as MMGNN [4] and CGIB [5] try to innovate in modeling the interaction between molecular pairs.

• However, in actual chemical systems, there are asymmetric and diverse relationships between molecules (such as solutes and solvents). Traditional GNNs are limited in their ability to model such heterogeneous information. To this end, heterogeneous graph neural networks (HetGNNs) [6] provide an effective solution. HetGNN expands the modeling capabilities of GNN, introduces multiple types of nodes (such as molecules of different categories) and edges (such as hydrogen bonds, hydrophobic interactions, etc.), and supports type-based message passing mechanisms. However, in the field of molecular relationship learning such as dissolution reaction prediction or drug interaction prediction, there is no mature solution for heterogeneous graph construction.

[1] The graph neural network model.

[2] Transfer learning for solvation free energies: From quantum chemistry to experiments.

[3] Graph attention networks.

[4] MMgnn: A molecular merged graph neural network for explainable solvation free energy prediction.

[5] Conditional graph information bottleneck for molecular relational learning.

[6] Heterogeneous Graph Neural Network.

W2. The difference between CSIE and existing jobs

Although methods such as Glister and InfoBatch have explored data subset selection strategies based on information gain, gradient diversity or label uncertainty in recent years, these methods still have two fundamental limitations when directly transferred to the field of molecular chemistry, which also constitute the starting point and core innovation of the CSIE framework proposed in this study:

(1) The weak alignment problem between chemical structure and semantics leads to misleading sample selection by existing methods. Glister uses the sensitivity estimate of model gradient to generalization error to measure the "importance" of samples, while InfoBatch relies on historical gradient trajectories to dynamically screen samples. However, these methods are heavily dependent on the representation or loss function feedback learned by the model in the early stages of training. In chemical molecular data, early representations often fail to capture complex and subtle changes in chemical properties, such as resonance structures, stereo effects, and non-covalent interactions. This leads to: 1. Over-attention to rare or difficult samples (e.g., pairs of molecules with extremely strong or weak polarity), thereby sacrificing the representativeness of the chemical space; 2. Ignoring important samples that are "fuzzy" in the representation space, such as molecules with similar topology but significantly different behaviors. The innovation of CSIE is introducing multimodal semantic embedding (MolCA) as the basic representation, which is independent of upstream and downstream tasks, and measuring whether the sample brings structural information gain by information entropy change, thereby avoiding overfitting dependence on early task models.

(2) The update computational cost under large-scale data is too high, making it difficult to use for high-throughput chemical screening. InfoBatch needs to frequently calculate indicators such as gradient change and loss difference, and Glister needs to evaluate the proxy objective function after each round of training, which significantly increases the computational overhead in the molecular space (usually tens of thousands of samples). Especially in practical scenarios, such as drug discovery or solubility screening, training cost is one of the deployment difficulties. The combined strategy proposed by CSIE has the following three efficiency advantages: the approximate nearest neighbor graph (HNSW) in sparse high-dimensional space can efficiently estimate the relative information density between samples; the submodular function structure ensures that the information gain of each round of subset optimization is monotonically constant and the optimization converges; the EM mechanism is introduced to iteratively correct the insertion bias to avoid the local optimum caused by one-time greedy selection. Through these designs, CSIE controls the time complexity to O(log(n)), which has the practical potential to process real-scale chemical data.

W3. Time and memory efficiency issues

Regarding the time efficiency, we explain as follows: During the model training process, the CSIE framework reduces the time and space costs mainly by reducing the size of the data set. For example, in the FreeSolv data set, by filtering samples through the CSIE framework, only 60% of the training data is needed to achieve performance that exceeds all training data. Therefore, for this data set, it can be said that CSIE saves 40% of the training cost and achieves better results. In Table 15 of Appendix E.5.2, the superiority of the CSIE framework in transfer learning is more intuitively demonstrated. It can achieve results that exceed 30% of the data with only 5% of the data, and the training time is more than 130 hours. Similarly, smaller data sets will also occupy less memory, but it should be acknowledged that CSIE is a core subset screening framework and does not affect the memory of the model.

Q1. discussion about sample selection of CSIE

It should be clarified that the amount of information is only related to the chemical semantic features of the sample itself, and has nothing to do with the prediction results. This article uses both molecular graph features and semantic features for encoding, and uses the general chemical model MolCA pre-trained on PubChem to fully extract sample features and ensure the completeness of sample chemical semantics as much as possible. Therefore, the CSIE framework does not prefer difficult samples. It prefers to select samples that are not yet included in the chemical space of the core subset or are in small numbers. In addition, in Section 3.3, we also visualized the screening effect of CSIE itself on the dataset: we divided the dataset according to solvent type, and evaluated the sample category distribution of the dataset before and after CSIE screening. It can be clearly seen that CSIE can indeed alleviate the redundancy problem of the dataset, obtain a more balanced dataset, and reduce overfitting.

Thank you again for your thoughtful and constructive review. Your feedback has been valuable in helping us improve the clarity, rigor, and positioning of our work. We have carefully revised the manuscript to implement your suggestions. We sincerely hope that our detailed response has adequately addressed your concerns. If the answer is yes, we would be grateful if you would consider revising the overall assessment of the manuscript. Of course, we are happy to provide any further clarification or participate in follow-up discussions as needed.

Dear Reviewer,

Thank you sincerely for your valuable time, thoughtful feedback, and constructive engagement during the review process. Your detailed comments and insightful suggestions have been crucial in enhancing our manuscript's clarity, technical rigor, and overall quality.

Your careful attention to key aspects of our work—such as the motivation and design of the CSIE framework, its relationship with existing GNN-based molecular modeling strategies, and the differentiation from information-theoretic subset selection methods—prompted us to refine our explanations and strengthen both the theoretical and empirical justifications of our approach.

We particularly appreciate your recognition of the novelty and potential impact of the CSIE strategy in chemical data modeling. Your willingness to engage with our point-by-point responses and your encouraging final assessment have motivated our team and reaffirmed the significance of our contribution. As you suggested, we have added a comparative description of CSIE, GloSTER, and InfoPatch in the revised manuscript to help readers clearly understand their differences in methodology, applicability, and efficiency.

Thank you again for your constructive support and guidance. Your feedback has been instrumental in shaping this work, and we truly value your contribution to its development.

Best regards, Authors

Thanks for your detailed and constructive comments. We appreciate the time and effort you dedicated to reviewing our manuscript. While we recognize that several aspects required clarification and improvement, we are encouraged by your acknowledgment of the potential value of our proposed subset selection strategy. Below, we provide a point-by-point response to your concerns.

W1. The connection between CSIE and AMGNN

• Both CSIE and AMGNN are designed to address data quality challenges in AI-driven chemistry. Theoretically generated datasets are often large but noisy, with significant class imbalance, while high-quality experimental data are scarce. To better leverage both types of data, we propose two complementary strategies:
  (1) Use CSIE to streamline theoretical datasets by reducing redundancy and improving distribution, thereby lowering pretraining costs.
  (2) Use AMGNN to better simulate real-world solute–solvent interactions and enhance model performance on experimental datasets.

• It is important to emphasize that CSIE is intended as a general-purpose framework for theoretical data screening. As stated in the title, dissolution free energy prediction serves merely as a case study. We also note in line 82 that Appendix E.5.1 includes additional validation on the material band gap prediction task, further demonstrating the versatility of CSIE.
  With such a general framework, we are able to select lighter, better-balanced pretraining datasets (Figure 3), which in turn improve the performance (Table 1) and generalization ability (Tables 2 and 3) of AMGNN.

W2 & Q3. Clarification and Validation of AMGNN Design

• Design Motivation

  The core of AMGNN is not to propose a new GNN architecture, but to introduce a more chemically faithful graph representation by constructing an asymmetric heterogeneous solute–solvent graph using DGL.

  Most existing approaches (e.g., Uni-Mol, Explainable GNN, CIGIN) treat the two molecules independently, and then model interactions via concatenation or attention, which results in a symmetric treatment of solute and solvent.

  However, molecular interactions are inherently directional. For instance:

  • Ethanol dissolves in water: −5.1 kcal/mol
  • Water dissolves in ethanol: −6.3 kcal/mol

  Such asymmetry is chemically meaningful but cannot be captured by symmetric designs. Therefore, AMGNN directly builds a directed heterogeneous graph, which encodes the interaction direction. This enables the model to learn directional dependencies from both structural and relational cues.

• Baseline Comparison & Validation

  To ensure AMGNN’s superiority is not due to model complexity, we implemented and compared against three strong, interpretable baselines:

  1. Flagged GNN: Adds solute/solvent flags to atom features in a standard GNN.
  2. Heterogeneous GNN: Uses node types to distinguish solute and solvent atoms.
  3. Embedding Concatenation: Separately encodes solute/solvent, then concatenates for prediction.

  To further evaluate AMGNN’s ability to model directionality, we constructed a new evaluation subset:

  • Opposite-Exp: Contains all 240 solute–solvent pairs in CombiSolv-Exp where both directions exist (i.e., A→B and B→A).
  • This allows direct testing of the model’s sensitivity to role reversal.

As shown in the table, our AMGNN consistently outperforms (MAE(↓)) all three approaches, suggesting that the proposed asymmetric design more effectively captures solute–solvent interactions.

W3. Performance improvement analysis

• Performance Analysis of AMGNN:

  • In Table 1, since CombiSolv-Exp contains a relatively large amount of data, the performance differences among models may not appear very pronounced. However, in Table 11 on page 20 of the Appendix, we conduct a necessary significance analysis, which confirms that the performance improvement of AMGNN is statistically significant.

  • Furthermore, for the comparison between AMGNN and other models, Table 12 on page 21 of the Appendix provides additional validation. On smaller datasets, AMGNN demonstrates a more noticeable advantage.

• Performance Analysis of CSIE:

  • In Table 10 on page 20 of the Appendix, we also present the significance analysis for CSIE.
  • In Observation 7 of Section 3.4, we note that the comparable performance of CSIE and systematic sampling stems from the dataset’s inherent solvent-type organization, which naturally balances categories. As shown in Table 9 of the Appendix, systematic sampling offers no significant advantage on CombiSolv-QM. To confirm this, we shuffled the sample order in Abraham and Compsol datasets and repeated the experiments; results (marked with *) are presented below.

the CSIE framework is robust when the original dataset is perturbed, while the performance of the method adopted by the system is significantly degraded.

Q1. Motivation for subset sampling

• In W1, we have thoroughly explained the motivation for applying the CSIE framework to dissolution free energy prediction, particularly in collaboration with AMGNN to leverage both theoretical and experimental data. We further demonstrated the generalizability of CSIE across various solute–solvent datasets and in the Material Band Gap Prediction dataset (Appendix E.5.1).

• Moreover, in solute–solvent prediction tasks, the CombiSolv-QM dataset exhibits notable redundancy and category imbalance, which can hinder model pretraining and transfer learning. As shown in Section 3.3 (Figure 3b), when samples are grouped by solvent type, aromatic hydrocarbons make up 29.5% of the dataset, while halogen-containing solvents account for only 7.6%, highlighting the skewed distribution.

Q2. Comparison with full data training

Thanks for your careful observation. However, we would like to clarify that for many datasets, the CSIE subset selection framework effectively balances the sample distribution, achieving better model performance with fewer training samples:

• First, we sincerely apologize for any confusion caused by the unclear design and analysis of Figure 4. In fact, the results shown in Figure 4 indicate that CSIE achieves better performance than using the full dataset on Compsol, Freesolv, and MNsol, while using only 80%, 60%, and 80% of the data respectively. We discuss and analyze this phenomenon in Obs.8 on page 9 of the main text.

• Second, on CombiSolv-QM, we tested the effect of different subset ratios on transfer learning performance. A similar trend was observed as described above. The detailed results and analysis can be found in Appendix Section E.5.2 and Table 15.

We greatly appreciate your insightful and helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research.

Thank you for your detailed and thoughtful rebuttal. I appreciate the effort invested in addressing each of the raised concerns. Below, I provide a brief assessment of each point.

W1. Connection between CSIE and AMGNN

While I appreciate the clarification and the broader context provided regarding the interplay between CSIE and AMGNN, I remain somewhat unsatisfied with this aspect. The manuscript could still benefit from a more explicit articulation of how these two components are jointly motivated and integrated from the outset. That said, I acknowledge the authors' efforts to generalize CSIE beyond a single application and their inclusion of a secondary validation task.

W2 & Q3. AMGNN Design and Validation

This part of the response was well-reasoned and convincing. The clarification regarding the asymmetric solute–solvent interaction modeling and the construction of a directed heterogeneous graph was particularly helpful. The additional baselines and the Opposite-Exp subset offer strong support for the efficacy of the design. I am satisfied with this explanation.

W3. Performance Improvement Analysis

While I appreciate the authors' providing statistical significance analyses, I would like to emphasize that statistical significance alone does not always translate into practical relevance. In particular, when absolute improvements are marginal, the real-world benefit may be limited. The explanation that systematic sampling performs worse after shuffling the dataset is interesting but not entirely clear to me. It would be helpful if the authors could elaborate on why shuffling the sample order causes such degradation in performance.

Q1. Motivation for Subset Sampling

The motivation and justification for CSIE are now clearly articulated. I am satisfied with the explanation and appreciate the attention given to redundancy and imbalance in the dataset.

Q2. Comparison with Full Data Training

The clarification on Figure 4 and the explanation of CSIE’s efficacy relative to full dataset training are appreciated. However, I note that the observed improvements, while statistically valid, remain modest in absolute terms. This somewhat limits the practical impact of the proposed approach. Still, the additional discussion and analysis are helpful.

Overall: The rebuttal effectively clarifies most of my previous concerns, particularly regarding the AMGNN design and the motivation for subset sampling. While I still have reservations about the integration of CSIE and AMGNN and the practical impact of performance gains, I recognize the thoughtful response and additional analysis provided. I will consider raising my evaluation score based on this revised understanding.

Thank you for your thoughtful and constructive feedback. We sincerely appreciate your careful consideration of our responses and your recognition of the clarifications and additional analyses provided. Your comments have helped us better articulate key aspects of our work, and we welcome the opportunity for further dialogue.

W1. Integrated Motivation of CSIE & AMGNN

We sincerely thank the reviewer for their continued engagement and constructive feedback. We now better understand that the core concern lies in the need for a more explicitly unified motivation for CSIE and AMGNN from the outset of the manuscript. We have added a new paragraph to the Introduction, before listing our contributions, to make the joint motivation crystal-clear:

To overcome the practical limitations of AI-driven chemistry, we begin by identifying two fundamental and complementary challenges: (i) theoretical datasets are large-scale but often noisy, redundant, and imbalanced; (ii) experimental datasets, while high in quality and critical for real-world applications, are limited in size and coverage. These challenges jointly motivate a two-stage solution, illustrated through the case study of dissolution free energy prediction: First, we introduce Core Subset Iterative Extraction (CSIE) to systematically reduce redundancy and improve distributional balance in theoretical datasets, enabling more efficient and informative pretraining. Second, we propose the Asymmetric Molecular Graph Neural Network (AMGNN), which captures directional solute–solvent interactions and fine-tunes model performance based on scarce yet reliable experimental data.

This framing shows that CSIE and AMGNN form a unified framework, jointly addressing the gap between theory and experiment rather than standing alone.

W2. Further analysis of the decline in System_Sampling

Thank you for recognizing our explanation and experiments. We now explain in detail why systematic sampling suffers when data order is randomized:

• As we previously mentioned, the effectiveness of systematic sampling in certain chemical datasets is largely due to the inherent ordering of the data.

  • Systematic sampling selects samples at fixed intervals, and its success relies on a key assumption: the original data has a relatively uniform class distribution along its ordering.
  • In some datasets (e.g., Abraham and Compsol, in the context of solvation free energy prediction), experimental data is often compiled in a series, typically organized by solvent type. Systematic sampling can exploit this structure — by tuning the sampling interval, it achieves relatively balanced coverage across categories.

• However, for datasets lacking this kind of ordering structure (such as the bandgap prediction dataset in the appendix or the shuffled versions of the solvation datasets), class distribution along the data order becomes random. In such cases, systematic sampling fails to ensure balanced category representation, leading to significant performance degradation.

• To validate the influence of data ordering, we additionally introduced a shuffled version of the Abraham dataset (denoted as Abraham*), and compared the category distributions selected by CSIE and systematic sampling under a 60% core subset ratio. The results are shown in the table below.

From the table, we observe that systematic sampling exhibits significant category bias on the shuffled Abraham* dataset. The category distribution after sampling is almost identical to that of the original dataset, which confirms our previous point.

Thanks for your thoughtful and constructive feedback. We sincerely appreciate the time and effort you invested in reviewing our manuscript. We are encouraged by your recognition of the principled basis of our subset selection strategy and its empirical effectiveness in bridging theoretical and experimental domains. At the same time, we acknowledge that some aspects require further clarification. Below we provide a detailed, point-by-point response to your comments and suggestions.

W1.&W3. Novelty and Generality of CSIE and AMGNN

Thank you for the valuable comments. While CSIE and AMGNN are two distinct components, they are proposed in tandem to address a core challenge in AI chemistry: the gap between large, imbalanced theoretical datasets and limited, high-quality experimental data. CSIE focuses on data quality and efficiency, while AMGNN targets accurate modeling of molecular interactions. Their integration offers a coherent solution for effectively leveraging both theoretical and experimental data.

We emphasize that both methods have standalone value:

• CSIE is not a typical active learning strategy. It avoids uncertainty-based sampling, which is unreliable in noisy chemical domains, and instead adopts a structure-aware, distribution-optimized selection mechanism. Its generality is validated on a separate task—material band gap prediction (Appendix E.5.1)—demonstrating that it is not limited to dissolution.

• AMGNN introduces a novel asymmetric graph structure to model directional interactions between solute and solvent, addressing a key limitation in prior work that assumed symmetric or simplistic combinations. This is crucial in chemistry, where interaction order significantly impacts properties (e.g., solvation, drug interactions).

Although dissolution prediction is the main benchmark in this paper, it was selected due to the availability of both theoretical and experimental data, making it suitable for evaluating both CSIE and AMGNN. However, the methods are general and can be readily extended to other molecular property prediction tasks, as demonstrated by our additional experiments.

In summary, CSIE and AMGNN are unified by a shared objective, each solving a different but complementary part of the same problem. Their combination is synergistic, but each also contributes meaningful innovation individually.

W2. Sampling quality of CSIE

Thank you for pointing out the problem about CSIE sampling quality.

• First of all, it should be noted that the encoder does affect CSIE's evaluation of samples to a certain extent, which is also a problem we hope to alleviate. For this reason, this paper uses both molecular graph features and semantic features for encoding, and uses the general chemical model MolCA pre-trained on PubChem to alleviate the problem of domain bias.

• Secondly, in Section 3.3 Fig 3, we also visualized the screening effect of CSIE itself on the dataset: we evaluated the sample category distribution of the dataset before and after CSIE screening according to the solvent type. It can be clearly seen that CSIE can indeed alleviate the redundancy problem of the dataset and obtain a more balanced dataset.

Q1. Effects of encoder and subset size on CSIE

• The following table uses AMGNN and CombiSolv-Exp as examples to supplement the impact of pre-trained encoders on CSIE (indicator is MAE). It can be seen that the encoder will change the effect of the subset selected by CSIE to a certain extent, and LLM is significantly better than GNN encoder.

• Regarding the sizes of different core subsets, we provide the experimental results and corresponding analysis on different data sets and different core subset sizes in Figure 4 in Section 3.4 of the main text and Table 15 in Section E.5.2 of the Appendix. We can draw the following conclusion: The larger the core subset, the better the effect. On multiple data sets, smaller core subsets have achieved better results, such as Compsol (80%), FreeSolv (60%), and CombiSolv-QM (5%).

Q2. Ablation experiment

We provide the corresponding ablation experiments in Table 4 on page 9 of our paper:

1. w/o CSIE versus no CSIE+AMGNN: Instead of using CSIE and performing multiple experiments, we use a randomly selected dataset and an AMGNN network.

2. Asymmetric network versus CSIE+standard GNN: We eliminate all interaction edges and use only the basic MPNN mechanism to extract individual molecular information, then obtain the overall bimolecular information through concatenation.

3. Our Method (CSIE+AMGNN) versus CSIE+AMGNN

The comparative results in Table 4 fully demonstrate the respective roles of CSIE and AMGNN. If you would like to add more ablation experiments on other datasets or models to further resolve your confusion, we are also happy to discuss with you and conduct more experimental verification to further improve our paper. Thank you very much for your help!

Q3. The impact of theoretical calculation deviation

Thank you for your concern about this. We will explain from two perspectives that our model is not easily affected by theoretical calculation bias. First, the CSIE framework evaluates information and redundancy based solely on sample characteristics, without considering the impact of theoretical dataset labels. Second, the pre-training task focuses on encoder pre-training, while the task prediction network is carefully fine-tuned on real data.

To verify this, we conducted additional experiments: We randomly varied the calculated results (label values) on the theoretical CombiSolv-QM dataset from 0 to 10%, 20%, 30%, 40%, 50% and presented the transfer learning results on AMGNN.

Generally speaking, for the same sample, the error between theoretical calculation and actual experiment is within ±25%, and the CSIE framework still maintains stable effectiveness. Of course, as random errors are amplified, the encoder will also be affected to a certain extent, resulting in deviations.

Q4. Explanation of some details

Band gap prediction is a single-molecule task. For data screening, CSIE only needs to modify the final bimolecular feature concatenation step and use only single-molecule features for information gain calculation. Regarding the model architecture, since this is a single-molecule task, we only use a simple GCN+MLP for prediction.

Thanks again for your time and effort in reviewing our manuscript. Your feedback is important to our efforts to improve our work. If our response adequately addresses your concerns, we sincerely hope that you will consider revising your rating. We are also happy to provide further clarification or continue the discussion.

Dear Reviewer,

Thanks once again for your careful reading and insightful feedback on our manuscript. We have now submitted a detailed response that we hope addresses your concerns thoroughly.

Specifically, we clarify the independent contributions of CSIE and AMGNN, justify our data and model design choices, and add new empirical results. We hope that these additions will help resolve previous ambiguities.

If there are any remaining questions or suggestions, we would greatly appreciate the opportunity to address them. Otherwise, if you find our response satisfactory, we would be grateful if you could consider updating your review accordingly.

Thank you once more for your valuable comments, which have significantly contributed to improving our manuscript.

Best regards,

The authors

Dear Reviewer 2Vf7,

We sincerely appreciate the time and effort you have devoted to reviewing our manuscript. Previously, we submitted a detailed response addressing all the concerns you raised, including additional experiments, clarifications, and details throughout the paper.

The discussion will end in about two days, but we have not yet received your feedback regarding whether our responses have adequately resolved your concerns so far. We would like to kindly follow up to confirm if our replies have sufficiently addressed your points. If there are still any unresolved issues or areas requiring further clarification, please do not hesitate to let us know—we remain fully committed to making additional improvements to ensure the work meets your expectations.

If our replies have satisfactorily resolved your concerns, we would be truly grateful if you could consider reflecting this in your evaluation by providing a positive score. Your feedback is of great significance to us, and your support would mean a great deal.

Thank you once again for your constructive comments and for contributing to the improvement of our work. We sincerely look forward to your response.

Warm regards,

The Authors

Dear Reviewer,

Thank you very much for your constructive feedback and for the time and effort you have invested in reviewing our manuscript. We are greatly encouraged by your recognition of the principled basis of our subset selection strategy and its empirical effectiveness.

In particular, we appreciate your valuable comments on the impact of encoders and theoretical biases in the dataset on the results of the CSIE framework, which are highly instructive for enhancing the quality of our research.

Thank you again for your support and careful guidance. We look forward to the opportunity to improve our work based on your suggestions and sincerely appreciate that you consider updating your rating of the manuscript.

Sincerely,

The Authors

Thanks for the rebuttal, I have raised my score to 4.

Response to Reviewer 24Q7:

Thank you for your thoughtful and constructive feedback. We sincerely appreciate the time and effort you dedicated to reviewing our manuscript. We are encouraged by your recognition of the theoretical grounding of the CSIE framework, its relevance to data-centric AI for Science, and the empirical performance of both CSIE and AMGNN. At the same time, we acknowledge the need for deeper clarification regarding the chemical relevance of information gain, the contributions of different components (e.g., MolCA vs. asymmetric message passing), and the broader generalizability of our approach. Below, we provide a detailed, point-by-point response to each of your insightful comments.

W1. Visual Analytic

Thank you for pointing out the issue. We explain as follows:

• First, in terms of theoretical foundations, the connection between the information gain metric and chemical properties is primarily established through the use of a more effective encoder. In this study, we employ both molecular graph features and semantic features for encoding. Additionally, we utilize MolCA, a general-purpose chemical foundation model pre-trained on PubChem 324k, to fully extract sample features and ensure the completeness of chemical semantics as much as possible.

• Second, we divide the dataset based on solvent types and evaluate the sample category distribution before and after CSIE filtering. It is clearly observable that CSIE effectively alleviates the data redundancy issue, resulting in a more balanced dataset in terms of class distribution. Detailed results can be found in Section 3.3, Figure 3, along with the corresponding analysis.

W2. Application of CISE in different network models

• CSIE consistently improves model performance across different network architectures: In Table 1 on page 7, the column CombiSolv-Exp → CombiSolv-Exp shows the performance of each model trained directly on real experimental data, while QM-mini → CombiSolv-Exp represents models pretrained on the CSIE-filtered QM-mini dataset and fine-tuned on the real data. A comparison of these two columns reveals that most models exhibit significant performance improvements, demonstrating the effectiveness of CSIE.

• Similarly, Tables 2 and 3 on page 8 present the performance under different solute/solvent type splits, illustrating CSIE's enhancement of model generalization ability. In Table 9 on page 20 (Appendix), we compare the performance gains of CSIE with other core-set selection frameworks across different network models. The results clearly show that CSIE outperforms the alternatives.

W3. Reasons for the performance improvement of AMGNN

• The performance improvement of AMGNN over MMGNN and CGIB is indeed attributed to its asymmetric message-passing mechanism. Under both experimental settings—CombiSolv-Exp → CombiSolv-Exp and QM-mini → CombiSolv-Exp—AMGNN achieves the best performance.

• In the former setting, the model is trained directly on the target dataset without any influence from CSIE, let alone the MolCA module within CSIE. In the latter transfer learning scenario, MolCA only plays a role during the data filtering process; the filtered QM-mini dataset is then fairly used across all models.

• In summary, the performance gains of AMGNN are independent of MolCA.

W4. Discussion on the generalizability of CSIE and AMGNN

We appreciate the reviewer’s concern regarding the generalizability of our proposed method. While the dissolution free energy prediction task serves as a focused case study, both CSIE and AMGNN are designed as general-purpose tools for broader applications in molecular machine learning.

(1) Generalizability of CSIE.
CSIE is a model-agnostic, data-centric framework aimed at improving the quality and efficiency of theoretical datasets. While it draws inspiration from active learning, it specifically addresses limitations in uncertainty-based sampling, which is often unreliable in noisy chemical datasets. To demonstrate its generality, we applied CSIE to a material band gap prediction task (see Appendix E.5.1), which is entirely distinct from dissolution. The performance gains observed there confirm that CSIE is not limited to solubility tasks, but rather serves as a universal dataset screening strategy for scientific machine learning domains where data redundancy and imbalance are common.

(2) Generalizability of AMGNN.

AMGNN proposes an asymmetric molecular graph construction paradigm that directly encodes the directional interactions between two molecules. Unlike traditional methods that treat molecular pairs symmetrically—either through feature concatenation or attention-based fusion (e.g., Uni-Mol, Explainable GNN, CIGIN)—AMGNN captures the fact that molecular interactions are not order-invariant.

For example, as stated in line 173, the solvation energy of ethanol in water differs significantly from that of water in ethanol. This asymmetry is also critical in other molecular relation learning tasks—such as drug–drug interaction

By explicitly capturing these asymmetries through heterogeneous graph modeling, AMGNN can be readily applied to other pairwise molecular tasks such as reaction prediction and toxicity analysis. As discussed in W2 & W3, we believe this paradigm opens a promising direction for more chemically realistic and interpretable molecular interaction learning.

In summary, while solubility serves as a representative task, the methods we propose—CSIE for data selection and AMGNN for asymmetric modeling—are both task-agnostic and widely applicable, offering general value to the molecular machine learning community.

We greatly appreciate your insightful and constructive comments, which have helped us clarify key aspects of our work and strengthen the manuscript. We hope that our point-by-point responses have addressed your concerns, especially regarding the chemical relevance of information gain, the independent contributions of CSIE and AMGNN, and the generalizability of our approach. If our clarifications and additional analysis have resolved the ambiguities you raised, we would be sincerely grateful if you would consider revising your evaluation. Of course, if you have any further questions or suggestions, we would be more than happy to continue the discussion. Thank you once again for your thoughtful review and valuable feedback—it has been instrumental in improving our work.

Dear Reviewer,

We sincerely appreciate the time and effort you dedicated to reviewing our manuscript. Your thoughtful and constructive comments significantly contributed to improving the clarity and rigor of our work.

In our recent response, we have carefully addressed your concerns. We would be grateful to know if our replies have addressed your concerns. If there are any remaining issues or areas where you feel further clarification or improvement is needed, we would be more than happy to provide additional details or make further revisions to strengthen the manuscript.

Thank you once again for your valuable feedback and for helping us improve the quality and clarity of our work.

Best regards,

The authors

Dear Reviewer,

We sincerely appreciate the time and effort you have dedicated to reviewing our manuscript. We recently submitted a detailed response addressing all the concerns you raised. We would like to kindly ask whether our responses have adequately resolved your concerns. If there are any further issues or areas that require additional clarification or improvement, please do not hesitate to let us know. We are more than willing to make further adjustments to ensure the work meets your expectations.

If our replies have addressed your concerns to your satisfaction, we would be grateful if you could consider reflecting this in your evaluation by providing a positive score.

Thank you once again for your constructive feedback and for helping us improve the quality of our work.

Warm regards,

The authors