W1&Q1: The authors should more clearly articulate their specific methodological contributions and the overall impact of their approach.

Thanks for your valuable comment. Current drug recommendation methods fall into two categories: (i) those relying on historical patient information (e.g., COGNet) and (ii) those leveraging drug substructures (e.g., SafeDrug, MoleRec). The former suffers from the complexity and sparsity of longitudinal EHRs, leading to high computational costs and difficulty in capturing transferable patterns, with COGNet showing the highest memory usage. GameNet also uses historical data but relies on lookup-based matching with dynamic memory, which is beyond deep learning. To address these issues, SubRec uses a vector quantization (VQ) module to cluster heterogeneous patient–drug interactions into a compact codebook, reducing modeling instability and enabling efficient similarity matching.

Our approach is fundamentally different from WISER (ICML 2024). , which learns discrete embeddings for individual drugs and predicts via weighted combinations, our model learns patient–drug association vectors. These vectors are randomly initialized and refined during training (Eq. 18), converging to cluster centers of patient–drug interactions (Fig. 6). This design captures interaction-level patterns rather than drug-level embeddings, enabling more personalized and context-aware recommendations.

The latter, represented by methods such as SafeDrug and MoleRec, uses rule-based decomposition (e.g., BRICS) to split molecules, which disrupts functional group connectivity and fails to capture the synergistic effects of multiple substructures. While clinicians rely on medical knowledge, molecular structure remains a key determinant of drug pharmacodynamics and mechanism of action [1,2]. Functional substructures (pharmacophores) drive therapeutic effects—for example, simvastatin and atorvastatin share an HMG-CoA reductase–inhibiting pharmacophore [3]—and also influence adverse drug interactions through shared or reactive motifs [4,5]. To address these factors, we incorporate an improved CIB to extract patient-specific core substructures, while the VQ module enforces a discrete latent space that preserves only task-relevant information. This design enables SubRec to balance accuracy, robustness, and interpretability, overcoming the limitations of prior approaches.

Indeed, existing works in this field, especially SafeDrug[1] and MoleRec [2], do not provide explicit demonstrations of the importance of drug substructures. Therefore, our original manuscript did not include such visualizations. In the revised version, following your advice, we have added drug substructure visualization results, which clearly show that our method captures chemically coherent and clinically relevant motifs. Due to policy restrictions, we are unable to include these figures in the rebuttal, and we kindly ask for your understanding.

[1] Why drugs fail—a study on side effects in new chemical entities. Nat. Rev. Drug Disc. 2005

[2] Pharmacophore modeling in drug discovery. Nat. Rev. Drug Disc. 2016

[3] Large-scale prediction and testing of drug activity on side-effect targets. Cell. 2012

[4] Learning motif-based graphs for drug–drug interaction prediction via local–global self attention. Nat mach Intell. 2024

[5] Emerging drug interaction prediction enabled by a flow-based graph neural network with biomedical network. Nat Comp Sci. 2023

W2: A more comprehensive explainability analysis to identified relevant substructures.

Thanks for your suggestions. However, existing works in this field, especially SafeDrug[1] and MoleRec [2], do not provide explicit demonstrations of the importance of drug substructures. Therefore, our original manuscript did not include such visualizations. In the revised version, following your advice, we have added drug substructure visualization results, which clearly show that our method captures chemically coherent and clinically relevant motifs. Due to policy restrictions, we are unable to include these figures in the rebuttal, and we kindly ask for your understanding.

W2&L3: Clarify the techniques contribution for drug structures nor the clinic patients.

Thank you for the thoughtful comment. We apologize for the misunderstanding—each module in our framework is carefully designed with domain-specific considerations for both molecular drug structures and clinical patient data.

• For drug substructure extraction, prior methods (e.g., BRICS) often break molecular connectivity, ignoring the contextual impact of patient conditions. In contrast, we propose a CIB framework that adaptively extracts substructures conditioned on patient-specific health profiles, taking comorbidities and personalized treatment patterns into account.

• Furthermore, motivated by the observation that similarity in current patient visits is often more informative than full-history similarity, we introduce a VQ module to store representative patient–drug interaction prototypes. This module acts as a compact, reusable memory that enhances recommendation in data-limited settings. The VQ also avoids the complexity and instability of high-dimensional variational modeling, making the framework lighter and more stable during training [1,2]. In addition, the discrete latent structure naturally aligns with the CIB principle by encouraging representations that retain only task-relevant information [3].

Overall, we present an advanced drug recommendation framework that achieves precise substructure-level reasoning under the guidance of well-designed CIB theory. The VQ module further condenses patient visit information into discrete, informative representations that support accurate and efficient recommendation. We carefully balance the contributions of each component to ensure high predictive performance while maintaining a low DDI rate.

Safe and accurate drug recommendation is an emerging and high-impact research direction with the potential to revolutionize pharmaceutical decision-making. Our work not only focuses on algorithmic optimization but also demonstrates how advanced machine learning techniques can drive real progress in healthcare and inspire future research in this critical domain.

In addition, in the revised version, we have also added test results in real scenarios and interpretable analysis of special pathological groups. Due to the limited characters, we hope to discuss and communicate with you in the next rebuttal process. Hope you can understand.

[1] Mole-BERT.ICLR, 2023. [2] Learning Invariant Molecular Representation in Latent Discrete Space. NIPS, 2024. [3] MAPE-PPI. ICLR, 2024.

W3&Q3: What is the underlying relationship between the CIB-derived subgraphs and the VQ module?

Thanks for your insightful question! specifically: 1) VQ enables end-to-end learning of discrete prototypes, which serve as interpretable and stable anchors for patient subtypes. 2) The quantization process avoids the complexity and instability of modeling variational distributions in high-dimensional latent spaces, making the model lighter and more stable during training[1,2]. 3) The discrete structure of the latent space naturally supports the CIB principle by enforcing a compact representation that preserves only the most task-relevant information for drug prediction[3].

Q1: How to handle the missing data?

Thank you for your kind reminder. In the data preprocessing stage, we follow the same strategy as SafeDrug, MoleRec. For missing data (e.g., NDC codes), we apply forward filling to propagate the last valid entry and remove rows with critical missing values to ensure data completeness. We will clarify this processing step explicitly in the revised version.

Q2&L1: The authors omit a specific loss function in Line 164?

The omission of the term $I\left(\mathbf{Y};\mathbf{e}^{(i)}\right)$ is a common practice in related work. Empirically, explicitly optimizing this term often leads to instability and performance degradation. So, some advanced work would omit it in the formulation[1,2].

[1] Conditional Graph Information Bottleneck for Molecular Relational Learning. ICML,2023

[2] Iterative substructure extraction for molecular relational learning with interactive graph information bottleneck. ICLR, 2025

Q3: The objectives described in Equations 8 and 9 are somewhat unclear.

We have prepared relevant proof, but the characters are limited. We hope to communicate with you further and present it to you.

Q4: What is the motivation for using discrete embeddings in this context?

Please refer to the responses of W1 and W3.

Q5: Have the authors considered incorporating gene expression or mutation data in their framework for drug recommendation?

Thank you for this valuable suggestion. We agree that integrating gene expression or mutation data could provide a richer biological context for drug recommendation, as these factors play critical roles in drug response and personalized treatment. However, our current work focuses on modeling EHR data and drug molecular structures, as publicly available datasets combining comprehensive clinical records with matched genomic profiles are still limited and this is not within the scope of our research.

L2: The absence of publicly available code.

The link is at the end of the abstract.

---

Due to the characters are limited, and we look forward to presenting it to you in the next rebuttal process. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score.

Thanks for the detailed comment. While I appreciate the effort of authors in addressing my comment. They are still largely unaddressed.

• My primary concern is around equation 9 in the draft. Since the authors have claimed that they have proof for this. I assume that authors are indeed trying to minimize the conditional log likelihood and have some mathematical reasoning around this. I would like to understand any intuitive reasoning behind this. What I was expecting in the paper was something like eq 13 (negative log likelihood) in [1]. I would like authors to clarify any misunderstanding on my part.

• Authors have claimed that the embeddings learn the drug-patient interaction. However, there is no empirical evidence apart from final results to validate that.

[1] Conditional Graph Information Bottleneck for Molecular Relational Learning, ICML 2023.

I am very pleased to receive your reply. Here, I am providing the proof for Eq. 9 to better address your concern.

Q1: Equation 9's Proof.

The proof are provided as: $-I(Y; \mathcal{G}_{sub}, e^{(i)}) = -\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right]$

Since the true conditional distribution $p(Y \mid X)$is intractable, we introduce a variational approximation $p_\theta(Y \mid X)$.

Then:$-\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right] = -\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \cdot \frac{p(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})} \right]= -\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right] - \mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})} \right]$

\left[ \log \frac{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right] - \mathbb{E}\_{\mathcal{G}\_{sub}, e^{(i)}} \left[ D\_{KL} \left( p(Y \mid \mathcal{G}\_{sub}, e^{(i)}) \| p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)}) \right) \right]$$ By the non-negativity of the KL divergence, we have: $$\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right] \leq -\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right]$$ Continuing, we can rewrite the right-hand side as: $$-\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log \frac{p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)})}{p(Y)} \right] \\ = -\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)}) \right] + \mathbb{E}\_Y [ \log p(Y) ] \\ = H(Y) - \mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)}) \right]$$ where $H(Y) = -\mathbb{E}\_Y [\log p(Y) ]$ is the entropy of $Y$. $-\mathbb{E}\_{Y, \mathcal{G}\_{sub}, e^{(i)}} \left[ \log p\_{\theta}(Y \mid \mathcal{G}\_{sub}, e^{(i)}) \right]$ is calculated by the prediction loss in Eq.10. $\mathcal{L}\_{\text {pred }}\left(\mathbf{Y}, \mathcal{G}\_{\text {sub }}, \mathbf{e}^{(i)}\right)$where $\mathbf{z} = \operatorname{pool}(\mathbf{H}\_{\text{sub}})$ is the graph-level embedding of $\mathcal{G}$, obtained from $\mathbf{H}\_{\text{sub}}$ of the subgraph $\mathcal{G}_{\text{sub}}$, and $f$ denotes the prediction head. $H(Y)$ is constant value that can be ignored.

Dear Reviewer,

Thank you very much for your reply. We are delighted to hear that we have addressed most of your concerns. For the remaining two issues, we would like to clarify as follows:

• Q1: You are absolutely right that this was a typographical error. We appreciate you pointing it out, and we will correct it in the main text and add the proof in the Appendix.

• Q2: After your response, we now have a clearer understanding of the issue. You suggested analyzing the similarity between embeddings for subgroups of patients and drugs, which we have indeed already shown in the manuscript. Specifically, Figure 1 illustrates our hypothesis: Patients with more similar diagnoses are more likely to be prescribed the same drugs by doctors. We performed a cosine similarity analysis between diagnosis and multi-hot encoded medications in the original MIMIC-III data. Please refer to Appendix E.2 and Figure 4 (a) for the details.

Additionally, we have now included results on the similarity between patient diagnoses and recommended drugs to further illustrate this point. The results are summarized in the table below, where we can clearly see that the higher the similarity between diagnosis results, the higher the overlap in drug recommendations:

---

If you have any further questions, please don’t hesitate to contact us. We are confident that we can resolve any remaining concerns and meet all of your expectations. Once again, we sincerely thank you for your valuable contributions to improving the quality of our manuscript. If our revisions have addressed your concerns, we kindly ask you to reconsider your evaluation and score.

Thank you again, and we wish you a wonderful day.

Best regards,

Authors

Q2 clarification

Thank you for the clarification regarding my second question. While I understand the intuition and reasoning the authors are trying to convey, I personally do not consider replacing the learned vector with noise to be a reliable method for evaluating whether it encodes medically relevant information. There could be several out-of-distribution effects or confounding factors contributing to the poor output, which this approach does not account for.

As an alternative, the authors might consider analyzing the similarity between embeddings for subgroups of patients and drugs (as illustrated in Figure 1) as a potentially more meaningful way to validate their intuition.

Thanks for the clarification. I noticed that the equation derived in the proof differs from the one presented as Equation (9) in the paper. Specifically, for subgraph ($\mathcal{G}$) and variational prob (p) with parameters (${\theta}$) the proof gives:

$H(Y) - \mathbb{E}_{Y, \mathcal{G}, e^{(i)}} \left[ \log p(Y \mid \mathcal{G}, e^{(i)}) \right]$

whereas Equation (9) in the paper states:

$H(Y) + \mathbb{E}_{Y, \mathcal{G}, e^{(i)}} \left[ \log p(Y \mid \mathcal{G}, e^{(i)}) \right]$

I believe this may be a typo in the paper. The derivation in the proof follows the standard form of variational approximation for subgraphs, and my question was primarily about the discrepancy in the final reported equation.

Response to Reviewer 3FcV:

Thank you very much for your valuable comments. We are deeply encouraged by your recognition of our paper's motivation, central theme, and sufficient experimental results. Below, we provide point-by-point responses to your suggestions and concerns.

W1&Q1: Lake computational complexity and resource requirements analysis.

Thank you for your suggestion, we have added relevant efficiency experiments. While our method does introduce some additional cost compared to lightweight baselines, we would like to emphasize that the computational overhead is manageable and can be effectively mitigated. Importantly, we found that this issue can be addressed in several practical ways:

• By applying an early stopping strategy, we observed more than 2× improvement in efficiency with minimal impact on performance.
• In a dedicated GPU setting, the training time can be reduced to approximately 7 hours, making the approach much more feasible in practice. Additionally, we believe that further optimization strategies, such as pruning or model distillation, could be applied to significantly reduce runtime, without sacrificing performance.

W2: Additional ablation studies are needed to be added.

Thank you for your helpful suggestion. In the revised version, we have conducted the requested ablation experiments, which are reported in Table 2. Specifically, the ablation of the substructure extraction module is denoted as “w/o substructure capture”, while the ablation of the VQ component is denoted as “w/o auxiliary codebook”. Furthermore, we performed an additional analysis on the effect of varying the codebook size, and the results are presented in Table 4.

W3&Q3: What is the underlying relationship between the CIB-derived subgraphs and the vector quantization (VQ) module?

Thanks for your valuable suggestion. Actually, we are motivated by the observation that similarity in current patient visits is often more informative than full-history similarity, we introduce a VQ module to store representative patient–drug interaction prototypes. This module acts as a compact, reusable memory that enhances recommendation in data-limited settings. The VQ also avoids the complexity and instability of high-dimensional variational modeling, making the framework lighter and more stable during training [1,2]. In addition, the discrete latent structure naturally aligns with the CIB principle by encouraging representations that retain only task-relevant information [3].

W4&Q2: How does SubRec perform if the noise injection is removed?

We have conducted the requested ablation by removing the noise injection. The results show training becomes unstable, and the performance variance significantly increases. This supports the necessity of the noise injection component for robust optimization.

From a theoretical perspective, we also provide the proof analysis.

• Without noise injection, the G_\{IB} selection objective reduces to: \min_{G_\{sub}}-I\left(G_\{sub}, Y\right)+\beta I\left(G_\{sub}, G\right) Let $G \in \mathbb{G}$ and $Y \in \mathbb{R}$ be the graph and its label．G_\{n} \in \mathbb{G} is the label-irrelevant substructure，independent to $Y$. If G_\{n} influences G_\{sub} only through $G$ , then：

I\left(G_\{sub}, G_n\right) \leq I\left(G_\{sub}, G\right)-I\left(G_\{sub}, Y\right)

When $\beta=1$ ,the objective encourages G_\{sub} to be minimally informative about G_\{n}. This inequality is supported by the theoretical results in [1].

• With noise injection, let $G_\epsilon \in \mathbb{G}$ is the injected noise and G_\{N} the resulting noised graph. We consider selecting G_\{sub}​ by filtering out $G_\epsilon$​ from G_\{N} ​, and show:

I\left(G_\{sub}, G_n\right) \leq I\left(G_N, G_n\right) \leq I\left(G_N, G\right)-I\left(G_N, Y\right)

Suppose $G, G_N, G_n$ and $Y$ satisfy the Markove condition $\left(Y, G_n\right) \rightarrow G \rightarrow G_N$. Then:

$$I\left(G_N ; G_n\right) \leq I\left(G_N ; G\right)-I\left(G_N ; Y\right)$$

$G_N$ is a deterministic function of $G_\epsilon$ and $G_{sub}$, since we can obtain $G_{sub}$ by filtering out $G_\epsilon$ from $G_N$. Then, we have:

$$\begin{aligned} I\left(G_N ; G_n\right) & =I\left(G_\epsilon, G_{sub} ; G_n\right) \\ & =I\left(G_{sub} ; G_n\right)+I\left(G_{sub} ; G_\epsilon \mid G_n\right) \\ & \geq I\left(G_{sub} ; G_n\right) \end{aligned}$$

Therefore:

$$I\left(G_{sub} ; G_n\right) \leq I\left(G_N ; G\right)-I\left(G_N ; Y\right)$$

Noise injection also facilitates a tractable variational upper bound for optimization (Eq. 16).

[1] Improving Subgraph Recognition with Variational Graph Information Bottleneck. CVPR. 2022

---

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.

Response to Reviewer Y23r:

Thank you very much for your valuable comments. We are sincerely encouraged by your recognition of our motivation, technical design, and experimental results. Below, we provide point-by-point responses to your suggestions and concerns.

W1&Q1: Limited Technical Novelty?

Thank you for the thoughtful comment. Each module in our framework is carefully designed with domain-specific considerations for both molecular drug structures and clinical patient data. We acknowledge that prior works, such as SafeDrug [1] and MoleRec [2], also adopt substructure-based modeling and consider DDI safety. However, SafeDrug and MoleRec rely on rule-based fragmentation (e.g., BRICS), which disrupts the natural connectivity of functional groups and fails to capture the synergistic effects of multiple substructures. In contrast, our CIB module learns condition-specific core substructures in a data-driven manner, conditioned on patient health context, ensuring that the extracted motifs are clinically relevant and task-dependent. Furthermore, motivated by the observation that similarity in current patient visits is often more informative than full-history similarity, we introduce a VQ module to store representative patient–drug interaction prototypes. This module acts as a compact, reusable memory that enhances recommendation in data-limited settings. The VQ also avoids the complexity and instability of high-dimensional variational modeling, making the framework lighter and more stable during training [1,2]. In addition, the discrete latent structure naturally aligns with the CIB principle by encouraging representations that retain only task-relevant information [3].

Overall, SubRec is a framework that unifies condition-aware substructure extraction and discrete patient modeling under a principled architecture. This synergy allows SubRec to achieve SOTA performance while providing chemically and clinically meaningful interpretability, which has not been addressed by prior works.

[1] Mole-BERT: Rethinking Pre-training Graph Neural Networks for Molecules. ICLR, 2023.

[2] Learning Invariant Molecular Representation in Latent Discrete Space. NIPS, 2024.

[3] Conditional Graph Information Bottleneck for Molecular Relational Learning. ICML, 2024.

W2: Insufficient Validation of Model Interpretability.

Indeed, existing works in this field, especially SafeDrug[1] and MoleRec [2], do not provide explicit demonstrations of the importance of drug substructures. Therefore, our original manuscript did not include such visualizations. In the revised version, following your advice, we have added drug substructure visualization results, which clearly show that our method captures chemically coherent and clinically relevant motifs. Due to policy restrictions, we are unable to include these figures in the rebuttal, and we kindly ask for your understanding.

W3: Lack of Pharmacological Alignment, provide concrete examples of the extracted substructures.

Following your suggestions，indeed, existing works in this field, especially SafeDrug[1] and MoleRec [2], do not provide explicit demonstrations of the importance of drug substructures. Therefore, our original manuscript did not include such visualizations. In the revised version, following your advice, we have added drug substructure visualization results, which clearly show that our method captures chemically coherent and clinically relevant motifs. Due to policy restrictions, we are unable to include these figures in the rebuttal, and we kindly ask for your understanding.

Regarding pharmacological alignment, we have addressed your concern from two perspectives. First, we present results showing that our model recommends similar drugs for patients with similar diagnostic codes, reflecting alignment with clinical prescribing practices (Table 6). Second, we provide real case studies of the recommended drugs, demonstrating that the extracted substructures correspond to known pharmacological mechanisms. These additions reinforce our claim that the proposed method offers clinically interpretable and pharmacologically meaningful predictions.

With regard to the pharmacological alignment analysis of molecular substructures, current research has rarely considered this issue, and thus there are not enough real-world cases for analysis. However, to further address your concern, we have re-adjusted our model to specifically target the prediction of metabolism-mediated diseases, while focusing on the active regions of drugs. Since this is an approximate analysis, we conducted testing in the form of drug pair inputs. This setting allows us to better examine whether the extracted substructures correspond to the regions where the drugs exert their effects. Specifically, we utilized a curated dataset of 73 chemicals (perpetrators) known to inhibit metabolic enzymes via specific functional groups, leading to metabolism-mediated DDIs. These chemicals formed 13,786 pairs with other drugs. Using SubRec, we conducted interpretability analysis by comparing the substructures identified by our model with the enzyme-inhibition functional groups reported in the literature.

[1] Learning motif-based graphs for drug–drug interaction prediction via local–global self attention. Nature machine intelligence. 2024.

W4: Charts need to be optimized.

Thank you for this helpful suggestion. In the revised version, we have optimized the figure 2 design by enhancing the color contrast to improve visual clarity and ensure a more intuitive understanding for readers.

Q2: Lack of citations to relevant literature.

In response to your comments, we have added relevant references.

Q3: Does the specific strategy for constructing the codebook in the vector quantization module shift significantly due to changes in the training set?

Thank you for this insightful comment. In our design, the codebook in the vector quantization module is constructed to capture shared latent patterns across the training set, clustering heterogeneous patient–drug interactions into a compact set of prototypes. This design inherently leverages the semantic similarities present across patients and drug interactions, rather than overfitting to individual samples.

Whether the codebook shifts significantly depends primarily on the distribution of the training data. When new training sets share similar distributions with the original data, the learned codebook remains largely stable, as the latent prototypes continue to capture the dominant interaction patterns. Even when novel or previously unseen conditions appear, the codebook still provides meaningful representations because it encodes generalizable structures (Eqs.16-18). Furthermore, its impact is jointly optimized with the overall learning objectives, rather than acting as an isolated component, which further mitigates sensitivity to data shifts.

Therefore, while extreme distributional changes could lead to adjustments in the learned prototypes, under typical conditions the codebook remains stable and continues to support robust and efficient modeling without requiring frequent reconstruction.

---

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.

Response to Reviewer Ch3Q:

W1&W2: Substructure-based modeling has already been explored in prior work.

Thank you for this comment. We acknowledge that SafeDrug [1] and MoleRec [2] are important prior works in substructure-based DDI modeling. However, these approaches rely on rule-based decomposition (like BRICS) of drugs, which may disrupt the intrinsic connectivity of functional groups, as illustrated in Figure 1. Drug activity usually arises from the synergistic effect of multiple substructures rather than isolated fragments. Moreover, meaningful substructure extraction should also account for patient-specific factors, such as comorbidities, which influence prescription decisions (Line 64–72).

To better clarify our contributions, we have added drug substructure visualization results in revised version demonstrating that our method captures chemically coherent and clinically relevant motifs. Due to policy restrictions, we are unable to include these figures in the rebuttal, and we appreciate your understanding.

W1: How SubRec overcomes the challenges.

Thank you for your valuable comment. Current research on drug recommendation can generally be divided into two categories: (i) methods that rely on historical patient information, such as COGNet, and (ii) methods that leverage drug substructures, such as SafeDrug and MoleRec. The former often suffers from the complexity and sparsity of longitudinal EHRs, resulting in high computational costs and difficulty in capturing transferable patterns (as shown in the following table). In particular, COGNet, as a typical historical information–based model, exhibits the highest memory consumption. GameNet is another early approach relying on historical data, but it primarily performs lookup-based matching with dynamic memory in key–value form, which falls outside the scope of deep learning models. To address these challenges, SubRec employs a vector quantization (VQ) module to cluster heterogeneous patient–drug interactions into a compact codebook, mitigating the instability of variational modeling and enabling efficient similarity matching.

The latter, represented by approaches such as SafeDrug and MoleRec, typically uses rule-based method (e.g., BRICS) to split molecules into fragments, which disrupts the natural connectivity of functional groups and fails to reflect the synergistic effect of multiple substructures. Here, we would like to emphasize that while it is true that clinicians typically rely on medical knowledge and pharmacological guidelines, the molecular structure remains a fundamental determinant of a drug’s pharmacodynamics and mechanism of action[1,2]. 1) Many drugs exert their therapeutic effects through functional substructures, or pharmacophores. For example, simvastatin and atorvastatin share a common pharmacophore that inhibits HMG-CoA reductase, making both effective for lowering cholesterol[3]. 2) Moreover, the latest research has present that substructures also contribute significantly to potential adverse drug interactions. Shared or reactive substructures can lead to overlapping metabolic pathways or toxicity profiles, which are crucial in assessing the safety of drug combinations[4,5]. Therefore, we incorporates an improved CIB to extract core drug substructures conditioned on patient health context, ensuring that the learned representations remain both concise and clinically relevant. The VQ mechanism further supports CIB by enforcing a discrete latent structure that preserves only the most task-relevant information. Through this design, SubRec overcomes the limitations of prior methods, achieving a better balance between predictive accuracy, robustness, and interpretability.

Indeed, existing works in this field, especially SafeDrug[1] and MoleRec [2], do not provide explicit demonstrations of the importance of drug substructures. Therefore, our original manuscript did not include such visualizations. In the revised version, following your advice, we have added drug substructure visualization results, which clearly show that our method captures chemically coherent and clinically relevant motifs. Due to policy restrictions, we are unable to include these figures in the rebuttal, and we kindly ask for your understanding.

[1] Why drugs fail—a study on side effects in new chemical entities. Nat. Rev. Drug Disc. 2005 [2] Pharmacophore modeling in drug discovery. Nat. Rev. Drug Disc. 2016 [3] Large-scale prediction and testing of drug activity on side-effect targets. Cell. 2012 [4] Learning motif-based graphs for drug–drug interaction prediction via local–global self attention. Nat mach Intell. 2024 [5] Emerging drug interaction prediction enabled by a flow-based graph neural network with biomedical network. Nat Comp Sci. 2023

W2: The complexity and sparsity of patient historical records.

Thank you for your insightful comment. We would like to further elaborate on how our framework addresses the complexity and sparsity of patient historical records.

First, regarding complexity, as noted in our response to the previous question, modeling the full visit history of patients can lead to significant memory consumption, especially in deep learning frameworks. To mitigate this, we designed a VQ module to encode heterogeneous patient–drug interactions into a compact codebook. This representation reduces computational overhead while retaining valuable information from historical visits.

Second, concerning sparsity, our dataset is split based on patients, but the number and frequency of visits vary significantly. Some patients have rich historical records, while others have only a few visit. This inconsistency introduces sparsity into the training process. In the extreme case of one-shot scenarios (i.e., patients with only one visit), we evaluate model performance and report the results in Figure 4, with detailed descriptions in Appendix E.3.

In addition, we considered a single-drug recommendation scenario, where each visit results in only one prescribed drug. This setting provides limited contextual information from longitudinal history. Our evaluation on the MIMIC-III dataset shows that models heavily dependent on patient history (e.g., COGNet) perform worse under this constraint. In contrast, structure-based models, MoleRec and SafeDrug, maintain superior performance, as they can capture intrinsic associations between molecular substructures and clinical conditions. The GIB module further strengthens this by extracting condition-specific drug substructures, allowing SubRec to remain effective even in sparse or low-information settings, and outperforming other baselines.

W3: What are the key differences between the SubRec and GIB (NIPS 2020) ?

Recently, information bottleneck (IB) theory (Arxiv., 2000) has been applied to learning significant subgraphs of the input graph for explainable GNNs (NIPS 2020, CVPR 2022; 2020; ICML 2022), which provides a principled approach to determine which aspects of data should be preserved and which should be discarded (AAAI 2021). However, directly applying GIB in the context of drug recommendation is non-trivial. Traditional GIB-based methods learn the core subgraph solely from the input graph itself, while in drug recommendation, the task requires considering patient-specific conditions to recommend optimal drug combinations. Therefore, our framework learns a condition-specific core subgraph $\mathcal{\widetilde{G}}_{\mathrm{IB}}$ that maximizes the mutual information between drug substructures and the prediction target $Y$ (Eqs. 6–7).

Moreover, our approach aims to go beyond existing GIB applications by integrating precise substructure-level reasoning with longitudinal patient records in an organic manner. This joint modeling allows SubRec to leverage both molecular structural knowledge and patient-specific clinical context, carefully balancing their contributions to achieve high predictive performance while simultaneously maintaining a low DDI rate.

W4: How the CIB is integrated into the proposed framework.

The framework models patient representations for prediction from three complementary aspects. First, by leveraging drug substructure information and the conditional drug CIB mechanism, the model focuses on extracting condition-specific core features from drug substructures while suppressing irrelevant information. This enables the generation of targeted molecular representations for all candidate drugs, thereby improving prescription accuracy. Second, by incorporating historical information, the framework employs a VQ module to construct a compact codebook that reduces complexity and efficiently encodes heterogeneous patient–drug interactions. Finally, through the interaction between condition-aware substructures and quantized patient states, the model produces reliable recommendation outcomes while maintaining a low DDI rate.

W5: Lack of Comparisons with Recent SOTA Methods.

Thank you for your suggestion. We have added the mentioned baseline model (Jaccard Index) in the revised version. The results show that our method maintains a clear advantage on the MIMIC-III dataset, while on MIMIC-IV it performs slightly lower than BSP, SeqCare, and HAR.