Self-supervised Blending Structural Context of Visual Molecules for Robust Drug Interaction Prediction

Thank you for your detailed review and constructive suggestions. We address the concerns and questions below.

Q1: What is the justification for accepting this potential loss of critical 3D conformational information compared to using 3D structural data?
Response: We acknowledge that 3D conformations play a critical role in drug interactions. While S$^2$VM does not aim to model full 3D structures, it leverages blended visual fragments to approximate pseudo-spatial proximity between functional groups. This design allows efficient and scalable modeling of local structural co-occurrence patterns. As shown in our interpretability analysis (Figure 5), S$^2$VM consistently highlights biologically meaningful substructures (e.g., 1,3-benzodioxole), indicating its ability to capture structure-dependent features. We also explicitly recognize in the Limitation section that extending to 3D or multimodal representations is a promising future direction.

Q2: Why were 2D images chosen as the input modality over molecular graphs?
Response: We chose 2D molecular images as the input modality because they provide an intuitive and spatially continuous view of molecular structures, making it easier to capture functional group arrangements [1] (Nature Machine Intelligence, 2022). Compared to graphs, images are often more robust—minor errors or missing bonds in molecular graphs may break connectivity, whereas images still preserve chemical context visually. This visual format also facilitates the use of powerful vision models, offering a natural path toward enhancing drug safety analysis through structure-aware attention. Empirically, S$^2$VM outperforms graph-based methods (e.g., SSI-DDI, MRCGNN) across standard and few-shot settings (Table 1–3 and Figure 4 in the main paper), validating the effectiveness of visual representations for DDI prediction.

Q3: Why does the model use two separate projection heads for the two drugs instead of a shared one, given the resulting asymmetry?
Response: The use of two distinct projection heads reflects a deliberate choice to model the inherent asymmetry in drug interactions. In real-world DDI scenarios, drug pairs often exhibit asymmetric roles. For example, one drug may alter the metabolism or concentration of the other. This asymmetric nature is well-documented in biomedical literature and has been emphasized in prior work [2] (Nature Machine Intelligence, 2024), which frames DDI prediction as a direction-sensitive task. An example of the 72-th Event type is: “The serum concentration of Drug2 can be increased when combined with Drug1” where Drug1 acts as the perpetrator and Drug2 as the victim. By assigning separate reconstruction heads to $d_u$ and $d_v$, the model learns role-specific structural representations, which are aligned with the biological reality of DDIs and contribute to more faithful pretraining dynamics. This confirms that explicitly modeling role-specific representations brings tangible benefits and better aligns with the asymmetric nature of real-world drug interactions.

Q4: Since the model uses stochastic blending during training, does this cause significant variance or inconsistency in predictions for the same drug pair during inference?
Response: We clarify that stochastic blending is applied only during pretraining to encourage robust representation learning. For downstream DDI prediction, the model receives deterministic inputs by concatenating the visual tokens of the two drugs in a fixed order, with no randomness involved during inference. This practice is consistent with standard MAE-style reconstruction pretraining, where masking is used only during training, while downstream tasks use the full input [3,4]. Additionally, in Appendix C.7, we report results across different blending strategies for DDI prediction, showing the model’s robustness to fusion variations.

Q5: Was there a rigorous check to prevent data leakage from the PubChem pretraining set into the benchmark test sets?
Response: We confirm that our pretraining process is entirely self-supervised and task-agnostic, with no access to DDI-related labels or annotations. This eliminates any risk of label leakage and follows standard practice in representation learning frameworks such as MAE [3] and MoCo [5], where the pretraining stage is decoupled from downstream tasks.

To assess potential data overlap, we rigorously analyzed the intersection between the 200K molecules used in pretraining and the test sets of benchmark datasets. We identified no shared drugs in the TWOSIDES dataset under the inductive S1 and S2 settings, and only one shared molecule (DrugBank ID: DB01351) in Deng’s and Ryu’s datasets (<0.001%, settings for both drugs are existing). This negligible overlap makes structural memorization extremely unlikely.

To further ensure strict evaluation, we removed all DDI instances involving the shared molecule from the test sets (34 out of 7,474 in Deng’s and 93 out of 38,349 in Ryu’s). As shown in Table 1, the model’s performance remains virtually unchanged, confirming that our results are unaffected by any potential data leakage.

Table 1. Performance comparison after removing the single shared drug from the test sets.

We will clarify this verification and filtering process more explicitly in the revised manuscript.

Q6: The paper claims that the joint reconstruction objective enables the model to learn structural correlations between the two drugs. Was a single-molecule pretraining control experiment considered to validate this claim?
Response: We conducted the suggested control experiment by pretraining a ViT-based masked autoencoder (MAE) using single-molecule reconstruction only. This baseline uses the same architecture, the same 200K PubChem molecules, and identical training epochs as S$^2$VM, ensuring a fair comparison. During downstream DDI prediction, we adopted a post-fusion strategy where each drug is encoded individually and the resulting representations are concatenated for classification.

As shown in Table 2, the single-molecule MAE baseline consistently underperforms S$^2$VM across all metrics and datasets. This demonstrates that joint reconstruction pretraining yields more expressive and interaction-aware representations, beyond what is achievable through standard single-drug encoding.

Table 2. Performance comparison between S$^2$VM and single-molecule representation baselines.

We will add this control experiment and detailed discussion in the revised manuscript to further strengthen the theoretical claims.

Q7: Why does the model use a blended input for DDI prediction, instead of encoding each full molecule separately and combining them later?
Response: The model uses full molecular structures during inference, with stochastic blending applied only in pretraining to enhance representation learning (see response to Q4). Our choice to maintain fusion-before-encoding is based on: 1) Many DDIs arise from localized interactions between functional substructures, not from global structural similarity [2,6]. By blending fragments early, the encoder can directly model cross-molecular spatial dependencies, which are crucial for detecting such interactions. 2) In Appendix C.1 (Table 7), we further compare pre- and post-fusion strategies using molecular fingerprints, again showing the advantage of early fusion. Similarly, in Figure 3, early fusion outperforms a shared-encoder baseline that encodes drugs separately and fuses representations afterward.

These results demonstrate that predicting from fused inputs is not a compromise, but a deliberate and effective design choice that enhances structural interaction modeling. We will make this rationale more explicit in the revised paper.

W1: The method requires high computational resources, which may hinder reproducibility.
Response: The model can be run on machines with >=16GB GPU RAM. To support reproducibility and broader adoption, we provide pretrained checkpoints and code for easy integration and evaluation.

[1] Zeng, Xiangxiang, et al. "Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework." Nature Machine Intelligence (2022)
[2] ZHONG, Yi, et al. Learning motif-based graphs for drug–drug interaction prediction via local–global self-attention. Nature Machine Intelligence, 2024
[3] HE, Kaiming, et al. Masked autoencoders are scalable vision learners. CVPR. 2022.
[4] ZHAO, Fan, et al. Self-supervised feature adaption for infrared and visible image fusion. Information Fusion, 2021
[5] HE, Kaiming, et al. Momentum contrast for unsupervised visual representation learning. CVPR. 2020.
[6] Nyamabo, A. K., Yu, H., & Shi, J. Y. (2021). SSI–DDI: substructure–substructure interactions for drug–drug interaction prediction. Briefings in Bioinformatics, 22(6).

Thank you once again for your valuable feedback, which has helped us improve our work. We hope these clarifications effectively address your key concerns and provide sufficient justification for a higher evaluation.

Thank you for your detailed review and constructive comments. We’re glad you found the idea interesting, the methodology sound, and the experiments robust.

Q1: It would make sense in biology to consider the effect of drugs on proteins and pathways that give rise to DDI, why the paper do not use this data?
Response: We focus on molecular structures from large-scale unlabeled data to address new or under-represented drugs, where protein or pathway data are often missing from biomedical knowledge graphs. Many DDIs, especially those related to drug metabolism such as enzyme inhibition or induction, are may driven by molecular substructures [1] (Nature Machine Intelligence, 2024). This makes structure-level modeling particularly relevant. Our results support this view. While S$^2$VM and MUFFIN perform similarly on common drugs (Table 1), S$^2$VM achieves 22.6% and 32.5% relative gains in rare-event settings on Deng’s and Ryu’s datasets, respectively. This suggests that knowledge-based models may struggle when entity coverage is low, whereas S$^2$VM generalizes more effectively by learning directly from structure.

We acknowledge that integrating biomedical knowledge into S$^2$VM is a valuable future direction and plan to explore it further.

Table 1. The Accuracy of S$^2$VM and MUFFIN on Deng's and Ryu's datasets under different settings.

[1] ZHONG, Yi, et al. Learning motif-based graphs for drug–drug interaction prediction via local–global self-attention. Nature Machine Intelligence, 2024, 6.9: 1094-1105.

Thank you once again for your valuable feedback, which has helped us improve our work. We believe these additions will strengthen our model's presentation.

Thank you for your thoughtful and encouraging review. The point-to-point answer is provided below.

Q1: How are the molecular images generated and kept consistent? Are they 2D diagrams with atoms and bonds, and is any augmentation used?
Response: We generate molecular images through a standardized and reproducible pipeline designed to ensure visual consistency and structural fidelity:

1. Canonicalization of SMILES: All molecules are first canonicalized using RDKit to obtain a unique and deterministic SMILES representation, eliminating variations due to atom ordering or tautomers.
2. Image Rendering via RDKit: The 2D molecular structures are rendered using RDKit’s MolsToGridImage function, which explicitly depicts atoms and bonds. Each molecule is rendered as a 224×224 pixel image without any stochastic augmentation to ensure deterministic and consistent visual representation across different runs.
3. Layout Standardization: We fix all layout-related parameters, including sub-image spacing, drawing style, and molecule alignment, to ensure that chemically identical molecules yield identical image representations.

The full image generation pipeline is available in our codebase for reproducibility. We will clarify this process in the revised manuscript and consider exploring augmentation strategies in future work to enhance generalization.

Q2: How was training on 200 million drug pairs made feasible, and how does model performance scale with dataset size?
Response: We designed S$^2$VM with scalability in mind.

• Pretraining efficiency: To ensure feasibility, we use a lightweight ViT encoder (12 layers, embedding dim 192) and patch size of 16×16, which significantly reduces the computational cost while maintaining performance. We also apply efficient data loading and batching strategies without extensive augmentations (Appendix B).
• Effect of Dataset Size: We conducted systematic experiments to evaluate how model performance scales with pretraining data size. Specifically, we varied the number of base molecules from 50K to 300K, which corresponds to approximately 50M to 300M drug pairs. As shown in Table 1 and Appendix C.4 (Figure 9 and Figure 10 in main paper), we observe consistent performance improvements with larger datasets, particularly from 50K to 200K molecules. Beyond 200K, the marginal gains begin to plateau, suggesting diminishing returns at very large scales.

Table 1. The performance of S$^2$VM on different pretraining data scales.

We will include further discussion and clarification in the revised manuscript.

Q3: The paper presents a theoretical justification for the effectiveness of S$^2$VM. How can this theory guide practical decisions? For instance, does it inform the choice of blending ratio, or the architecture of the decoder?
Response: Our theoretical analysis, which frames S$^2$VM as maximizing mutual information between masked and blended molecular fragments, provides valuable guidance for several design choices.

• Blending ratio: The theory supports the idea that mutual information is maximized when both source distributions contribute sufficiently to the reconstruction target. This insight informed our choice of a balanced blending ratio (e.g., 0.5:0.5), and we validated this empirically in Appendix C.1 (i.e., Table 9 in the main paper), where extreme ratios (e.g., 0.7:0.3) yielded lower performance as shown in Table 2.
• Decoder architecture: The theory emphasizes the need to recover both intra- and inter-molecular dependencies. This motivated us to adopt a lightweight but expressive decoder, sufficient to reconstruct fine-grained structural features from the fused latent space without overpowering the encoder’s role in learning meaningful representations.

We will incorporate a brief discussion of this connection in the revised paper.

Table 2. The performance of S$^2$VM with various mask ratios.

W1: The current method focuses solely on structural features, potentially overlooking pharmacological or biological factors involved in DDIs. Please discuss this trade-off.
Response: S$^2$VM focuses on structural learning to support new or under-annotated drugs, where biological context is often missing. While knowledge-enhanced models like MUFFIN [1] perform well on common drugs, they rely on entity coverage and may struggle with rare or unseen cases. Our comparison in Table 3 shows that S$^2$VM achieves comparable results on common settings, but significantly outperforms MUFFIN on rare events with up to 32.5% higher accuracy. This highlights that structure-driven self-supervised learning offers stronger generalization under low-resource scenarios. We will elaborate on this trade-off in the revised manuscript and consider integrating knowledge graphs into S$^2$VM in future work.

Table 3. The Accuracy of S$^2$VM and MUFFIN on Deng's and Ryu's datasets under different settings.

W2: The blending process lacks clarity. How are fragments selected and positioned, and could a visual example help replication?
Response: We adopt an anchor-based replacement strategy: one molecule (e.g., $d_u$) serves as the structural anchor, and a fixed proportion of its visual fragments (e.g., 30%) are randomly masked. The corresponding fragments from the second molecule $d_v$ are then inserted at the same positions, forming the blended input. This leads to an effective fragment ratio of 0.7:0.3 between $d_u$ and $d_v$ in the final representation. This strategy ensures spatial alignment while introducing meaningful structural variation across training pairs. We will include a visual illustration of this blending process in the revised paper to enhance clarity and reproducibility.

[1] CHEN, Yujie, et al. MUFFIN: multi-scale feature fusion for drug–drug interaction prediction. Bioinformatics, 2021, 37.17: 2651-2658.

Thank you again for your valuable insights, which will help us further strengthen the presentation and impact of our work.

Thank you for your detailed review and constructive comments. We address the concerns and questions below.

Q1: Could the authors share their thoughts on why a JEPA-based approach was not considered to better understand their architectural choices?

Response: We appreciate your insightful suggestion and interest in JEPA-style architectures. Our initial choice of a ViT-based encoder–decoder design was motivated by its demonstrated success in visual molecular representation learning [1] and its compatibility with pixel-level reconstruction, which aligns well with our objective of capturing fine-grained spatial correlations between drug fragments.

That said, we also agree that JEPA is a compelling alternative for structure-level pretraining. To explore this direction, we implemented an image-level JEPA (I-JEPA [2]) variant of S$^2$VM, trained under the same settings and scale (50K base molecules, 50M drug pairs) for fair comparison. As shown in Table 1, I-JEPA achieves comparable performance on the Ryu dataset and slightly underperforms on the Deng dataset. These results suggest that JEPA-style models can capture complementary information and may serve as a promising backbone for future extensions of S$^2$VM.

We will include these comparative results and release the I-JEPA-based pretrained checkpoints in our revised manuscript. Due to time constraints, we have evaluated I-JEPA only on the small-scale setup, but we plan to further scale up the training to fully assess its potential.

Table 1. Performance (Macro-F1) comparison between ViT-based and JEPA-based S$^2$VM models under small-scale pretraining.

Q2: I encourage the authors to report downstream performance as a function of pretraining dataset size. This would help quantify the benefit of large-scale pretraining, which seems underexplored in the current paper.

Response: We have conducted additional experiments to evaluate the impact of pretraining dataset size on downstream DDI performance, as detailed in Appendix C.4 (Figures 9 & 10).

As shown in Table 2, we observe that S$^2$VM benefits consistently from scaling up the number of pretraining drug pairs. The Macro-F1 scores increase steadily from 50M to 300M molecular pairs, with diminishing returns beyond 200M. Based on this trend and computational considerations, we selected 200K base molecules (yielding ~200M drug pairs) as our default setting.

Table 2. Macro-F1 scores of S$^2$VM on Deng's and Ryu's datasets across different pretraining data scales.

W1: The theoretical analysis in Section 4.2 lacks clarity. A concise summary in the main text and clearer notation in the proof would improve readability.
Response: We will add a concise summary of the theoretical insight directly in the main text, rather than deferring entirely to the Appendix. Specifically, we will highlight the core idea as follows:

Our training objective maximizes the mutual information between the blended visible fragments and the original unobserved parts of each molecule, thereby encouraging the model to learn both intra- and inter-molecular dependencies critical for DDI prediction.

Meanwhile, we will:

• Explicitly clarify that $X_1, X_2$ denote the visible fragments from drug A and drug B (e.g., $A_1, B_2$) used as input;
• $Y$ denotes the reconstruction target consisting of the missing parts (e.g., $A_2, B_1$);
• The mutual information term $I(X_1, X_2; Y)$ thus quantifies how well the latent representation captures information necessary to reconstruct the original molecules.

We will align the theoretical notation with the data flow used in the model (e.g., using $A_1, A_2, B_1, B_2$) and briefly explain how the decomposition supports the design of our fusion-before-encoding framework.

[1] XIANG, Hongxin, et al. A molecular video-derived foundation model for scientific drug discovery. Nature Communications, 2024, 15.1: 9696.
[2] ASSRAN, Mahmoud, et al. Self-supervised learning from images with a joint-embedding predictive architecture. CVPR. 2023. p. 15619-15629.

We sincerely appreciate your thoughtful suggestions, which will help us further strengthen the impact of our work. We hope our responses have effectively addressed your concerns and provided the necessary evidence to support a more favorable assessment.