CIDD: Collaborative Intelligence for Structure-Based Drug Design Empowered by LLMs

We sincerely thank the reviewer for their thoughtful feedback. Below are our responses to the main points:

Regarding Code

We apologize for not including the code in the initial submission. Due to file-upload restrictions during the rebuttal phase, we cannot share it publicly now but are fully committed to open-sourcing upon acceptance. In LLM-based systems like ours, prompt design is central to model behavior. To support reproducibility, we provide detailed prompt templates and example outputs in Appendix, and step-by-step pipeline descriptions (main text). If any part of the prompt design is unclear, we’re happy to clarify during the rebuttal.

Regarding Novelty and Contribution

We thank the reviewer for their thoughtful comments. Our contribution may have been misunderstood as a heuristic combination of tools, but in fact, the LLM plays a central and decisive role in our framework, reasoning over structured molecular knowledge and coordinating all key design decisions. Rather than a simple toolchain, our method is definitely a novel collaboration between LLMs and 3D SBDD models, where the LLM drives the refinement process with domain-informed reasoning.

Integrating LLMs into SBDD is Non-Trivial and Scientifically Novel

LLMs are not inherently designed to understand molecular graphs, 3D structures, or protein-ligand interactions. Their reasoning is based on symbolic text, not geometry or chemistry. Therefore, applying LLMs to tasks that require spatial reasoning and chemistry knowledge—as is essential in SBDD—is a major challenge. Our contribution lies in bridging this gap through a carefully designed pipeline.

Ablation Study: Effectiveness Comes from LLM Reasoning Design, not external rule-based libraries.

Following the reviewer’s suggestion, we conducted an ablation study to further support the non-triviality and novelty of our LLM-SBDD integration. Specifically, we replaced our original chain-of-thought, domain-informed prompts with a simple instruction: “Based on the pocket and interaction analysis, modify the original molecule.”

Notably, despite using the same LLM and tools, the simplified prompting strategy performs significantly worse than the CoT-guided version and is only marginally better than the MolCRAFT baseline. This underscores that the effectiveness of our framework lies not in tool usage, but in domain-informed, structured reasoning via prompt design.

We are prepared to conduct additional ablation studies should reviewers request further clarification or analysis.

The Core Novelty: A Domain-Guided, LLM-Centered Framework

Our key innovation is a framework where the LLM serves not just as a molecule generator, but as a decision-making coordinator that mimics expert reasoning in structure-based drug design (SBDD). We encode domain knowledge into prompts, allowing the LLM to interpret chemical and structural constraints in a human-like, structured way. Specialized tools (e.g., PLIP, RDKit) handle atomic-level tasks like interaction extraction or fragment generation, while the LLM focuses on high-level reasoning and decision-making. This division of labor plays to each component’s strength and offsets the LLM’s limitations with raw 3D geometry.

This mirrors how medicinal chemists operate—translating 3D interactions into symbolic reasoning before modifying structures. Our framework replicates this workflow, refining SBDD outputs into drug-like molecules without relying on oracle-based optimization.

Regarding oracle budget

We thank the reviewer for raising this concern and offer the following clarifications:

CIDD is not an oracle-based optimization framework.

Our goal is to preserve the 3D interaction patterns from the initial SBDD output while improving drug-likeness (e.g., stability, synthesizability, plausibility). We do not query any drug-likeness-related oracles (e.g., QikProp, MRR, SA, QED) during generation; these metrics are used only post hoc for evaluation. This clearly distinguishes our method from optimization pipelines that rely on repeated oracle feedback.

PLIP enables interaction representation—not optimization.

Since LLMs cannot directly process 3D structures, we use PLIP to extract symbolic interaction patterns (e.g., hydrogen bonds, hydrophobic contacts) from the SBDD-generated molecule. These are passed to the LLM in natural language to support reasoning. PLIP is used solely as a structural-to-symbolic translator—not a scoring or feedback oracle.

Relations with existing SBDD model

As we are not oracle based optimization method, we believe it is still a fair comparasion with existing SBDD model. And actually we are not "competitive" with existing SBDD models. Instead, we are trying to build upon, or to say help them to solve unresolved problems such as structure reasonability and syntehzi ability.

Regarding computational costs

We acknowledge the reviewer’s concern regarding the runtime and token-related costs associated with repeated LLM invocations (e.g., scaffold analysis, molecule design, reflection, and selection). Indeed, LLM-agent pipelines—like ours, and as seen in many other LLM-agent use cases—can incur non-trivial computational overhead. However, we would like to offer the following justifications:

Cost Is Becoming a Lesser Barrier

The cost of multi-step LLM reasoning is rapidly shrinking thanks to advances in quantization, batching, and optimized inference. DeepSeek-V3/R1 achieves ~$0.55M/input and ~$2.19M/output tokens—a 60–70% drop from GPT-4’s 2023 cost. Frameworks like vLLM[1], vTensor[2], and Splitwiser[3] further boost efficiency: vLLM offers up to 24× throughput, vTensor gives 2× speedups with 70% less GPU memory, and Splitwiser reduces latency by 1.7× via prompt-generation colocation. With accelerators like TensorRT-LLM, domain-aware multi-step pipelines are becoming increasingly scalable. Overall, computational overhead is shrinking and is unlikely to be a major limitation in the future.

[1] Kwon et al., SOSP 2023 — PagedAttention for efficient memory management in LLM serving.

[2] Xu et al., CoRR 2024 — vTensor: virtual tensor management for efficient LLM inference.

[3] Aali et al., arXiv 2025 — Splitwiser: latency-optimized LLM inference under resource constraints.

Traditional Drug Development Is Far More Costly

Drug development is slow and costly, often taking years and billions of dollars to bring a single drug to market. The true bottleneck lies not in design time, but in the lengthy, expensive trial-and-error process of testing candidate molecules. Unlike real-time tasks like recommendation or autonomous driving, drug discovery prioritizes precision over speed. What matters is generating molecules with higher chances of experimental success—a goal our framework supports by improving the quality and viability of SBDD-generated candidates.

Our framework does not replace traditional SBDD models but accelerates their downstream refinement through a structured, multi-agent pipeline that preserves key interactions while improving structural plausibility. Although multi-step reasoning introduces some computational overhead, the significant reduction in human effort more than compensates—especially in large-scale or high-throughput scenarios.

We thank the reviewer for raising this point. We will discuss this tradeoff in the Limitations section and note that ongoing advances in LLM efficiency continue to reduce these barriers.

Regarding new metric, MRR

Defining strict criteria for drug-likeness remains elusive, yet long-standing empirical evidence has delineated recurrent structural motifs. In cyclic scaffolds, extranuclear electron distributions can enforce stability when specific electronic configurations, notably continuous sp² hybridization, are satisfied[1, 2]. Carbonyl and imine substituents often conform to sp² hybridization yet fail to establish a fully delocalized π-system around the ring—a feature frequently observed in approved drugs[3, 4, 5]. The MRR formalizes these principles, accurately distinguishing drug-like from non-drug-like molecules: 85.90% of FDA-approved small molecules comply with MRR criteria, substantially exceeding that of current 3D-SBDD models. By focusing on ring systems, present in 83.31% of approved drugs, MRR captures a fundamental chemical constraint: stable rings adopt either fully saturated or fully aromatic configurations, thereby conferring thermodynamic stability, synthetic tractability, and favorable interaction profiles—properties systematically exploited in both natural products and rationally designed therapeutics.

[1] Morsch et al., Organic Chemistry, LibreTexts, Ch. 15.2.

[2] Aromatic Compounds, ChemistryTalk, accessed 30 Jul 2025.

[3] Saudagar PS et al., Biotechnol Adv, 26(4):335–351, 2008.

[4] Poupaert JH et al., J Med Chem, 27(1):76–78, 1984.

[5] Stezowski JJ, J Am Chem Soc, 98(19):6012–6018, 1976.

Regarding hallucination or reasoning errors.

We thank the reviewer for raising this important concern. We mitigate hallucination and reasoning errors via:

Controlled decoding (low temperature) to reduce randomness and improve consistency.

Instructional prompting with clear reasoning steps and domain constraints, leveraging strong instruction-following models (GPT-4o, DeepSeek).

Self-correction, where invalid molecules trigger reflection and regeneration, enhancing robustness without external feedback.

We will include this discussion and future directions in the final version.

Regarding Limitation

We thank the reviewer for raising these important concerns. We will include discussions on the societal impact, as well as the limitations mentioned above, in the final version.

We sincerely thank the reviewer for their thoughtful evaluation and positive feedback on our paper. We appreciate the insightful comments and are happy to address the points raised below.

Regarding costs

We acknowledge the reviewer’s concern regarding the runtime and token-related costs associated with repeated LLM invocations (e.g., scaffold analysis, molecule design, reflection, and selection). Indeed, LLM-agent pipelines—like ours, and as seen in many other LLM-agent use cases—can incur non-trivial computational overhead. However, we would like to offer the following clarifications:

Cost Is Becoming a Lesser Barrier

The cost of multi-step LLM reasoning is rapidly declining thanks to advances like quantization, batching, and optimized inference engines. For example, DeepSeek-V3/R1 supports inference at $0.55M/input and$2.19M/output tokens—a 60–70% reduction from GPT-4’s 2023 pricing. Efficiency-focused frameworks such as vLLM[1], vTensor[2], and Splitwiser[3] further accelerate inference: vLLM achieves up to 24× higher throughput, vTensor offers 2× speedups with 70% less GPU memory, and Splitwiser reduces latency by 1.7× via co-located prompt and generation. Combined with hardware accelerators like TensorRT-LLM, these developments make domain-informed, multi-step LLM workflows increasingly practical and scalable. For example, with inference frameworks like vLLM, the DeepSeekV3 model can be deployed on 8×A100 80GB GPUs.

In addition, attention-level optimizations have significantly accelerated LLM inference: FlashAttention‑2 achieves up to 4–5× speedup over standard attention [4], DuoAttention reduces memory usage by 2.55× and speeds up decoding by 2.18× [5], and MInference delivers up to 10× acceleration in pre-filling long contexts [6].

These examples confirm that the computational and token overhead associated with multi-step LLM pipelines is rapidly shrinking and is unlikely to pose a significant obstacle in the foreseeable future.

[1] Kwon, Woosuk, et al. "Efficient memory management for large language model serving with pagedattention." Proceedings of the 29th symposium on operating systems principles. 2023. [2] Xu, Jiale, et al. "vTensor: Flexible Virtual Tensor Management for Efficient LLM Serving." CoRR (2024). [3] Aali, A., Cardoza, A., & Capo, M. (2025). Splitwiser: Efficient LM inference with constrained resources. arXiv preprint arXiv:2505.03763. [4] Dao, T. FlashAttention‑2: Faster Attention with Better Parallelism and Work Partitioning. arXiv preprint arXiv:2307.08691. [5] Xiao, G. et al. DuoAttention: Efficient Long‑Context LLM Inference with Retrieval and Streaming Heads. In International Conference on Learning Representations (ICLR), 2025. arXiv preprint arXiv:2410.10819. [6] Jiang, H. et al. MInference 1.0: Accelerating Pre‑filling for Long‑Context LLMs via Dynamic Sparse Attention. In Conference on Neural Information Processing Systems (NeurIPS), 2024. arXiv preprint arXiv:2407.02490. 

Traditional Drug Development Is Far More Costly

Drug development is slow and costly, often taking years and billions of dollars to bring a single drug to market. The true bottleneck lies not in design time, but in the lengthy, expensive trial-and-error process of testing candidate molecules. Unlike real-time tasks like recommendation or autonomous driving, drug discovery prioritizes precision over speed. What matters is generating molecules with higher chances of experimental success—a goal our framework supports by improving the quality and viability of SBDD-generated candidates.

Our framework does not replace traditional SBDD models but accelerates their downstream refinement through a structured, multi-agent pipeline that preserves key interactions while improving structural plausibility. Although multi-step reasoning introduces some computational overhead, the overall time savings from eliminating manual, expert-driven intervention far outweighs the additional inference cost—especially in large-scale or high-throughput scenarios.

Case Study: Runtime and Token Usage

To provide a concrete estimate, we profiled a representative design iteration with the following statistics:

Average runtime per design : ~1 minutes; Total token usage per design: ~10000 tokens

In this paper for each initial molecules we generate 5 designs sequentially so the cost would be 5 times. This can be adjusted based on requirements. In practice we found that one design is able to create molecules with better drug likeness and syteh

Importantly, these operations are highly parallelizable, and large-scale molecules screening can be executed efficiently in batch across multiple machines, and still support generating large scale number of molecules in reasonable time.

In summary, we thank the reviewer for raising this important point. We acknowledge that cost is a common concern for frameworks involving multiple LLM agents, and we will include a discussion of this limitation in the revised version of the paper. That said, we believe recent advances in model efficiency and inference infrastructure are rapidly reducing this barrier, as discussed above.

Regarding more challenging targets

We thank the reviewer for the helpful suggestion to evaluate CIDD on more challenging targets with diverse chemotype distributions. In response, we tested CIDD on additional targets: 1sa1, 1t8i and 7rpz, all known for their structural complexity and variability in ligand scaffolds. For 1sa1, Tubulin presents a challenging drug target because its binding sites are highly dynamic, involving extensive protein–protein interfaces; furthermore, the ligand in this complex, podophyllotoxin, is a complex natural product scaffold[1]. For 1t8i, Human DNA Topoisomerase I presents a challenging drug target because its catalytic mechanism involves a highly dynamic binding site that accommodates DNA; moreover, the ligand in this initial complex is camptothecin, a natural product scaffold rather than a traditionally rationally designed small molecule[2]. For 7rpz, KRASG12D is difficult to inhibit due to its high GDP/GTP affinity and absence of a nucleophilic residue near the switch II pocket, which hampers covalent strategies and limits prior chemotypes. The ligand uniquely overcomes these barriers with a pyrido[4,3-d]pyrimidine scaffold, fully engages the pocket, and achieves picomolar noncovalent inhibition with strong in vivo efficacy[3].

Results on 1sa1:

Results on 1t8i:

Results on 7rpz:

As shown, CIDD achieves substantially higher success ratios than the MolCRAFT baseline on these targets, even when the initial molecules generated by the basele model are of low quality. This demonstrates CIDD’s ability to be effective for chemically complex and challenging protein targets.

We thank the reviewer again for this insightful suggestion, and we will include these new results and discussions in the revised paper.

[1] Ravelli, R.B.G. et al. Nature 2004, 428, 198–202.

[2] Staker, B.L. et al. J. Med. Chem. 2005, 48, 2336–2345.

[3] Wang, X., et al. J. Med. Chem. 65.4 (2021): 3123–3133.

Regarding prompt sensitivity

We thank the reviewer for highlighting the importance of prompt sensitivity.

First, as shown in Table 2(a), even when using different large language models (GPT-4o, GPT-4o-mini, DeepSeek-V3), the outcome metrics remain fairly consistent. This suggests that our framework is robust and not overly sensitive to the choice of underlying LLM. In other words, model updates or substitutions are unlikely to cause failures or unpredictable behaviors in the system.

Table 2(a):Different LLM Backends in CIDD

To further assess prompt sensitivity, we conducted an additional experiment in which we asked ChatGPT to automatically rephrase the core design prompts, while preserving their original intent. We use deepseek-v3 as our LLMs.

As shown in the following table, the performance remained largely stable across the two prompt versions, indicating that the method is not highly sensitive to moderate rephrasings in Chain-of-Thought design instructions.

These results suggest that while prompt formulation can influence LLM responses, CIDD remains robust under reasonable variations in prompt phrasing.

Regarding synthetic feasibility of our top candidates

As shown in Table 1, the SA score reflects the synthetic feasibility of the generated molecules, and our method improves this metric compared to baselines.

We sincerely thank the reviewer for their thoughtful evaluation and positive feedback on our paper. We appreciate the insightful comments and are happy to address the points raised below.

Regarding the diversity

Thank you for pointing out the diversity metric. Initially, we included diversity in the main results table, but later decided to remove it as it did not offer significant insights—most methods showed only marginal differences in this aspect. For completeness, we provide the diversity scores of the baseline models and CIDD below:

We also evaluated the diversity when generating 100 molecules per target. As shown, CIDD achieves diversity comparable to its baseline model, MolCraft. Moreover, increasing the number of generated molecules from 10 to 100 does not significantly affect the average diversity, and diversity is not a concern of CIDD framework.

Regarding re-docking SBIG-generated molecules

We appreciate the reviewer’s insightful comment on the use of re-docking in our pipeline. The reason we perform re-docking, rather than directly using the generated molecular conformations, is that the original conformations produced by generative models may contain issues—such as atomic clashes or physically unstable geometries—that can negatively impact both interaction analysis and subsequent steps. This issue has been discussed in recent paper[1]. Re-docking helps ensure that we begin from a more stable and physically reasonable pose.

That said, we fully agree with the reviewer that it is important to test whether this re-docking step is necessary. To this end, we conducted an ablation study where we directly used the originally generated conformations without re-docking. We use deepseek-v3 as our LLMs. The results are shown below:

As shown, the performance differences are relatively small, suggesting that our framework is robust to this design choice. We thank the reviewer again for the valuable suggestion and will include this analysis and discussion in the appendix of the revised manuscript.

Regarding 2D generation methods guided by docking scores

Thank you for pointing us to these works. We appreciate the references and agree that those 2D molecular generation is an important and active line of research.

We would like to clarify that our method is fundamentally different in formulation from the works cited. Specifically, CIDD is not an optimization framework: we do not use reinforcement learning, docking-score-based fine-tuning, or iterative optimization guided by oracle feedback. Instead, our framework focuses on repairing and improving imperfect 3D SBDD-generated molecules by preserving key protein-ligand interactions while enhancing drug-likeness and synthesizability—without relying on explicit reward signals or gradient guidance.

Moreover, the referenced methods are not evaluated on the CrossDocked dataset, which we adopt as our main benchmark to ensure consistency with our chosen baselines. Likewise, these methods are not included as baselines in the prior works we compare against. As a result, a fair and meaningful comparison would require non-trivial adaptation and retraining of those methods in our setting—especially considering that some of them are specifically designed for narrow domains such as antibiotic discovery—which is unfortunately infeasible within the limited rebuttal period. In addition, several of the referenced methods can be computationally intensive, which is why they are typically evaluated only on a small number of selected targets.

Importantly, we view CIDD as complementary to these optimization-based methods. In fact, many of them could be incorporated into our pipeline as components within the SBIG (Structure-Based Initial Generation) stage. Our framework is modular by design, and can serve as a downstream interaction-preserving refinement module, enhancing the outputs of other generative methods that may not explicitly account for protein context or structural plausibility.

We thank the reviewer again for bringing these methods to our attention. We will cite the relevant works and include a discussion of their methodological differences, as well as their relationship to CIDD, in the Related Work section of the revised manuscript.

Regarding similarity metrics

Thank you for pointing out the metric on similarity. Yes, we do evaluate the similarity between the raw molecule from the SBIG module and the final molecule generated by CIDD. This choice is intentional and aligned with the design goal of the LEDD step: to make minimal but meaningful modifications that improve chemical quality while preserving the structural basis provided by the SBDD model.

Specifically, we treat the SBIG-generated molecule as a structure-based proposal that carries important protein-ligand interaction information. Rather than generating a completely new molecule, which may disrupt these interactions, we aim to refine the initial structure in a way that retains its interaction intent. Therefore, a higher similarity reflects a successful case where the molecule is improved in terms of drug-likeness or synthesizability, while maintaining the original design.

This design choice also supports a data-centric view. By producing pairs of molecules with small structural differences but large differences in quality, our framework naturally generates valuable data for future training of optimization models. These minimally perturbed, property-improved pairs can serve as useful examples for learning structure-to-property relationships.

We will clarify the motivation and implications of this similarity metric in the revised version of the manuscript.

Regarding hit-to-lead tasks

We thank the reviewer for highlighting the potential of applying the CIDD framework to hit-to-lead optimization. Although CIDD was not originally designed for this purpose, it can indeed be adapted to such scenarios. To explore this, we initialized the generation process—and the input to the LEDD module—using ground-truth ligands from the CrossDocked test set. As shown in the following table, despite these ligands already exhibiting good synthetic accessibility and a high reasonable ratio, CIDD was still able to generate molecules with further improvements in potency, synthesizability, and structural reasonability, demonstrating its effectiveness and potential in hit-to-lead optimization.

Notably, as we mentioned, the original CIDD (LEDD) framework was not specifically designed for hit-to-lead optimization tasks. Fully supporting such applications would require adjustments to the pipeline design, modifications to the chain-of-thought prompts, and changes in tool integration. Nonetheless, the current results already demonstrate the potential of our proposed domain model and LLM collaboration paradigm in broader scenarios such as hit-to-lead. We will incorporate this direction into the discussion of future work.

We sincerely thank the reviewer again for the valuable feedback.

Regarding generating 1,000 or 10,000 molecules

Thank you for the thoughtful question. We confirm that the CIDD pipeline supports large-scale candidate generation, including sets of 1,000 or even 10,000 molecules.

The end-to-end generation time for a single molecule—including interaction parsing, prompt construction, LLM inference, and generating for N=5 new designs—is approximately 5 minutes when run sequentially. However, the pipeline is inherently parallelizable. When deployed via cloud-based API services (such as OpenAI or DeepSeek), there is virtually no limit on parallelization, assuming adequate API throughput. In practice, this allows for generating 1,000 molecules within just a few minutes using parallel asynchronous requests.

Compared to the time required for downstream experiments such as molecular dynamics simulations, or wet-lab synthesis and validation, this generation time remains extremely efficient, and does not constitute a computational bottleneck in the overall drug discovery pipeline.

Regarding SA scores

Thank you for the question. Yes, we re-scaled the original synthetic accessibility (SA) scores for consistency and interpretability in our evaluation.

The original SA score ranges from 1 (easy to synthesize) to 10 (difficult to synthesize). To make the interpretation consistent with other normalized scores (where higher is better), we applied the following linear transformation:

sa_norm = round((10 - sa) / 9, 2)

This maps the SA score to a normalized range between 0 and 1, where a higher value indicates easier synthesis. This rescaled score was used for reporting purposes in our tables and analysis.

This reviewer would like to thank the authors for their detailed response.

The authors highlight that the method makes interaction-preserving repairs and improvements to molecules generated by the SBIG. While it is true that the proposed method does not rely on reward signals directly, the interactions are hypothesized via either re-docking or the original generated structure, both of which are dubious assumptions.

Despite this, it is compelling that the method is able to obtain cheap and substantial improvements to molecules without additional training, and this reviewer finds the paired-data idea for training future structure-to-property models an interesting application. For this reason, the score will be raised to 5.

We sincerely thank the reviewer for their thoughtful evaluation and positive feedback on our paper. We appreciate the insightful comments and are happy to address the points raised below.

Regarding CIDD not directly generating 3D conformations and taking longer computation time

We acknowledge the reviewer’s observation that our proposed CIDD framework may incur longer computational time compared to traditional 3D structure-based drug design (SBDD) methods. While this is indeed a relevant concern, we would like to offer additional context to clarify this trade-off.

Drug discovery is inherently a time-intensive and resource-demanding process, often spanning several years and incurring significant financial costs. A substantial portion of this effort is driven by human tasks—such as expert assessments, manual filtering, and iterative decision-making—rather than computational bottlenecks alone.

CIDD explores a different trade-off: although it introduces additional computational steps, it significantly reduces human involvement by leveraging large language models (LLMs) to enhance chemical reasonability and drug-likeness. These are aspects that traditionally rely on deep domain expertise. In our experiments, CIDD consistently improves the quality of generated molecules in these dimensions, delivering results that are more chemically reasonable with less need for manual inspection.

We believe that, in practice, this translates to a net gain in efficiency: the additional generation time is more than offset by the reduction in expert labor and manual validation efforts.

We appreciate the reviewer’s observation and will incorporate this discussion into the limitations section of our revised manuscript.

Regarding the code

We apologize for not providing the codebase during the initial submission. We are fully committed to releasing the code and supporting reproducibility. However, due to file-upload restrictions during the rebuttal phase, we are currently unable to share the code publicly. We will release the full codebase upon acceptance.

That said, in LLM-based systems like ours, prompt design plays a central role in governing the model’s reasoning behavior. In the CIDD framework, prompts dictate how the LLM interprets structural context, coordinates tool usage, and executes multi-step molecular design decisions. As such, we have placed particular emphasis on reproducibility through prompt transparency.

To support this, we provide detailed prompt templates and example model responses for each stage of the pipeline in the Appendix. Coupled with the step-by-step descriptions in the main text, we believe these materials offer sufficient detail for readers to reproduce and understand the behavior of our system, even prior to code release.

If there are any ambiguities or questions regarding the prompts we’ve provided, we are more than happy to clarify them during the rebuttal process.

Regarding the 3D-SBDD baselines

We apologize for any confusion caused by the terminology and formatting used in the table. We understand the reviewer’s concern that “3DSBDD + LLM” may be interpreted as an individual method rather than a category.

To clarify: in our work, “3DSBDD” refers to the collection of all baseline methods we compare against, rather than a specific standalone approach. In particular, we do employ MolCRAFT as our SBIG (structure-based Interation Generation) module to generate the initial raw molecules, with the corresponding results shown in Table 1. And as shown in the table, the Vina score of generated molecules actually becomes better compared to MolCRAFT(the second-to-last line in Tab. 1) after modifications from LLMs. （-8.496 vs -7.783）

To avoid ambiguity, we will revise the table caption and terminology to make this categorization clearer and more precise in the updated version.

Thank you again for pointing this out—your careful review and valuable feedback have been very helpful in improving the clarity and rigor of our work.