Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule

We sincerely thank the reviewer for the careful reading and insightful feedback. Below, we address each question and concern to improve the clarity and completeness of our work, as well as demonstrate its generality.

Questions

Q1: Integration with More Frameworks

Why not integrate the proposed VOS with more frameworks? It's necessary to demonstrate its practical applicability within various frameworks.

Thank you for this valuable suggestion. To demonstrate VOS's broader applicability, we integrated it with the diffusion-based framework TargetDiff. Specifically, we reimplement TargetDiff with our generalized training objective and train 140k steps following the default training configuration (TargetDiff*, with code provided), and then derive the test-time optimal schedule that resembles the shape in Figure 4B.

The results on CrossDock show that VOS successfully enhances conformation quality for diffusion models too, with generated poses achieving Vina Scores closer to Vina Min values, indicating near-optimal realistic poses.

Q2: CrossDock Test Set Consistency

Is the ID CrossDock test set identical to those used in previous studies?

Yes, we confirm that we used the identical CrossDock test set as proposed in previous studies, ensuring fair and consistent comparison with existing methods.

Weaknesses

W1: Clarity of Implementation Details

The description of the proposed MolPilot is not clear, lacking many implementation details.

We appreciate this feedback and will enhance the clarity of our description by adding more detailed description to Section 4, Algorithm 1, and Appendix A. We will move key implementation details from the Appendix to the main text where appropriate. Thanks again for the advice, and we will revise our manuscript accordingly.

W2: Essential References

This paper ignores many essential references, limiting its impact and connection to existing research. For example, many essential references are not discussed in this paper, such as IPDiff [1], DecompOpt [2], IRDiff [3], DecompDPO [4], and BINDDM[5].

Thank you for highlighting these important references. We agree that IPDiff, DecompOpt, IRDiff, DecompDPO, and BINDDM represent meaningful advancements in controllability and practicality by incorporating guidance signals (e.g. binding affinity, interactions) to enhance molecular properties of generated ligands.

We will expand our related work section to include thorough discussion of these approaches and their contributions. Additionally, we report the results of IPDiff and BindDM on CrossDock, where we calculated PB-Valid and other metrics based on the samples obtained by using the official code, showing that they indeed enhance affinities through binding signals from pretrained predictors, yet also with the issue of conformation plausibility.

W3: Code Repository

We sincerely apologize for the oversight regarding our code repository. We had uploaded our code to GitHub before the reviewing process began, but failed to verify the auto-update mechanism for the anonymous repository. This unintentional mistake has been rectified now, and we have force-updated the repository with all necessary code to ensure full reproducibility, including the newly added experiment with TargetDiff.

W4: Ethical Review Concerns

We appreciate the attention to formatting requirements. The missing content was due to that we have commented out printAffiliationsAndNotice{}, yet it was not intended to create extra space. As noted in the ICML 2025 Peer Review FAQ (March 18, 2025 update), the Program Chairs have acknowledged that papers with this specific issue will not be desk-rejected "due to confusion in the LaTeX template." We will restore this element in our revision to ensure full compliance with formatting requirements.

We thank the reviewer again for the constructive feedback that helps us improve our manuscript, and we welcome further discussion.

We sincerely thank the reviewer for the thorough evaluation and valuable feedback. Below, we address each point raised to improve the clarity, reproducibility, and depth of our work.

Questions

Q1: Architectural Changes

The backbone of MolPilot aligns with DecompDiff, where we only replaced the MLP in the q, k, v calculation with LinearNoBias to reduce GPU memory consumption. This decreases memory usage from 42177MB to 34699MB for batch size 8 in the first 500 steps, enabling more efficient training without compromising performance. We report this baseline trained by default loss in Figure 6.

Q2: Sensitivity to Optimal Path Choice

Thank you for this insightful question. We conducted additional experiments interpolating between linear $t_l$ (coef $c=0$) and our optimal time-rescaling functions $t_o$ ($c=1$), by setting time functions $t=c\cdot t_l + (1-c)t_o$. Our findings show a clear trend of improving performance as we move toward the optimal schedule on both datasets.

For CrossDock:

For PoseBusters:

Additional Clarifications

Q3: Repository and Reproducibility

We sincerely apologize for the oversight regarding the code repository. We had uploaded our code to GitHub before the reviewing process, but failed to verify the auto-update mechanism for the anonymous repository. This unintentional mistake has been rectified, and we have now force-updated the repository with necessary code to ensure reproducibility.

Q4: Notation Clarifications

We appreciate your careful reading that identified these notation inconsistencies. $\alpha$ in Equation 2 is actually the discretized $\alpha_i$ for time step $i$ in the sequence, derived from time-dependent function $\alpha(t)$, not a constant. $K$ is defined as the number of classes for one-hot encoding (with $K_h$ and $K_A$ denoting atom and bond types, respectively). $\mathbf{e}_\mathbf{x}$ is kronecker function, i.e. the projection from a class index $\mathbf{x}=j$ to a one-hot vector $\in \mathbb{R}^K$ with the $j$-th value equal to 1. We will revise the manuscript to improve clarity as requested.

Q5: Molecular Docking Study and Baselines

We thank the reviewer for this important point about training data differences. Following this suggestion, we trained a version of our model on PDBbind from scratch (120 epochs over 1.5 days) to enable a more direct comparison.

Although the comparison remains imperfect, the results are still informative, for our primary goal was to demonstrate that ours is the first with genuine docking capability that can repurpose an SBDD model for the docking task. This capability stems from our decoupled training strategy rather than architectural changes. We will add clarification to the manuscript.

Q6: Novelty of Generated Compounds

Thank you for this valuable suggestion. We have calculated comprehensive metrics to assess the novelty. As shown in the table below, MolPilot achieves the highest combined metric (0.864) among all methods, demonstrating its ability to generate novel, valid, and unique compounds.

In addition to these metrics, we have calculated the Fréchet ChemNet Distance (FCD), which measures distribution similarity between generated molecules and training distribution (similar to FID for images, lower is better). MolPilot achieves competitive FCD scores that demonstrate its ability to generate novel compounds while faithfully capturing the training distribution.

We sincerely thank the reviewer for the thorough reading of our manuscript and the positive evaluation. We are pleased that the reviewer recognized the key challenges and contributions of our work on deriving optimal schedules for structure-based drug design, and we certainly welcome further discussion.

Regarding Theoretical Contributions

We appreciate the reviewer's acknowledgment of one of our main contributions - the dynamic programming algorithm for VLB-optimal noise scheduling when dealing with the twisted discrete-continuous probability path. We believe the theoretical foundation is sound and provides an important advancement for handling the challenging joint generation of molecular graphs (discrete) and their 3D structures (continuous).

Regarding Experimental Results

We are particularly encouraged by the reviewer's recognition of our experimental design and results, especially highlighting our improvement in PB-Valid. As the reviewer correctly noted, this metric demonstrates our method's ability to generate consistent 2D and 3D molecular structures - addressing a key limitation in previous approaches that often favored 3D structure at the expense of 2D topology.

We thank the reviewer again for the supportive evaluation. We remain committed to further developing and refining our method to contribute to the field of structure-based drug design.