Scalable and Cost-Efficient de Novo Template-Based Molecular Generation

Thank you for your thoughtful feedback. We have carefully addressed all the concerns raised and have incorporated your suggestion regarding a more fine-grained study of component contributions. The corresponding ablation experiments are underway, and we will include their results in the revised camera-ready version of the paper. We hope these additions may help to further clarify our approach and underscore our commitment to a thorough evaluation. We would greatly appreciate your kind consideration of these improvements when reassessing the paper.

Yields and stock prices for MEDIUM and LARGE

For the MEDIUM and LARGE settings, stock prices were obtained by automatically scraping Enamine’s publicly available catalog. These prices represent a static snapshot at the time of data collection and do not reflect temporal fluctuations or pricing differences across vendors. Yield estimates are derived from literature-based averages, which provide reasonable robustness across diverse reaction classes. However, as discussed in the Cost Estimates Generalization section, these estimates may lack precision in the context of specific experimental setups or lab conditions.

Validation via retrosynthesis software

SCENT is a de novo generative model that rarely reproduces molecules previously tested in the lab—typically, only ~1 in 150,000 generated compounds appears in ChEMBL. Thus, validating SCENT using retrosynthesis tools better reflects its intended use case than somehow testing it on a set of molecules with known synthesis routes. As shown in Figure 7, SCENT achieves a higher AiZynthFinder success rate than RGFN (likely due to its preference for shorter synthetic trajectories) and outperforms the non-template-based FGFN+SA, even though FGFN+SA is explicitly optimized for synthetic accessibility.

Fine-grained components’ contributions

We appreciate the suggestion and agree that a more fine-grained ablation would provide valuable insights. In response, we initiated experiments with SCENT (w/o C + P) and SCENT (w/o C + D) to better isolate the individual contributions of the Dynamic Library and the Exploitation Penalty across all tasks and settings.

Unfortunately, these experiments are computationally intensive and have been delayed due to limited server availability, particularly during the rebuttal period, when compute resources are in high demand. While we are aiming to complete these runs before the end of the discussion phase on August 6th, we acknowledge that there is a chance they may not finish in time.

Regardless, we are committed to including the full results in the camera-ready version of the paper and will notify the reviewers as soon as the results are available.

Cost Guidance hyperparameters

In the original paper, we have not explored the sensitivity, neither the automated tuning (we performed semi-manual small grid searches). Following the suggestion, we performed a detailed sensitivity analysis for each component (including temperature α for costs models, and γ in Exploitation Penalty). The results can be found in the response to Reviewer QKfB (Section Hyperparameters). Our results indicate that SCENT is generally resilient to variations in hyperparameters - its overall performance remains stable under a wide range of settings. That said, some hyperparameters have a notable impact on specific aspects of behavior; for example, changing $\alpha_D$ temperature for Decomposability Guidance can lead to a ~20% increase in the number of discovered modes. This highlights the potential for further improvements through systematic hyperparameter optimization using Bayesian optimization, which we leave as future work.

Trajectory analysis

As suggested by the reviewer, we conducted a detailed analysis of the generated synthesis trajectories and found that the Synthesis Cost Guidance module influences generation in two key ways: (1) it defers the use of expensive fragments to later stages in the trajectory, effectively reducing the quantity of such fragments needed to produce the same amount of final product; and (2) it favors cheaper fragment alternatives when available. This behavior aligns with the module’s objective of minimizing overall synthesis cost. A thorough analysis, along with concrete examples, is included in our response to Reviewer 9Lb3 (see Pathways Analysis section). We are unable to include figures or PDFs during the rebuttal, but we will incorporate the corresponding visualizations in the revised version of the paper.

Cost estimates generalization

We agree that the yield estimation and cost model may not generalize well across different labs. It is, however, a consequence of the inherent chemical synthesis variability and the lab-specific factors (e.g. contracted vendor) being inherently hard to account for. To our knowledge, there is no universally applicable yield estimation method [1], neither a gold-standard scheme of pricing the building blocks.

Our framework is as general as it can be, in the sense that it can accommodate any cost estimation method, including those that are accurate yet very lab-specific.

Moreover, in a study requested by Reviewer 9Lb3 (see Warm-start study section), we demonstrate that SCENT’s cost prediction module can effectively adapt to random perturbations in pricing during training. When price changes occur, whether due to market fluctuations or vendor contract updates, they can be directly incorporated into the model. SCENT is able to adapt to these changes within approximately 1,000 training iterations, eliminating the need for retraining from scratch each time prices are updated. This warm-start capability significantly improves the model’s practicality in dynamic pricing environments.

[1] “When Yield Prediction Does Not Yield Prediction: An Overview of the Current Challenges”

Costs and yield deferred to appendices

Thank you for the comment, we considered this and would like to argue that the main focus of the paper is the generative framework, and the cost prediction model, while important for the framework to work, is in the end secondary. As such, we’d argue that keeping the details of this model in the appendix is appropriate. However, if the Reviewer feels strongly that shouldn’t be the case, we can incorporate additional details in the main body of the paper.

Limitation section

We apologize for the lack of limitation section, we are going to include the following limitation summary in the revised paper:

“While our study demonstrates the effectiveness of SCENT in generating cost-efficient and synthesizable molecules, several limitations remain.

First, the synthesis cost estimates are simplified and heuristic, omitting factors such as solvent usage, purification steps, hazardous reagents, reaction conditions, and scalability. Consequently, the current cost model should be regarded as a first-order approximation. More accurate estimation would require lab-specific calibration, including robust yield prediction models and vendor-specific building block pricing.

SCENT’s performance varies across different biological targets, indicating that its effectiveness may depend on task-specific factors. Identifying and addressing these bottlenecks is an important direction for future work. Key areas of improvement include refining the exploitation behavior of the Dynamic Library, and enhancing the generalizability of the cost predictor model used in Synthesis Cost Guidance. Additionally, we do not assess performance as the number of reaction templates increases, an important scalability consideration, as this expansion significantly enlarges the action space and introduces additional learning challenges.

Our framework emphasizes synthesizability and cost, but does not incorporate toxicity or broader ADMET properties, which are critical for downstream drug development. Finally, we acknowledge the absence of wet-lab validation, which remains the definitive benchmark for evaluating the practical utility of our approach. We leave this to future work.”

Thank you very much for your thoughtful and constructive review. We’ve carefully addressed all the issues you raised and have implemented your suggestions related to computational resource metrics and cost model evaluation. In particular, the evaluation of the Synthesis Cost Guidance is now complete, and we are actively working on the analysis of the Decomposability Model.

We believe these enhancements have significantly strengthened the paper and trust that the revisions will provide additional clarity and depth to our contributions.

Experiments with random cropping

As shown in Figure 6, reducing the fragment library size in SCENT (w/o C) generally leads to decreased or stagnant performance. Since the smaller libraries are strict and random subsets of the larger ones, this reduction effectively “crops” the available fragments. By contrast, adding cost biasing (C) and exploitative penalty (P) yields significant performance gains.

From these results, we conclude that expanding the fragment library is more beneficial than cropping it, particularly when combined with cost biasing and exploitative penalty. The largest improvements are observed for SCENT (C+P), which underscores the importance of maintaining a rich and diverse fragment library.

Narrowed number of modes in 128k

Thank you for the insightful question. As the number of building blocks increases, the state space grows exponentially, which makes learning more challenging under a fixed training iteration budget. This may explain why the number of modes narrows at M=128k, despite having more fragments available. This trend is consistent with prior observations in related work (e.g., RGFN). Notably, SCENT is able to maintain strong performance even at higher fragment counts, significantly extending the range over which improvements are observed compared to RGFN.

Building block libraries analysis

Thank you for your helpful question. To characterize the SMALL, MEDIUM, and LARGE building blocks libraries beyond their size, we analyzed molecular diversity, complexity, and key physicochemical properties relevant to drug discovery. Specifically, we computed:

• Diversity: Number of unique molecular modes and scaffolds, using leader clustering (with a Tanimoto similarity threshold of 0.3) to define clusters.
• Complexity: Metrics including molecular weight and rotatable bonds.
• Physicochemical features: Properties such as logP, hydrogen bond donors (HBD), hydrogen bond acceptors (HBA), and topological polar surface area (TPSA), which influence solubility, permeability, and binding potential.

The summary statistics are as follows:

These data show that the curated SMALL dataset represents a focused set of chemically simpler molecules with fewer scaffolds and lower molecular weight, while the MEDIUM and LARGE datasets encompass increasingly diverse chemical scaffolds and more complex molecules. The LARGE set in particular covers a broad chemical space, as indicated by the large number of unique scaffolds and modes and the wider range of physicochemical properties.

This comprehensive characterization supports our claim of “broad coverage of molecular space” by demonstrating substantial diversity in core scaffolds and molecular features critical to drug discovery across the datasets.

Computational resources

As requested by the reviewers, we report GPU runtimes (in minutes) for all baselines on the sEH proxy, averaged over 3 random seeds using a V100 32GB GPU. All methods used a batch size of 64, except SynFlowNet, which required batch size 32 on the LARGE setting even after optimizing its policy code to fit within memory limits.

RGFN shares a part of the codebase with SCENT, allowing for direct comparison. We observe consistent improvements across all settings, particularly in the SMALL configuration. While SynFlowNet is faster on SMALL and MEDIUM, it struggles on LARGE, even with our memory optimizations. However, differences in codebases make direct comparisons less definitive.

Note that our implementation is not optimized for runtime and could benefit from further improvements.

Explain sEH

Following the suggestion of the reviewer we added a brief explanation of sEH proxy to the revised paper: “The sEH proxy is a graph neural network trained to predict the binding affinity of small molecules to soluble epoxide hydrolase (sEH), using docking scores computed by AutoDock Vina as training labels. The higher the reward defined by sEH proxy, the more strongly a given molecule binds to sEH.”

Performance of cost models

Thank you for the valid suggestion. In response to Reviewer QKfB, we conducted an analysis of the cost model used in Synthesis Cost Guidance (see Performance of Cost Models). A similar analysis for the Decomposability Cost Guidance is currently in progress, and we will provide the results as soon as they become available.

MCES instead of Tanimoto

Thank you for drawing our attention to MCES. We agree that it can offer a more nuanced measure of molecular similarity compared to Tanimoto similarity. However, we opted for Tanimoto primarily because (1) it is widely used in related work, (2) it is significantly faster to compute.

Limitation section

We apologize for the lack of limitation section, we are going to include the following limitation summary in the revised paper:

“While our study demonstrates the effectiveness of SCENT in generating cost-efficient and synthesizable molecules, several limitations remain.

First, the synthesis cost estimates are simplified and heuristic, omitting factors such as solvent usage, purification steps, hazardous reagents, reaction conditions, and scalability. Consequently, the current cost model should be regarded as a first-order approximation. More accurate estimation would require lab-specific calibration, including robust yield prediction models and vendor-specific building block pricing.

SCENT’s performance varies across different biological targets, indicating that its effectiveness may depend on task-specific factors. Identifying and addressing these bottlenecks is an important direction for future work. Key areas of improvement include refining the exploitation behavior of the Dynamic Library, and enhancing the generalizability of the cost predictor model used in Synthesis Cost Guidance. Additionally, we do not assess performance as the number of reaction templates increases, an important scalability consideration, as this expansion significantly enlarges the action space and introduces additional learning challenges.

Our framework emphasizes synthesizability and cost, but does not incorporate toxicity or broader ADMET properties, which are critical for downstream drug development. Finally, we acknowledge the absence of wet-lab validation, which remains the definitive benchmark for evaluating the practical utility of our approach. We leave this to future work.”

We thank the reviewer for their feedback. We have discussed all the issues raised and incorporated the suggestion to include case studies on generated synthesis trajectories which strengthen our paper. Given these revisions, we kindly hope the reviewer will consider updating their score.

Ground-truth synthesis cost of a molecule

As SCENT is designed for de novo molecular discovery, most of its outputs are novel and have not been synthesized before. Consequently, empirical synthesis costs are not available. The only definitive way to validate these costs would be through wet-lab experiments, which we agree would be valuable, but fall outside the scope of this work.

Wet-lab experiments

We appreciate the reviewer’s concern regarding the absence of wet-lab validation, which we fully acknowledge would significantly strengthen the benchmark. However, we believe this falls outside the typical scope of a machine learning conference. Our goal is to introduce and evaluate SCENT as a computational framework; while experimental validation is important, we argue it is not essential at this stage for demonstrating the method’s value in algorithmic and modeling terms. We see wet-lab validation as an important next step and a promising direction for future collaboration with experimental groups.

Case studies

We appreciate the reviewer’s suggestion and have conducted a detailed case study analyzing synthesis trajectories proposed by our model. While we are unable to include any images at this rebuttal stage, specific examples, including the generated molecules, the reactions used to construct them, and the associated synthesis costs, are provided in the next Pathway Examples section. The detailed analysis of the pathways can be found in our response to Reviewer 9Lb3, under the section Pathways Analysis. It demonstrates that the Recursive Cost Guidance module effectively reduces synthesis cost by deferring the use of expensive fragments until later stages in the trajectory and favoring cheaper fragment’s alternatives when available. We will include the entire analysis, along with visualizations in the revised paper.

Pathways Examples

The extensive analysis of those synthesis pathways can be found in response to Reviewer 9Lb3, under the section Pathways Analysis.

Example 1, SMALL settings.

SCENT (w/o C)

Step 1:

OB(c(cc1)cc2c1[nH]cc2)O.Nc3c(O)cc(Br)cc3>>Nc4c(O)cc(c(cc5)cc6c5[nH]cc6)cc4

Reactant_1 cost: 12.56

Reactant_2 cost: 0.43

Reaction Yield: 75%

Product cost: 17.32

Step 2:

Nc1c(O)cc(c(cc2)cc3c2[nH]cc3)cc1.O=C(C4CNC4)O>>c5(cc(c(cc6)cc7c6nc(C8CNC8)o7)cc9)c9[nH]cc5

Reactant_1 cost: 17.32

Reactant_2 cost: 0.19

Reaction Yield: 70%

Product cost: 25.01

Step 3:

ClCC1CC1.c2(cc(c(cc3)cc4c3nc(C5CNC5)o4)cc6)c6[nH]cc2>>c78n(CC9CC9)ccc7cc(c(cc%10)cc%11c%10nc(C%12CNC%12)o%11)cc8

Reactant_1 cost: 0.33

Reactant_2 cost: 25.01

Reaction Yield: 75%

Product cost: 33.80

Step 4:

O=C(C1CNC1)O.c23n(CC4CC4)ccc2cc(c(cc5)cc6c5nc(C7CNC7)o6)cc3>>O=C(C8CCC8)N(C9)CC9c%10oc%11cc(c%12ccc%13n(CC%14CC%14)ccc%13c%12)ccc%11n%10

Reactant_1 cost: 0.19

Reactant_2 cost: 33.80

Reaction Yield: 75%

Product cost: 45.32

SCENT (C)

Step 1:

O=C(C1CNC1)O.Nc2c(O)cc(Br)cc2>>Brc(cc3)cc4c3nc(C5CNC5)o4

Reactant_1 cost: 0.19

Reactant_2 cost: 0.43

Reaction Yield: 70%

Product cost: 0.89

Step 2:

Brc(cc1)cc2c1nc(C3CNC3)o2.OB(c(cc4)cc5c4[nH]cc5)O>>c6(cc(c(cc7)cc8c7nc(C9CNC9)o8)cc%10)c%10[nH]cc6

Reactant_1 cost: 0.89

Reactant_2 cost: 12.56

Reaction Yield: 75%

Product cost: 17.93

Step 3:

ClCC1CC1.c2(cc(c(cc3)cc4c3nc(C5CNC5)o4)cc6)c6[nH]cc2>>c78n(CC9CC9)ccc7cc(c(cc%10)cc%11c%10nc(C%12CNC%12)o%11)cc8

Reactant_1 cost: 0.33

Reactant_2 cost: 17.93

Reaction Yield: 75%

Product cost: 24.35

Step 4:

O=C(C1CNC1)O.c23n(CC4CC4)ccc2cc(c(cc5)cc6c5nc(C7CNC7)o6)cc3>>O=C(C8CCC8)N(C9)CC9c%10oc%11cc(c%12ccc%13n(CC%14CC%14)ccc%13c%12)ccc%11n%10

Reactant_1 cost: 0.19

Reactant_2 cost: 24.35

Reaction Yield: 75%

Product cost: 32.72

SCENT (C+D)

Step 1:

Reactant_2 (c2(cc(c(cc3)cc4c3nc(C5CNC5)o4)cc6)c6[nH]cc2) was taken from the Dynamic Library

ClCC1CC1.c2(cc(c(cc3)cc4c3nc(C5CNC5)o4)cc6)c6[nH]cc2>>c78n(CC9CC9)ccc7cc(c(cc%10)cc%11c%10nc(C%12CNC%12)o%11)cc8

Reactant_1 cost: 0.33

Reactant_2 cost: 17.93

Reaction Yield: 75%

Product cost: 24.35

Step 2:

O=C(C1CNC1)O.c23n(CC4CC4)ccc2cc(c(cc5)cc6c5nc(C7CNC7)o6)cc3>>O=C(C8CCC8)N(C9)CC9c%10oc%11cc(c%12ccc%13n(CC%14CC%14)ccc%13c%12)ccc%11n%10

Reactant_1 cost: 0.19

Reactant_2 cost: 24.35

Reaction Yield: 75%

Product cost: 32.72

Example 2, SMALL settings.

SCENT (w/o C)

Step 1:

O=C([C@H]1CCCNC1)O.O=C([C@H]2CCCN2)O>>O=C([C@H]3CCCN3C([C@H]4CCCNC4)=O)O

Reactant_1 cost: 19.37

Reactant_2 cost: 0.10

Reaction Yield: 75%

Product cost: 25.97

Step 2:

O=C([C@H]1CCCN1C([C@H]2CCCNC2)=O)O.NCc(c[nH]3)c4c3cccc4>>O=C([C@H]5CCCN5C([C@H]6CCCN(C(NCc(c[nH]7)c8c7cccc8)=O)C6)=O)O

Reactant_1 cost: 25.97

Reactant_2 cost: 16.81

Reaction Yield: 75%

Product cost: 57.04

Step 3:

O=C([C@H]1CCCN1C([C@H]2CCCN(C(NCc(c[nH]3)c4c3cccc4)=O)C2)=O)O.NC5CC5>>O=C([C@H]6CCCN6C([C@H]7CCCN(C(NCc(c[nH]8)c9c8cccc9)=O)C7)=O)NC%10CC%10

Reactant_1 cost: 57.04

Reactant_2 cost: 0.02

Reaction Yield: 75%

Product cost: 76.07

Step 4:

O=C(NC1CC1)[C@H]2CCCN2C([C@H]3CCCN(C3)C(NCc4c5ccccc5[nH]c4)=O)=O.ClCC6CC6>>O=C(NC7CC7)[C@H]8CCCN8C([C@H]9CCCN(C9)C(NCc%10c%11ccccc%11n(CC%12CC%12)c%10)=O)=O

Reactant_1 cost: 76.07

Reactant_2 cost: 0.33

Reaction Yield: 75%

Product cost: 101.87

SCENT (C)

Step 1:

NC1CC1.O=C([C@H]2CCCN2)O>>O=C([C@H]3CCCN3)NC4CC4

Reactant_1 cost: 0.02

Reactant_2 cost: 0.10

Reaction Yield: 75%

Product cost: 0.16

Step 2:

O=C([C@H]1CCCN1)NC2CC2.O=C([C@H]3CCCNC3)O>>O=C([C@H]4CCCN4C([C@H]5CCCNC5)=O)NC6CC6

Reactant_1 cost: 0.16

Reactant_2 cost: 19.37

Reaction Yield: 75%

Product cost: 26.04

Step 3:

O=C([C@H]1CCCN1C([C@H]2CCCNC2)=O)NC3CC3.NCc(c[nH]4)c5c4cccc5>>O=C([C@H]6CCCN6C([C@H]7CCCN(C(NCc(c[nH]8)c9c8cccc9)=O)C7)=O)NC%10CC%10

Reactant_1 cost: 26.04

Reactant_2 cost: 16.81

Reaction Yield: 75%

Product cost: 57.14

Step 4:

O=C(NC1CC1)[C@H]2CCCN2C([C@H]3CCCN(C3)C(NCc4c5ccccc5[nH]c4)=O)=O.ClCC6CC6>>O=C(NC7CC7)[C@H]8CCCN8C([C@H]9CCCN(C9)C(NCc%10c%11ccccc%11n(CC%12CC%12)c%10)=O)=O

Reactant_1 cost: 57.14

Reactant_2 cost: 0.33

Reaction Yield: 75%

Product cost: 76.63

SCENT (C+D)

Step 1:

NC1CC1.O=C([C@H]2CCCN2)O>>O=C([C@H]3CCCN3)NC4CC4

Reactant_1 cost: 0.02

Reactant_2 cost: 0.10

Reaction Yield: 75%

Product cost: 0.16

Step 2:

O=C([C@H]1CCCN1)NC2CC2.O=C([C@H]3CCCNC3)O>>O=C([C@H]4CCCN4C([C@H]5CCCNC5)=O)NC6CC6

Reactant_1 cost: 0.16

Reactant_2 cost: 19.37

Reaction Yield: 75%

Product cost: 26.04

Step 3:

Reactant_2 (NCc1c2ccccc2n(CC3CC3)c1) was taken from the Dynamic Library

O=C([C@H]1CCCN1C([C@H]2CCCNC2)=O)NC3CC3.NCc4c5ccccc5n(CC6CC6)c4>>O=C(NC7CC7)[C@H]8CCCN8C([C@H]9CCCN(C9)C(NCc%10c%11ccccc%11n(CC%12CC%12)c%10)=O)=O

Reactant_1 cost: 26.04

Reactant_2 cost: 22.86

Reaction Yield: 75%

Product cost: 65.20

Example 3, MEDIUM settings.

SCENT (w/o C)

Step 1:

N#CCCNc1cc(OCCO2)c2cc1.Cn3cc(CNc4cc(C(O)=O)ccc4)cc3>>Cn5cc(CNc6cc(C(N(c7cc(OCCO8)c8cc7)CCC#N)=O)ccc6)cc5

Reactant_1 cost: 125.31

Reactant_2 cost: 279.35

Reaction Yield: 75%

Product cost: 539.55

SCENT (C)

Step 1:

N#CCCNc1cc(OCCO2)c2cc1.Nc3cc(c4cc(C(O)=O)ccc4)ccc3>>N#CCCN(C(c5cc(c6cc(N)ccc6)ccc5)=O)c7cc(OCCO8)c8cc7

Reactant_1 cost: 125.31

Reactant_2 cost: 27.27

Reaction Yield: 75%

Product cost: 203.45

We truly value the reviewer’s detailed feedback and thoughtful recommendations. We have integrated all the suggestions, with the GPU resource metrics, cost model performance analysis, and the hyperparameter sensitivity study significantly enhancing the quality of our paper. We have addressed all the concerns raised and ask the reviewer to take these revisions into account during their evaluation.

Computational resources

As requested by the reviewers, we report GPU runtimes (in minutes) for all baselines on the sEH proxy, averaged over 3 random seeds using a V100 32GB GPU. All methods used a batch size of 64, except SynFlowNet, which required batch size 32 on the LARGE setting even after optimizing its policy code to fit within memory limits.

RGFN shares a part of the codebase with SCENT, allowing for direct comparison. We observe consistent improvements across all settings, particularly in the SMALL configuration. While SynFlowNet is faster on SMALL and MEDIUM, it struggles on LARGE, even with our memory optimizations. However, differences in codebases make direct comparisons less definitive.

Note that our implementation is not optimized for runtime and could benefit from further improvements.

Scaling with number of reactions

We agree that scaling reaction-based generative models with the number of reaction templates is an important factor, omitted from the paper. However, curating a large, high-quality reaction set is a resource-intensive process that cannot be completed within the rebuttal period. Even with automated extraction (which would reduce reaction quality) the computational demands of evaluating all baselines exceed what is feasible in one week. We believe that our experiments already demonstrate the merits of our approach. While scaling with more templates could add value, it would not alter the key conclusions of our work - we propose this as a promising direction for future research.

Lack of limitation section

We apologize for the lack of limitation section, we are going to include the following limitation summary in the revised paper:

“While our study demonstrates the effectiveness of SCENT in generating cost-efficient and synthesizable molecules, several limitations remain.

First, the synthesis cost estimates are simplified and heuristic, omitting factors such as solvent usage, purification steps, hazardous reagents, reaction conditions, and scalability. Consequently, the current cost model should be regarded as a first-order approximation. More accurate estimation would require lab-specific calibration, including robust yield prediction models and vendor-specific building block pricing.

SCENT’s performance varies across different biological targets, indicating that its effectiveness may depend on task-specific factors. Identifying and addressing these bottlenecks is an important direction for future work. Key areas of improvement include refining the exploitation behavior of the Dynamic Library, and enhancing the generalizability of the cost predictor model used in Synthesis Cost Guidance. Additionally, we do not assess performance as the number of reaction templates increases, an important scalability consideration, as this expansion significantly enlarges the action space and introduces additional learning challenges.

Our framework emphasizes synthesizability and cost, but does not incorporate toxicity or broader ADMET properties, which are critical for downstream drug development. Finally, we acknowledge the absence of wet-lab validation, which remains the definitive benchmark for evaluating the practical utility of our approach. We leave this to future work.”

Why SCENT underperforms on JNK3?

Thank you for the thoughtful feedback. One notable characteristic of the JNK3 task is its sparse reward landscape - the number of discovered modes is significantly lower than in sEH and GSK3β. We hypothesize that in such settings, excessive exploitation driven by strong cost guidance may hinder mode discovery. Reducing the Synthesis Cost Guidance temperature ($\alpha_S$) could help balance exploration and exploitation, potentially improving diversity at some cost trade-off. SCENT was not specifically tuned for JNK3, and experiments to test this hypothesis are ongoing - we will share the results as soon as they are available.

Hyperparameters

To assess SCENT (C+D+P) sensitivity to hyperparameters, we performed an ablation study on sEH MEDIUM, varying each hyperparameter around the paper’s chosen values (bolded in tables) over 3k iterations. Results indicate SCENT is generally robust, with no catastrophic performance drops. However, some hyperparameters notably affect outcomes; for example, simple tuning of $\alpha_D$ can boost discovered modes by ~20%. This highlights potential for further improvements through targeted tuning, left for future work. Detailed results follow.

Why Dynamic Library underperforms in some settings?

While it is true that the dynamic library sometimes shows lower mode discovery in the online setting, it consistently yields higher-reward samples on average during inference. This behavior arises because dynamic library updates can occasionally favor fragments that bias exploration toward a narrower region of the search space. While this region may be high-reward, this comes at the cost of reduced diversity, which explains the lower mode discovery numbers highlighted in the review.

This trade-off suggests that the main limiting factor is the fragment selection criterion. Carefully designing this criterion to maintain diversity while still encouraging promising fragments is key to improving exploration and fully realizing the benefits of the dynamic library.

Performance of cost models

Thank you for the helpful suggestion regarding the cost prediction model. To better understand its performance, we trained it separately on a subset of molecules from SCENT trajectories in the SMALL (C) setting. For validation, we used a held-out set from the same setting and computed ground-truth costs using the full recurrence from Equation 7, whereas the training set relied on approximate estimates to simulate the online learning scenario.

We observed steady improvement during training, with validation R² peaking at 0.77 before overfitting set in, eventually declining to ~0.6, while training R² continued to rise to 1.0. While we’re unable to include plots at this rebuttal stage, we’ve added them to the revised version of the paper.

It’s worth noting that in the online setting, the model is trained on trajectories sampled by SCENT itself, helping it learn to identify low-cost backward trajectories. However, during inference, SCENT only uses forward sampling and does not rely on the cost predictor. Thus, while the model’s validation performance may influence trajectory selection during training, it does not affect SCENT’s inference procedure directly.

Consequences of inaccurate prediction model

The consequences of inaccurate cost predictions are assessed in Appendix G.2.2. We compared our model with a low-quality cost estimator that outputs a constant value for every input molecule, where this constant is the mean cost of all the building blocks in the given setting. It is important to note that we only modify the recursive component of the $C^S$ function from Equation (5), keeping the fragment costs unchanged. As shown in Table 8, this constant estimator can reduce costs and boost SCENT’s performance in MEDIUM and LARGE settings, but the machine learning-based cost model used in the main paper delivers far greater improvements, especially in cost reduction. This confirms that cost prediction quality is crucial for SCENT’s

Following up on our earlier hypothesis regarding SCENT’s underperformance on JNK3, we conducted additional experiments to test whether reducing the Synthesis Cost Guidance temperature ($\alpha_S$) could improve mode discovery. Specifically, we ran SCENT (C + D + P) with $\alpha_S=2$ instead of the default $\alpha_S=5$ on all settings.

The results did not support the hypothesis: lowering the temperature increased synthesis cost but did not consistently improve diversity.

The results clearly refute the hypothesis: lowering the temperature led to an increase in synthesis cost but did not produce consistent improvements in diversity.

While we do not have a definitive explanation for the limited diversity gains on JNK3, it is important to emphasize that SCENT’s performance remains robust. The number of discovered modes on JNK3 is comparable to RGFN, yet SCENT achieves this with a lower synthesis cost, demonstrating its efficiency even in challenging sparse-reward environments. We will incorporate these findings into the final version of the paper.

We sincerely appreciate the reviewer’s thoroughness and insightful suggestions. We have incorporated all of them and believe that the warm-start study and qualitative analysis of the pathways, in particular, have significantly strengthened our paper. Regarding the concerns raised, we have carefully addressed each one and kindly encourage the reviewer to consider this in their evaluation.

Warm-start study

We found the “warm-start” scenario compelling and conducted an experiment on SCENT (C+D+P) to evaluate system robustness to changes in fragment costs.

At iteration 3000, we randomly reshuffled fragment costs and continued training without resetting model parameters, only updating the costs estimates used to train the cost predictor. As expected, the trajectory cost spiked immediately after the change but quickly decreased, converging to the same level reached by a model trained from scratch with the new cost assignments (within ~1000 iterations). This behavior was consistent across three random seeds and both SMALL and MEDIUM settings.

Conclusion: The system does not require full retraining when costs change. Updating the cost predictor is sufficient.

Plots showing the evolution of cost and reward over time have been added to the revised paper (not included here due to rebuttal format constraints).

How Dynamic Library expands the reachable set?

It is true that, by definition, any intermediate molecule is reachable via some trajectory. However, once a molecule is added to the dynamic library, it becomes available as a reactant. This enables synthesis paths (trajectories) to become non-linear, as a reactant can now be a previously synthesized molecule—rather than being limited to the pre-defined reactant pool. This is illustrated in Figure 3: the molecule highlighted in red is itself a synthesis tree and could not have been reused as a reactant without the Dynamic Library. More formally, SCENT without the dynamic library is restricted to synthesis trees in which each node has at most one non-leaf child. Under this constraint, not all molecules are reachable. To experimentally validate that, we randomly generated 10000 full synthesis trees (balanced binary trees) using fragments and templates from MEDIUM settings, and 43% of the obtained molecules were observed to be non-decomposable with standard tree-like trajectory, confirming our claim.

Dynamic Library hyperparameters sensitivity

We conducted an extensive ablation study to assess the sensitivity of SCENT to hyperparameter choices across all components. Detailed results are provided in the response to Reviewer QKfB (Section Hyperparameters). We find that reducing the number of intermediates can slightly improve performance, while changing the update frequency of the dynamic pool has negligible effect. Overall, SCENT exhibits strong robustness to the hyperparameters of the Dynamic Library.

How are the reaction yields obtained?

Thank you for noting the imprecision in Appendix C. For reactions where sufficient in-house experimental data was available (amide coupling, nucleophilic additions to isocyanates, Suzuki reaction, Buchwald–Hartwig coupling, Sonogashira coupling, and azide–alkyne cycloaddition), yields were calculated as the average of all recorded experimental yields corresponding to each SMARTS-based reaction template. For the remaining reactions in the dataset, yield estimates were derived from the average of reported literature yields for reactions matching the respective SMARTS templates.

Constant yields

We fully acknowledge the reviewer's concern that our yield assignment method is simplified. However, in our view, this simplification is necessary, as no more accurate and universally applicable yield estimation methods currently exist [1]. Importantly, our cost‑biasing framework is fully modular and can accommodate any recursively defined cost estimation, including those with sophisticated, laboratory-specific yield‑prediction models.

[1] “When Yield Prediction Does Not Yield Prediction: An Overview of the Current Challenges”

Solvents and reagents are ignored

Including solvent and reagent costs in template cost estimation is a valuable suggestion and may be important in practical applications. However, accurately estimating these costs is nontrivial, as they are often lab-specific and less standardized than building block prices. In our work, we focus on building blocks, which typically dominate overall cost, and adopt a simplified approximation that excludes solvents and reagents. While this may reduce accuracy for any specific lab setting, it offers a more robust and broadly applicable estimate.

Importantly, our framework is flexible: more detailed or lab-specific cost models, including those that incorporate reagent pricing, can be readily integrated. We view this as a promising direction for future work.

Pathways Analysis

To visualize the cost guidance in practice, we gathered training trajectories from SCENT (C) and SCENT (w/o C) that lead to the same high-rewarded molecule in the SMALL setting. We added them to revised paper, and included them in a form of Reaction SMILES in response to Reviewer ptir under Pathways Examples section.

We observe that, on average, cost-guided trajectories are ~10% cheaper. We selected a few example trajectory pairs and noticed that the cost reduction usually occurs by changing the order of the reactions: introducing expensive fragments later in the synthesis path, which decreases the product losses due to imperfect yields, and as a result increases the cost-efficiency. In Example 1 (response to Reviewer ptir under Pathways Examples**)**, SCENT (C) chooses the most expensive building block in the sequence (OB(O)c1ccc2[nH]ccc2c1) at the beginning of the synthesis, while SCENT (C) utilizes it in the second step, effectively reducing the synthesis cost by almost 28%. Cost-guided model also occasionally prefers cheaper fragments, e.g., Brc1c[nH]cn1 instead of Ic1c[nH]cn1. It is worth emphasizing that the fragments in the SMALL setting are already low cost, which is the reason why most gains are observed due to reaction order swapping.

Similarly, we compared the Dynamic Library, SCENT (C+D), to SCENT (C), and observed that on average the trajectories are 12% shorter and 6% cheaper. We selected few examples that visualized the compressed trajectories. We have found that the Dynamic Library enables the generation of convergent synthetic pathways (Example 2), which are responsible for the reduction of the synthesis cost. The application of the convergent synthetic strategy decreases the longest linear sequence (LLS) of steps, thereby increasing the overall yield.

In the MEDIUM settings, the SCENT variations diverged early on, so the set of shared molecules is very limited. Therefore, we matched high-rewarded molecules from SCENT (C) to closest high-rewarded molecule from SCENT (w/o C) with Tanimoto similarity > 0.6. We observed that, on average, the cost-guided trajectories are 20% cheaper. Thorough investigation of few selected trajectories pairs revealed that, differently from SMALL setting, in MEDIUM setting, the cost reduction comes mostly from the selection of cheaper fragments (Example 3).

No limitation section (and future work)

We apologize for the lack of limitation section, we are going to include the following limitation summary in the revised paper:

“While our study demonstrates the effectiveness of SCENT in generating cost-efficient and synthesizable molecules, several limitations remain.

First, the synthesis cost estimates are simplified and heuristic, omitting factors such as solvent usage, purification steps, hazardous reagents, reaction conditions, and scalability. Consequently, the current cost model should be regarded as a first-order approximation. More accurate estimation would require lab-specific calibration, including robust yield prediction models and vendor-specific building block pricing.

SCENT’s performance varies across different biological targets, indicating that its effectiveness may depend on task-specific factors. Identifying and addressing these bottlenecks is an important direction for future work. Key areas of improvement include refining the exploitation behavior of the Dynamic Library, and enhancing the generalizability of the cost predictor model used in Synthesis Cost Guidance. Additionally, we do not assess performance as the number of reaction templates increases, an important scalability consideration, as this expansion significantly enlarges the action space and introduces additional learning challenges.

Our framework emphasizes synthesizability and cost, but does not incorporate toxicity or broader ADMET properties, which are critical for downstream drug development. Finally, we acknowledge the absence of wet-lab validation, which remains the definitive benchmark for evaluating the practical utility of our approach. We leave this to future work.”

Unrealistic one-step fragments

The reaction Dynamic Library does take into account the cost of multi-step synthesis. Each intermediate molecule is assigned the lowest-cost synthesis path identified at the time it is added to the building block set. We acknowledge that this important detail was not clearly stated in the original manuscript, and we are grateful for the reviewer’s observation. We will clarify this point in the revised version of the paper.

I appreciate the detailed authors' responses, and after reading the replies to the other reviewers' comments, I believe the changes, particularly those regarding the dynamic library hyperparameter sensitivity and warm-start, will increase the quality of the draft. I maintain my positive outlook on the paper and my positive recommendation remains unchanged. Thank you.