Procedural Synthesis of Synthesizable Molecules

Thank you so much for acknowledging the importance, vitality and careful execution of our work!

1. Since this work heavily involves discussion on search space, and amortized search within the synthesis tree, similar things are also explored deeply in the retrosythesis community. Thus, it would be beneficial to incorporate some discussions in the paper to discuss the connections and differences between this work and previous work.

• Self-Improved Retrosynthetic Planning, Kim et al., ICML 2021

• Retrosynthetic Planning with Dual Value Networks, Liu et al., ICML 2023

• Retrosynthesis Zero: Self-Improving Global Synthesis Planning Using Reinforcement Learning, Guo et al., JCTC 2024

Thank you for pointing us to these interesting works! We have read these works and included a thorough discussion of the similarities and differences in App. D.3. We are familiar with and acknowledge that the retrosynthesis community has adopted sophisticated techniques like iterative policy improvement under self-play and we included discussion of potential ways our work can incorporate them in the next iteration. We also included a deeper dive into how the difference between forward and retrosynthesis informed the development of our method in App. D.2. We hope you find the discussions interesting!

2. The paper is clear and well-written. However, it would be helpful if the introduction includes a reference to the detailed description of the two tasks in the related work section. This will help readers understand the tasks before they delve into the techniques used.

Thank you for this helpful suggestion! We added a few sentences in the Introduction with references to prior works that describe the tasks more precisely. We also modified certain descriptions, e.g. “synthetic pathways” => “synthetic procedures” to avoid ambiguity with the retrosynthesis problem (since sometimes “pathways” is used to describe backward templates). We’ve also expanded Section 3.1 to define the problems more precisely for our program synthesis based formulation.

3. Regarding the experimental setup, could you explain the choice of the 91 reaction templates and the 147,505 building block compounds? Are they forward reaction templates? Do these choices reflect real-world applications in molecular design?

Yes, yes, and yes!

The building block compounds come directly from Enamine, a supplier of purchasable compounds (https://enamine.net/). These building blocks are used to construct make-on-demand libraries by Enamine REAL (REadily AccessibLe) Space [1], so they are reliable starting materials for robust synthesis protocols. The REAL Space has an average synthesis rate of 80% for 1-3 synthesis steps, attesting to the validity of the building blocks.

The 91 reaction templates are curated by Button & Hartenfeller and were defined in close collaboration with experts. They include well-known organic transformations such as the Wittig reaction [2], Buchwald-Hartwig reaction [3], and various functional group conversions. By category, 13/91 are uni-molecular and 78 are bi-molecular reactions. 63 are skeleton formation, 23 for peripheral modifications and the remaining 5 used for either. For the skeleton formation reactions, 45 are ring formation reactions, with 37 heterocycle formations and 8 carbocycles. These templates seek to codify the most robust “rules of chemistry” and approximate real laboratory organic synthesis procedures. They can be found under data/assets/reaction-templates/hb.txt of our anonymous repo, which we just made accessible. The templates are written in SMARTS. You can visualize them with (https://smarts.plus). As a fun sidenote, you can prompt ChatGPT with the template, asking it to describe what it does, and it does a good job describing the reaction and its common use cases. We can contrast this set with the much larger (~380K) set of retrosynthesis templates derived from retrosynthesis databases like USPTO. They can be automatically extracted using tools like RDChiral and complement search algorithms, but they have not been individually vetted to be robust chemical transformations that serve real-world applications. In the (moved) App. D.2., we explain why this matters for developing our methodology.

[1] Grygorenko, Oleksandr O., et al. "Generating multibillion chemical space of readily accessible screening compounds." Iscience 23.11 (2020).

[2] Wittig, Georg, and Georg Geissler. "Zur Reaktionsweise des Pentaphenyl‐phosphors und einiger Derivate." Justus Liebigs Annalen der Chemie 580.1 (1953): 44-57.

[3] John F Hartwig. Transition metal catalyzed synthesis of arylamines and aryl ethers from aryl halides and triflates: Scope and mechanism. Angewandte Chemie International Edition, 37(15):2046–2067, 1998.

Thank you so much for acknowledging the innovation and excellence of our approach!

1. I think the possible weakness is the dependency on the tree and the grammar components. On one hand having a very large amount of templates will increase the computational complexity of the model (it is not clear how for instance the MCMC algorithm would handle this) and on the other hand a more efficient smaller set will not allow to generate all desired solutions

Thanks for raising these points! We study whether a larger set of templates can be incorporated by studying whether our framework extrapolates to unseen templates at inference time. Note this argument relies on the observation that whether the method can handle a much larger template set at test time is reducible to whether the method can extrapolate on any given new, unseen template. We’ve added new sections (4.3.3 and App. B) for a new ablation study. Specifically, we hold out 25% of templates, retrain our model, and see if our surrogate model is still performant. In particular, we set up our study so all original templates can still be used at inference time (they are only excluded from training). This will test whether 1) our method can incorporate unseen and potentially a much larger set of templates for downstream purposes, and 2) whether this matters for the final results. Here are some takeaways from the study:

• Our ablation surrogate model can reconstruct programs from templates it has not seen before to a comparable level of accuracy as the original model.
• Similarity and diversity of generated analogs only drops marginally on the Test Set. What’s surprising is that performance actually held up on the 25% set of templates which were held out from training. We suspect removing the templates induces a regularization effect, implying the model is not underfitting to those templates.
• Performance drops slightly when tested with the ChEMBL, which comprises mostly unsynthesizable molecules. Diversity decreases, to some extent confirming your suspicion.
• Interestingly, the downstream molecular design results are not hurt at all, and in fact sample efficiency is improved in some cases. This suggests the ablation model can still generate sufficiently many analogs within our bilevel optimization framework.

Also, an alternative to MCMC we propose for navigating a large number of templates is to “recognize” the template with a learned predictor (introduced in Section 3.3). Although efficient, we show it maintains respectable analog generation performance for both downstream tasks, as a baseline (Table 1) and an ablation (Table 4).

Also, we added new sections 4.1.3, Appendix F5 and Appendix F7 to discuss the computational complexity and its dependency on the grammar, as well as a new sampling strategy to mitigate the complexity explosion from using larger templates.

(continued in next reply)

Thank you for recognizing the novelty, effectiveness and validity of our work!

1. The current approach uses a limited number of templates, and it is unclear how this framework could be expanded to include a broader range of templates, which could limit its flexibility.

Thank you for bringing up this essential question about our method. A large-scale template set expansion is the natural next step following this initial pilot framework. Before going there, we answer a key question first: does our framework extrapolate well to a broader range of templates at test time? If the answer is yes, we can expand the framework to handle a broader range of templates, demonstrating the flexibility of our method. We added new sections Section 4.3.3 and Appendix B to answer that question. Specifically, we hold out 25% of templates, retrain our model, and see if our surrogate model is still performant. Some findings include:

• Our ablation surrogate model can reconstruct programs from templates it has not seen before to a comparable level of accuracy as the original model.
• Removing a fraction of the templates and all program occurrences from the dataset does not detriment downstream analog generation performance by much.
• We actually see marginal improvements in sample-efficiency for synthesizable molecular design, while upholding similar overall Top k performance. The findings show the framework is flexible enough to include new templates while maintaining similar performance to if the framework had used all the templates during training. We hope this is a positive indication that our framework would be able to handle a much larger set of templates, provided enough similar templates are supplied for training. We leave the template set scaling laws as directions for future work.

2. Although the authors claim efficiency, the paper lacks direct comparisons to demonstrate this advantage against other methods.

Thank you for bringing up another essential question about our method. We apologize for not including this in the initial submission. We hope the list of following additions in our Revision sufficiently address your concern:

• Added 4.1.3: We describe the complexity of dataset construction and training and explain why despite the seemingly high complexity, it is still efficient in practice. We also introduce a new stratified sampling strategy designed to scale to larger templates. It has the desirable linear complexity and emits a tunable parameter C to trade-off efficiency with performance.
• Revised App. A: We include statistics from our dataset to make the discussion around efficiency grounded in actual numbers.
• Added App. F.5: We study the performance using the simplified training strategy across downstream tasks, observing a performance drop when C=1 but competitive performance when C=10, and discuss the implications.
• Added App: F.7: We analyze the computational complexity under the simplified training strategy, observing equivalent big O complexity to SynNet. For C=1, we include convergence plots and do some back-of-the-envelope calculations to show the efficiency of our training. Interestingly, we also show convergence for the standard training is only slower by a factor of 1-2x compared to the simplified training strategy.

3. The comparison between tasks in Section 3.1 could be enhanced with mathematical notation alongside chemistry examples. While the method draws on program synthesis concepts, the explanation may be confusing without using clear chemical illustrations.

We apologize for this omission from our initial manuscript. The revised version includes mathematical notation for the two tasks and elaborates on their mechanics. The description refers to Figure 1 for illustrations of the terminologies used. We hope this improves the readability of that Section. Due to current space limitations, we refer the reader to Figure 3 of a (cited) prior work for chemical illustrations. If this is not enough, let us know and we can do the best to move other parts of the main text around to incorporate our own chemical illustrations.

The primary concern lies in the method's scalability under the current scheme. It remains unclear how the approach would handle a significantly larger design space with more templates, and how its performance would hold up as template diversity increases.

Thanks for bringing this up again. We address these questions with our new study in 4.1.3 and address the implications in App. B. We note our argument relies on the following observation: whether the method can handle a much larger template set at test time is reducible to whether the method can extrapolate on any given new, unseen template. We hope our findings support the rest of our argument following this premise.

Feel free to let us know if you have further suggestions or comments!

Thank you for recognizing the many strengths of our work!

1. This paper failed to mention the source code / anonymous repository and also in Appendix E . 6 ATTENTION VISUALIZATION figure number is missing.

Here’s our anonymous repo! We’ve included detailed instructions, illustrations and commands in the README. We added the correct label to the figure -- nice catch!

2. Results are compared against the 2022 paper; The authors have not compared the results against any recent publications.

Here are additional comparisons we did with ChemProjector [1] and SynFormer [2], recent follow-up works to the 2022 SynNet paper that propose Transformer-based solutions for synthesizable analog generation (ChemProjector, SynFormer) and synthesizable molecular optimization (SynFormer). Note that a fair comparison of SynthesisNet with them is tricky due to a few factors (that work to their favor):

• The building block set is supplied by Enamine, which regularly updates their catalogues. Both works use the Oct. 2023 version which contains 211,220 molecules, but that set is not publicly released by them. Plus, that version has also been updated since. Meanwhile, we use the older version adopted by SynNet. Post discussion period, we will update the final (hopefully stronger) results on the newest catalogue data for the camera ready.

• The reaction template set used by SynFormer is a different, larger set (115 templates) than the set used by SynNet, ChemProjector and Ours (91). They one-hot encode the templates, so it cannot be changed at inference time.

We tried to resolve these factors as follows:

• To standardize comparisons, we tried retraining those models on the building block and reaction sets used by SynNet and us. However, SynFormer reported doing distributed training across 8 A100 GPUs for over 6 days, spanning 730k training steps. Our work required nowhere close to that much…. For reference, our models were trained on a single NVIDIA RTX A6000 and required only a few hours. Unfortunately, we could not get our hands on that much resources so we cannot complete the retraining during the discussion period window.
• Instead, we set up the following comparison. We use their released model checkpoints, but constrain their inference to the building block set used by us and SynNet. This has the following downstream effect:
  • For ChemProjector, ignoring any positive (or negative, but unlikely) transfer from pretraining on a larger superset of building blocks, this is as close to a fair comparison as we can achieve.
  • For SynFormer, the same applies, but they still get to use their larger set of 115 reaction templates.

We will do the best to resolve these disparities by including fair comparisons with SynFormer and ChemProjector in the camera-ready version, but for now, here’s the preliminary results:

Results on Synthesizable Analog Generation (vs ChemProjector)

We run ChemProjector’s default command for projecting molecules. Other than rebuilding the fingerprint index with the smaller building block set, all hyperparameters are the same. Although ChemProjector has the edge on reconstructive similarity, our method generates significantly more diverse (nearly 2x internal diversity) sets of analogs. Since the paper focuses on projecting to synthesizable space and evaluates only on hit expansion tasks, it may not consider the same desiderata for analog generation as us (e.g. localized edits vs structural diversity). However, a common desideratum that is crucial to consider is the simplicity / ease of synthesis, not just synthesizability. Our method controls this by design due to control over the structure of the synthetic pathway, and our MCMC framework biases towards simpler templates, and our results show up to 0.56 lower average SA score.

(continued in next reply)