Efficient Evolutionary Search Over Chemical Space with Large Language Models

In replacing the crossover and mutation operations within GraphGA, did the LLM genuinely perform these operations, or did it retrieve known SMILES strings based on the provided inputs? For mutation, does the LLM modify only certain tokens in the existing SMILES, similar to the actual mutation operation, or does it reconstruct the entire sequence? Regarding crossover, how closely does the LLM's output resemble true crossover?

The reviewer raises excellent points here. In general, LLMs are actually not that good at generating fit molecules from prompting alone (see Appendix C1, Table A3, where prompting LLMs to generate molecules de novo results in essentially a random pool). The question about what edits are actually taking place is a very interesting one – one way to answer this question is to analyze the Tanimoto similarity between molecules after and before editing, vs. with a random molecule and see if the edited molecules are as far away as the random ones, which is analogous to reconstructing the whole sequence. We show these similarities in the table below for all LLMs we used in the study. Because GPT-4 uses crossover instead of mutations, we report the max and min Tanimoto similarities, where max Tanimoto similarity is the closer parent and min Tanimoto similarity is the further parent. In all cases, the Tanimoto similarity between the molecules before and after editing are larger than after editing and with a random molecule, which strongly indicates that the edits are preserving molecular substructures. Interestingly, molecules edited by BioT5 exhibit lower similarity to their pre-edit versions compared to other LLMs, suggesting that BioT5 may reconstruct more of the sequences. We have also added this to Appendix D5 in the updated manuscript. We thank the reviewer again for this interesting suggestion!

The objectives in the PMO benchmark are well-known objectives, with extensive experimental data available. Is there a possibility that the LLM is retrieving and outputting known high-performing molecules rather than generating novel ones? Verifying this would be crucial to assess the true innovative capacity of the approach.

We verify this for open-source LLMs. Specifically, we compared the final molecules generated for the JNK3 task against two comprehensive datasets: ZINC20, which was used to create the training set for BioT5, and PubChem, which served as the training set for MoleculeSTM. Our analysis revealed no overlap between the generated molecules and these datasets, indicating that the models do not retrieve and output known high-performing molecules (also – App. Table A3 shows that LLMs are not great at retrieval in the first place!). We also evaluated the framework's performance in a scenario resembling a discovery setting, where we initialized the optimization process with the best currently known JNK3 inhibitors (Table 3, main paper). Our method outperformed the baseline by identifying better new molecules, demonstrating its potential utility in novel discovery settings.

Building on the previous question, did you intentionally leverage the LLM's existing knowledge to rapidly improve performance, perhaps aiming for a zero-shot learning scenario? If so, this strategy has merit. However, it raises concerns about the method's applicability to entirely new objectives where the LLM lacks prior knowledge. How do you address this potential limitation?

Yes, we intentionally use LLM’s existing knowledge to guide the search process in zero-shot settings without the need for finetuning. The reviewer raises an important point. Following the reviewer’s suggestion, we analyzed which protein-drug interaction tasks are represented in the training datasets of MoleculeSTM and BioT5, as detailed in Appendix Table A14. Notably, their training datasets do not explicitly include captions for two specific targets, DRD3 and Adenosine receptor A2a. However, they do contain information on related concepts, such as other adenosine and dopamine receptors. This suggests that knowledge from related tasks can be leveraged to optimize molecules for novel, unseen targets. A major strength of LLMs lies in their ability to generalize across diverse target functions. When optimizing for a genuinely novel task that the LLM has not encountered before, it is assumed that the model’s training on related tasks (e.g., optimizing for Jun N-kinase 3 while having been trained on other kinases) provides generalizable insights into the functional space of kinases. This transfer learning capability enables meaningful, functionally informed edits. In cases where the task is entirely novel—where no relevant or related information exists in the training data—the model may generate random molecules, akin to random mutation operations. Integrating LLMs into an optimization framework mitigates this limitation, as non-meaningful edits are filtered out.

We appreciate the reviewer’s constructive feedback and questions. We are glad that the reviewer considers our method to be “strategic” and “effective”, leveraging the advantages of both LLMs and GAs. We wholeheartedly agree with the reviewer that this method can “open avenues for further advancements”.

We focus on increasing the breadth of our experiments to new GA settings, as well as additional analyses requested by the reviewer. We have updated our PDF in blue text to highlight these changes and address them in detail below.

While the main contribution appears to be the replacement of evolutionary algorithm (EA) components, the paper lacks a comprehensive analysis of existing EA methods. Particularly in multi-objective optimization, evolutionary algorithms like MOEA/D and NSGA-III [1, 2]. The study introduces Pareto Optimality (PO) using the Pareto Front but does not compare it with these established multi-objective optimization EAs. Additionally, the paper DyMol [3] has proposed a Pareto Sampling method similar to the PO in this study, which should be acknowledged.

We thank the reviewer for bringing these works to our attention. We agree that there are numerous intriguing multi-objective optimization algorithms worth exploring. However, as the reviewer rightly noted, this is not the focus of our contributions in this paper. Here, we have chosen to employ the simplest widely-used multi-objective optimization methods commonly used in practice (PO and summation). To our understanding, PO differs from [1,2] by simply keeping all the solutions in the Pareto set, while the other methods incorporate some selection criteria of clustering for selecting the set. That said, the works mentioned [1,2,3] by the reviewer are indeed compelling; inspired by this, we have added a new related works section on multi-objective algorithms (Appendix A1) where we discuss all these works, as well as some additional ones we found. We believe these approaches would be valuable to investigate in future research.

The MolLEO framework is applied only to GraphGA, which limits its demonstrated versatility. Recent works like Genetic GFN [4] and Saturn [5] have enhanced performance by incorporating crossover and mutation into methods like Augmented Memory and GRU-based GFlowNet. Applying MolLEO to these methods could further validate its effectiveness.

We thank the reviewer for this valuable suggestion. We selected GraphGA as the backbone model because its simplicity as a baseline allows us to isolate and directly evaluate the effects of the LLM-based genetic operators in our study. .

Additionally, we have incorporated LLM-based genetic operators in Augmented Memory and integrated them into other EAs including Genetic GFN [4] and JANUS [6] as suggested by you and Reviewer TwFV. The results are shown below, as well as in Appendix D.2 and Appendix D.3 of our updated manuscript.

We hope these results demonstrate the versatility of our proposed method.

The paper could better articulate why specific objectives were chosen, such as maximizing JNK3 and minimizing GSK3β in objective selectivity tasks. Providing rationale for these choices would strengthen the study's context and applicability.

We appreciate the reviewer's insightful comment. JNK3 and GSK3β are widely used oracles for evaluating binding affinities. One valuable task in drug design is selectively designing molecular structures that exhibit high binding affinity for one target while minimizing affinity for another [7]. In our study, we chose to maximize JNK3 and minimize GSK3β as a possible selectivity design task, although the direction was chosen randomly.

All the score curves in the paper compare GraphGA and MolLEO. Including Augmented Memory in these comparisons would provide a more comprehensive evaluation of MolLEO's performance relative to other established methods.

We agree with this very reasonable point; we have updated Figure A1 in the manuscript to include optimization curves for two additional top-performing methods: Augmented Memory (as requested by the reviewer) and REINVENT (a widely-used RL framework).

In Table 2, the "Sum" appears to represent the sum of AUC top-10 scores. To prevent confusion—since row “sum“ and column “sum“ are denoted same—please clarify this in the caption or relabel it as "Sum(AUC)." Alternatively, consider using a term like "Scalarization" to more accurately reflect the data presented.

We thank the reviewer for this suggestion; this is much better. We have updated the caption in the manuscript.

Since the objectives involve both maximization and minimization, summing their scores may not straightforwardly reflect overall performance. Maybe you applied simple (e.g., using 1−score) score for minimization objectives for scalarization. It's important to explicitly detail these methods to ensure the validity and reproducibility of your results.

The reviewer is correct–for the objectives involving minimization, we applied a simple transformation by using 1−score. Also, we ensure all objectives remain within the range of 0 to 1 by normalizing values to allow for comparability across objectives (except for docking scores; those are not scaled). We have updated these details in the manuscript (Appendix B).

Closing comment

We hope that our responses were sufficient in clarifying all the great questions asked by the reviewer. We sincerely thank the reviewer for their time and consideration, and we kindly encourage them to reevaluate their score if they find that our rebuttal and the new experiments presented warrant it.

References

[1] Zhang et. al. (2007) "MOEA/D: A multiobjective evolutionary algorithm based on decomposition." IEEE Transactions on evolutionary computation

[2] Deb and Jain, (2013) "An evolutionary many-objective optimization algorithm using reference-point-based nondominated sorting approach, part i: solving problems with box constraints." IEEE transactions on evolutionary computation

[3] Shin et. al. (2024) "DyMol: Dynamic Many-Objective Molecular Optimization with Objective Decomposition and Progressive Optimization." ICLR 2024 Workshop on Generative and Experimental Perspectives for Biomolecular Design.

[4] Kim et. al. (2024) "Genetic-guided GFlowNets: Advancing in Practical Molecular Optimization Benchmark." Neurips.

[5] Guo, and Philippe, (2024) "Saturn: Sample-efficient Generative Molecular Design using Memory Manipulation." arXiv preprint arXiv:2405.17066.

[6] Nigam, AkshatKumar, Robert Pollice, and Alan Aspuru-Guzik. "JANUS: parallel tempered genetic algorithm guided by deep neural networks for inverse molecular design." arXiv preprint arXiv:2106.04011 (2021).

[7] Schneuing, Arne, et al. "Structure-based drug design with equivariant diffusion models." arXiv preprint arXiv:2210.13695 (2022).

We thank the reviewer for their positive feedback! We are thrilled the reviewer finds our method to “yield superior performance” and “showcase promise in early drug discovery”.

Below, we address the reviewer’s questions about computational costs as well as combining with other evolutionary mechanisms. We have updated our PDF in blue text to highlight these changes and address them in detail below.

Could the authors provide a computational cost comparison between MOLLEO and traditional methods, given the LLM integration?

The most expensive part of chemistry experiments is often the oracle call. Below, we show the computational costs of MolLEO(BioT5) vs GraphGA for two experiments: JNK3, which uses a lightweight oracle, and docking against human dopamine D3 receptor (PDB ID 3pbl), which is more expensive.

In this table, we show that for lightweight oracles, incorporating LLM edits indeed results in high runtimes since calling the LLM is more expensive than random edits. However, as experiments become more expensive, such as with docking, the cost of the LLM call becomes insignificant in comparison to the docking time, hence the runtime is governed by the oracle call. For MolLEO (GPT-4), the runtime is further constrained by OpenAI API rate limits, which impose restrictions on the number of input/output tokens processed per minute.

How does MOLLEO balance chemical space exploration with optimization, and could additional mechanisms for diversity be beneficial?

In our experiments, we find that incorporating structural filters is really useful, which is an exploitation mechanism. Our filters were inspired by those implemented in [1], which remove any generated molecules that are extremely dissimilar from the current top molecules. The point of incorporating diversity mechanisms is interesting. Per this suggestion, we implemented our operators in a second genetic algorithm framework, JANUS, which maintains an exploration population. We find that incorporating our operators into JANUS improves its performance, which we show below and in Appendix D3. This indicates that our method works synergistically with a range of evolutionary mechanisms.

Closing comment

We are happy to address any additional comments or concerns the reviewer may have. Otherwise, we would appreciate if the reviewer continues to view this work positively. Thank you once again for your time and thoughtful feedback.

References

[1] Nigam, AkshatKumar, Robert Pollice, and Alán Aspuru-Guzik. "Parallel tempered genetic algorithm guided by deep neural networks for inverse molecular design." Digital Discovery 1.4 (2022): 390-404.

I am uncertain about the evaluation function. Could you clarify what is meant by 'black-box'? In my experience, fitness is typically determined either through practical experimentation or by calculation using a specific function.

Certainly, we are happy to provide this clarification. By "black-box," we mean that the function's closed-form expression is unknown, and its internal workings are not directly accessible. In practical scenarios, fitness is typically obtained through experimentation (e.g., a bioassay) or simulation (e.g. DFT). As we cannot assume the availability of a closed-form function or its gradients, our approach treats it as a black-box optimization problem. In our study, we use simulations (protein-ligand docking) and simpler computational models, where we assume we don’t have access to gradients.

How to turn molecule into text description so that it can be processed by LLMs?

The transformation of molecular structures into text-based representations is commonly achieved using the Simplified Molecular Input Line Entry System (SMILES) notation [3]. SMILES provides a compact string representation of 2D chemical structures that is both human-readable and amenable to computational processing. LLMs trained on datasets containing SMILES notations learn the joint probability distribution of these molecular representations alongside corresponding text descriptions of their properties and functions. This allows LLMs to encode both molecular structure and textual information together.

I am curious the scale of the molecule and its text description. If the scale is too big, will LLMs suffers the risk of correctness? I mean, the generated molecule is probably infeasible?

The reviewer's concern about the correctness of molecules generated by LLMs is very reasonable, and is a challenge with generative models in general. As noted in Appendix C.1, we cannot guarantee that an LLM generates a valid molecule with a single query. To address this, we fall back to random mutations if the LLM-generated molecule is invalid; we also experiment with using chemical languages that always guarantee generating valid molecules (SELFIES, which is used in BioT5). This LLM generates the strongest results from the open-source models.

Regarding the feasibility of generated molecules, it is important to note that this is a broader challenge inherent to all generative models, irrespective of molecule size or length. While synthesizability is a recognized issue in molecular generation, it falls outside the primary scope of this work.

Closing comment

We sincerely thank the reviewer for their valuable feedback. We hope our rebuttal has effectively addressed their questions and concerns. If the reviewer is satisfied with our responses, we kindly ask them to consider a reassessment of our paper. We remain more than happy to address any additional questions or concerns that may arise.

References

[1] Microsoft AI4Science et al. The impact of large language models on scientific discovery: a preliminary study using gpt-4. arXiv, 2023.

[2] Nigam, AkshatKumar, Robert Pollice, and Alán Aspuru-Guzik. "Parallel tempered genetic algorithm guided by deep neural networks for inverse molecular design." Digital Discovery 1.4 (2022): 390-404.

[3] Weininger, David. "SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules." Journal of chemical information and computer sciences 28.1 (1988): 31-36.

We thank the reviewer for their feedback and thoughtful questions. We are ecstatic that the reviewer finds our method “innovative” and “enhances proposal quality”. We now turn to address the main points raised by the reviewer, with a focus on improving the clarity of our presentation and addressing the question of surrogate models. We have updated our PDF in blue text to highlight these changes as well.

Based on the description, this work employs LLMs designed solely for text, but Figure 1 appears to suggest the use of an MLLM.

We utilize SMILES/SELFIES representations to convert molecules to textual representations for processing by LLMs, which is consistent with how the underlying LLMs trained their models. The molecule graph depicted in the original figure was included solely for visualization; we do not use MLLMs and apologize for the confusion. Per your suggestion, we have updated Figure 1 to show the conversion from molecule graph to SMILES string before model ingestion.

The presentation of the proposed method should be improved. For instance, there is no justification provided for using different prompts for LLMs like GPT-4, BIOT5, and MOLSTM. Introducing these LLMs could be moved to the appendix, and the comparison of the designed strategies would benefit from a clearer presentation, such as in a table format.

We thank the reviewer for providing these suggestions to improve the clarity of our work, although we think it is valuable to describe how each LLM works in the main body since they rely on different inference schemes (i.e., next-token prediction vs. maximizing cosine similarity using gradient ascent). To design the prompts, we followed the best practices proposed in each of the corresponding models and did not deviate from these. For GPT-4, we used similar prompts as in [1], which tested the model in several chemistry settings. We show all prompts used in Appendix E. Per the reviewer’s suggestion, we have mentioned this explicitly in Section 3.2; we hope the reviewer finds this more clear.

How can we ensure LLMs can produce better offspring based on the description and fitness value?

The reviewer raises an important question – indeed, we cannot guarantee that LLMs produce better offspring based on an objective description. We have documented this in Table A1, where we show that 20-50% of offspring have higher fitness. However, this value is significantly higher than the number of offspring with higher fitness generated by GraphGA (see Figure 2). We also know that LLMs on their own are not able to generate extremely fit molecules from prompting alone (see Appendix C1, Table A3). However, by incorporating LLMs into an optimization framework, it’s perfectly fine if some offspring are not better, since they will not proceed through the evolutionary framework. What’s important is that LLMs produce better offspring with higher likelihood than by random mutations; these offspring will be propagated through the optimization framework to find better molecules faster.

The abstract states that the proposed method addressed the problem of expensive evaluation. However, I didn't see any relevant discussion on it. It is solved by surrogate model normally in evolutionary computation.

For the experiments, we use AUC top-10 as the evaluation metric, which is the area under the curve of top-10 average property values versus the number of oracle calls. This metric takes into account both the objective values and the computational budget spent. MolLEO achieves better performance under the same oracle budget, demonstrating improved sample efficiency compared to traditional approaches.

The reviewer is correct in that surrogate models are indeed commonly used in evolutionary frameworks. At the reviewer’s suggestion, we have integrated our method into JANUS [2], which incorporates surrogate models within its framework to filter out crossover offspring by replacing its mutation operator with our LLM-based MolLEO mutation operator on two tasks, which we show in a table below. We find that our method is able to achieve better performance than JANUS on its own, which suggests that our method can be combined with surrogate models to find better molecules within a fixed experimental budget.

[1] Hiener, Danielle C., and Geoffrey R. Hutchison. "Pareto optimization of oligomer polarizability and dipole moment using a genetic algorithm." The Journal of Physical Chemistry A 126.17 (2022): 2750-2760.

[2] Romera-Paredes B, Barekatain M, Novikov A, et al. Mathematical discoveries from program search with large language models[J]. Nature, 2024, 625(7995): 468-475.

[3] Fernando C, Banarse D, Michalewski H, et al. Promptbreeder: Self-referential self-improvement via prompt evolution[J]. arXiv preprint arXiv:2309.16797, 2023.

[4] Nigam, AkshatKumar, Robert Pollice, and Alán Aspuru-Guzik. "Parallel tempered genetic algorithm guided by deep neural networks for inverse molecular design." Digital Discovery 1.4 (2022): 390-404.

[5] Abramson, Josh, et al. "Accurate structure prediction of biomolecular interactions with AlphaFold 3." Nature (2024): 1-3.

[6] Guo, Jeff, and Philippe Schwaller. "Augmented Memory: Sample-Efficient Generative Molecular Design with Reinforcement Learning." Jacs Au (2024).

[7] Kim, Hyeonah, et al. "Genetic-guided GFlowNets: Advancing in Practical Molecular Optimization Benchmark." arXiv preprint arXiv:2402.05961 (2024).

[8] Nigam, AkshatKumar, Robert Pollice, and Alán Aspuru-Guzik. "Parallel tempered genetic algorithm guided by deep neural networks for inverse molecular design." Digital Discovery 1.4 (2022): 390-404.

Zhu et al. (2024) introduced a method to decompose the optimization process into a progressive sequence based on the order of objectives. Shin et al. integrated GFlowNets with a preference-conditioned sum of objective functions, further advancing the optimization landscape.

One key strength of LLMs is their ability to generalize to several target functions. In the event that we are optimizing for a truly novel task that the LLM has never seen before, the assumption is that the LLM could have been trained on other tasks that are related (for example, say we are optimizing for Jun N-kinase 3 but the LLM has only seen other kinases). These related tasks can still encode some useful information about kinases in general, which can allow for meaningful, functionally-informed operations. This is an example of transfer learning. If the task is truly novel (i.e., all tasks that the LLM has been trained on have orthogonal information), then at worst the LLM spits out random molecules, which can be seen as random mutation operations. This is also the benefit of plugging LLMs into an optimization framework – edits that are not meaningful are filtered out.

We also test going beyond what chemical information is currently known by initializing our optimization framework with the best molecules currently known that are JNK3 inhibitors in Table 3 of the main paper, a situation that most resembles a new discovery setting. We find that our method is able to find better new molecules than the baseline.

The authors employ GraphGA as the backbone model, but why was this specific choice made? Other genetic algorithm-based methods, such as GPBO, or even other generative models like Augmented Memory, could also be suitable.

We selected GraphGA as the backbone model because its simplicity as a baseline allows us to focus on and directly evaluate the effects of the LLM-based genetic operators in our study. Our focus in this work was on understanding the impact of different LLM-based genetic operators within a simple framework.

As the reviewer has rightly suggested, it would be interesting to explore our mutations in other settings. To this end, we have incorporated our LLM-based genetic operators in three additional backbones: Augmented Memory [6], Genetic GFN [7] (as suggested by Reviewer bF3X), and JANUS [8] (which incorporates exploration mechanisms, as suggested by Reviewer TwFV). The results are shown below; we also added them to Appendices D2 and D3 in the updated manuscript.

We hope the reviewer finds these results interesting that LLM-based generation operations are beneficial in several existing frameworks.

Authors do not explicitly explain how they handle the differences between minimization and maximization objectives. Do they simply flip the sign of the minimization objectives to treat them as maximization objectives? If so, is this approach sufficient for effectively balancing objectives with inherently different goals?

We confirm that we do indeed flip the signs of the minimization objectives to treat them as maximization objectives. While this approach is a straightforward transformation, it is a setting that we compare against other methods to treat objectives with the same weight, which is a reasonable assumption with no additional prior information (i.e., determining different weights for different objectives is nontrivial).

In Figure 2, there appears to be a typo where "LLM editting" is used. Should this perhaps be corrected to "LLM editing"?

We thank the reviewer for pointing out the typo in Figure 2; we have corrected it.

Closing comments

We thank the reviewer for the extremely valuable feedback. We hope our rebuttal sufficiently addresses their questions and concerns, and we kindly ask the reviewer to consider a revised evaluation of our paper if the reviewer is satisfied with our responses. We are happy to address any further questions or points of clarification.

The reviewer asks a very interesting question. Our initial hypothesis was that LLMs themselves are a useful deep-learning based approach for generating mutation and crossover operations because they have been trained on large-scale functional text. The cool thing about this is LLMs can link molecular structures with a description of their functions, and the outcome is that molecules with similar functions are closer together in their latent representations.

The challenge with existing deep-learning based operators is that they are typically trained on small-scale datasets for very specific tasks, and introducing a new task might mean needing to re-train the network. On the other hand, the breadth of training data in LLMs also offers an advantage; if there is data scarcity for labels in some task, there can be information shared from other tasks that generate meaningful molecular representations.

In our evaluation, we do compare our method to DNN-based approaches (REINVENT, DST) and find that our method outperforms both. If the reviewer has a specific work they are referring to, we would be happy to take a look.

While the reviewer raises a very reasonable point about computational costs, we find that for tasks where the oracle call is expensive, the call time to the LLM becomes negligible. For example, the docking experiments for DRD3 (PDB ID 3pbl) on average took 8 h 23m ± 1 h 46 m on average for 5 runs for the baseline GraphGA, while it tool MolLEO(BIOT5) 6 h 57m ± 1 h 24 m on average. Thus, the bottleneck in those experiments was not the LLM call, but rather the time to run the docking simulation. We hope this quantitative evaluation provides justification for the reviewer as to why the runtime of LLMs is not prohibitive.

How can we be confident that LLMs naturally incorporate … domain-specific knowledge and execute these crossover and mutation operations accurately, rather than merely presenting empirical results? What evidence supports the notion that LLMs inherently understand and replicate these chemically-informed operations?

We thank the reviewer for raising this valid concern. Our goal is not actually to necessarily replicate these inductive biases. Instead, our motivation is that LLMs are able to provide functionally-informed operations because they were trained on large amounts of paired molecule-function data, and so LLMs can learn a meaningful landscape of molecular structure representations that are conditioned on their functions and encode domain-specific knowledge. Thus, when prompted with a new molecule structure, the LLM can edit the molecule in a way that improves its function, either by gradient descent or maximizing token probabilities.

We measure the ability of LLMs to provide good functionally-informed operations through several ablations and experiments (Figure 2, Table A1) by having our operations perform a single round of edits on a set of molecules and measuring the functions of molecule pairs before and after, and find that they outperform pre-defined, domain-expert operations. This may also suggest that it may not actually be necessary to encode these inductive biases, which interestingly is similar in spirit to what was found in the AlphaFold3 paper [5].

If LLMs have learned about these well-known objectives, they may inherently recognize and prioritize molecular structures that are known to perform well for these targets. If LLMs have learned about … well-known objectives, they may inherently recognize and prioritize molecular structures that are known to perform well for these targets. If the LLM’s prior knowledge is being leveraged intentionally, how does this approach handle genuinely novel molecular objectives that the LLM has never encountered?

We thank the reviewer for these thoughtful and nuanced questions. While LLMs have been trained on extensive chemical literature, they are actually not good at generating top candidate molecules on their own, which we show in App. Table A3. Thus, it is doubtful that they are simply retrieving instances from the training data.

Per the reviewer’s suggestion, we did an analysis for which protein-drug interaction tasks appear in the training datasets of MoleculeSTM and BioT5 in Appendix Table A14. Interestingly, we find that their training datasets do not explicitly contain captions related to two of the targets, DRD3 and Adenosine receptor A2a, although they contain information to related concepts, such as other adenosine receptors and dopamine receptors. This indicates that we can leverage information from related tasks to optimize molecules on new, unseen tasks.

We are extremely appreciative of the reviewer’s feedback and suggestions to improve our work. We are glad that the reviewer finds our work “easy to follow” and that our framework is “flexible” for “different resource constraints”. We are also heartened that the reviewer considers our work to have “potential value for real-world drug discovery pipeline”.

Below, we address the reviewer’s concerns about task familiarity and include several additional baselines, at the reviewer’s suggestions. We also highlight all these changes in blue in the updated PDF.

The core idea of MOLLEO is essentially a straightforward substitution, where genetic operators in the existing evolutionary algorithm framework (Graph-GA) are replaced with LLMs, which could be seen as a relatively simple extension of existing approaches.

We thank the reviewer for this constructive feedback, although we argue that incorporating LLM-based mutations into black-box optimization frameworks has not been done before in scientific discovery to our knowledge. Evolutionary algorithms (EAs) are widely used in chemistry [1], although they traditionally rely on random manipulations of chemical structures, which can be inefficient since chemical space is large and experiments are expensive. LLMs have been trained on large amounts of chemical knowledge and it has previously been shown that they can tweak molecular structures based on text descriptions, and so we wanted to investigate whether or not these tweaks are informative enough to be incorporated into optimization algorithms. Up until now, the incorporation of LLMs in EAs has primarily been shown for code/algorithm design [2, 3], but not in scientific discovery, which we argue is a much harder domain and it was not obvious whether language-guided mutations would be useful. Figuring out how to incorporate them is nontrivial; we did several ablation studies and showed that our model is able to beat many existing methods on an established benchmark and on complicated property optimization tasks, including protein-drug interaction tasks. We hope this answer demonstrates to the reviewer that our work is both novel and useful for the community.

Authors acknowledge that determining the weight for each objective function in a sum-aggregate objective could be challenging, they do not specify how these weights are assigned in their experiments.

In our experiments, we assigned equal weights to each objective in the sum-aggregate objective function. This was to avoid introducing additional hyperparameters since determining objective weights can be nontrivial. We instead compute this simple metric, as has been done in other works [4], and focus on more robust multi-objective optimization algorithms such as Pareto set selection.

[Pareto set selection] is a standard practice in evolutionary algorithms, as seen in well-established methods like NSGA [1] and MOEA/D [2], which leverage Pareto ranking … Though the authors do not claim a novel contribution in using Pareto set selection, the paper would benefit from a more thorough discussion of their approach to multi-objective optimization and clearer positioning within the broader landscape of Pareto-based optimization approaches.

We thank the reviewer for pointing us to these works. We agree that there may be several interesting multiobjective optimization algorithms to explore, although as the reviewer pointed out, this is not a contribution we claim in this work. In this paper, we opt for the most standard multi-objective optimization methods that are widely used in practice. However, the works that the reviewer points out are interesting. To our understanding, PO differs from these by simply keeping all the solutions in the Pareto set, while the other methods incorporate some selection criteria of clustering for selecting the set. Motivated by this, we have created a new related works section on multiobjective algorithms (Appendix A1) and think they would be interesting to investigate in a separate work.

What unique capabilities do LLMs bring to molecular generation that cannot be achieved through existing deep learning-based mutation and crossover operations? The authors may consider that LLMs do not require additional training for genetic operator learning, as they can be adapted simply by modifying prompts. However, I believe that the inference process for LLMs could introduce a computational cost comparable to training deep neural networks for genetic operators. Furthermore, while the authors may suggest that LLMs can facilitate the generation of molecules aligned with objective fitness functions through descriptive prompts, deep neural networks trained specifically for genetic operations can offer distinct advantages. By learning from a dataset of molecules, they can capture meaningful molecular representations and generate novel combinations that are data-driven rather than randomly assembled.