3M-Diffusion: Latent Multi-Modal Diffusion for Language-Guided Molecular Structure Generation

Thanks for your time and feedback!

Q1. These results on the GuacaMol benchmarks are absent from the subsequent sections.

A1. Thanks for your comments! We believe there are misunderstandings. The results of unconditional molecule generation with GuacaMol benchmarks are put in Table 5 of Appendix.

Q2. Why are the evaluation metrics from MolT5 missing?

A2. Thanks for your question! The metrics like BLEU, Exact, Levenshtein employed in MolT5 are used to assess similarity based on SMILES representations rather than graph similarity. We have used MACCS similarity metrics in our paper. Additionally, our paper conduct molecule generation with a focus not just on reproducing known molecules, but on generating diverse and novel molecules that have not yet been discovered, aiming to advance drug discovery.

Q3. The comparison should include baselines in [1]. Table 5 should be moved to the main body.

A3. Thanks for your comments! We have included the baseline from the source code provided in [1]. The results presented below demonstrate that 3M-Diffusion outperforms all baselines across all evaluated metrics.

Our paper mainly focuses on text-guided molecule generation. While unconditional generation isn't our main focus, it serves to validate our model's effectiveness in learning diverse and novel molecules. Therefore, we include the results of unconditional molecule generation in the Appendix rather than in the main body.

Q4. Table 1 and Table 2 have the same caption.

A4. Thanks for your comments! We will correct this problem in our revised version.

We appreciate the reviewer's perception of our contributions to both empirical analysis and motivations, and are grateful for the encouraging comments. Please see our responses below:

Q1. It is more natural for me to understand the necessity of molecule-to-text generation instead of text-guided molecular generation.

A1. Thank you for your comments. Traditional Molecule generation based on value guidance specifying a single or a handful of target properties might be insufficient to capture intricate conditions. Using natural language can greatly utilize biomedical knowledge from humans for molecule generation [1], [2]. Specifically, text-guided molecule generation allows us to encompass conditions adeptly and flexibly, like substructure (benzene ring) or the novel properties like categories of the molecule (“flavones” as shown in Figure 2).

Q2. The technical novelty is limited, but it is acceptable.

A2. Thank you for your feedback on the technical novelty of our work! We believe that our approach introduces several innovative aspects. Specifically, our paper highlights:

• novel problem: we explore the previously unaddressed problem of text-guided molecular graph generation.

• novel perspective: we present an approach from a novel perspective based on diffusion models for the text-guided molecular graph generation.

• novel results: our 3M-Diffusion model successfully generates diverse and high-quality molecules that accurately match textual descriptions, which outperform other baselines for both text-guided and unconditional molecule generation.

Q3. How about using the 3D geometry for molecular modeling and introducing the protein or pocket information as a prompt for drug discovery?

A3. Thanks for your questions! We agree with your opinions. However, our current paper is focused on text-guided molecule generation and this concept falls outside our scope. We acknowledge the potential of this approach and suggest exploring these promising settings in future work.

[1] Edwards, Carl, et al. "Translation between Molecules and Natural Language." EMNLP-2022.

[2] Christofidellis, Dimitrios, et al. "Unifying molecular and textual representations via multi-task language modeling." ICML-2023.

We thank the reviewer for the comments.

Q1. Figure 1 is confusing, the legends in Table 1 are incorrect and Table 1 and 2 could be merged.

A1. Thanks for your comments! We will update the Figure and Tables following your suggestion.

Q2. How to obtain the ground truth graph of a generated molecule if two texts are very similar.

A2. We believe there may be some misunderstandings. Each sample in the dataset consists of a text paired with a molecular ground truth graph (included in the dataset).

Q3. Retrieval tasks should be included.

A3. We present the molecule-to-text retrieval results on PubChem, comparing them with Sci-BERT:

We measure performance with accuracy and recall@20, where our model outperforms Sci-BERT.

Q4. Does the case study from the training set?

A4. No, we select the samples from the test set.

Q5. It is inappropriate to express improvements as the percentage increases.

A5. Thank you for your comments. We will use absolute relative improvement.

Q6. Plot the distribution of logP and QED.

A6. Thanks for your comments! We have plotted the logP and QED distribution on the ChEBI-20 dataset for ground truth data, MolT5-large and 3M-Diffusion in the link. Our model shows smaller or similar distribution differences with the ground truth compared to MolT5-large, highlighting its effectiveness.

Q7. Why is there such a big difference between the prompts in Figure 2 and 3?

A7. Thanks for your questions! The text in Figure 3 is derived from the original dataset, and the property-related words used in Figure 2 is designed by ourselves. We try to explore our model's ability for text-guided molecule generation of different cases with different prompts.

Q8. The paper's statement requires supporting evidence. If the method generate consistent SMILES, would this issue not exist?

A8. Thanks for your question! The structure of ethanol can be specified using CCO, OCC, or C(O)C,' is in the link. While the model delivers consistent results, variations in SMILES can yield molecules with distinct properties, as noted in JT-VAE, making it also unsuitable to use SMILES for molecule representation.

We gratefully appreciate your time in reviewing our paper.

Q1. Overclaim and SOTA.

A1. Thanks for your comments. We add Mol-instruct (ICLR 2024) [1] as a baseline. The results show that 3M-Diffusion outperforms Mol-instruct (Due to word limits, we just put results on PCDs, all other datasets have similar conclusions).

For the unconditional generation results, we put them in Table 5 of the Appendix.

Q2. No technical contribution.

A2. We respectfully disagree that our model isn't novel or has no technical contributions because it is not surprising, and uses components in previous works. We believe that discouraging authors from building on others' work due to fear of reviewer criticism hinders community progress. We argue that we explore text-guided molecule generation from a novel perspective that involves generating molecular graphs based on texts with diffusion models.

While the diffusion model has shown promise in various fields, it has not yet been applied to the novel problem of text-guided molecular graphs. We stress the fact that our method can be used to discover novel, diverse and high-quality molecules.

Q3. Trade-off rather than improvement.

A3. We respectfully disagree that our model's performance is merely a trade-off rather than an improvement. Trade-offs are inherent in molecule generation. Generating molecules that match text is important, but creating diverse and novel molecules is also crucial in drug discovery. A model that only reconstructs known molecules but cannot discover new ones fails to meet drug discovery needs. Therefore, the target of text-guided molecule generation maintains an acceptable level of similarity while ensuring diversity and novelty. Overall, our average results on all metrics outperform other methods as shown below (just put PCDes for world limit and all datasets have similar conclusion) :

[1]Mol-instructions: A large-scale biomolecular instruction dataset for large language models.