Text-guided Diffusion Model for 3D Molecule Generation

Dear Reviewer e3Wq,

Thank you for your valuable feedback. We appreciate the time and effort you put into reviewing our paper. Here is our response to your comments:

Comparison with IVLR (W1, W5): We would like to clarify that our work fundamentally differs from IVLR. While both methods aim at generating 3D molecules, we propose a fine-tuning approach using reference geometry for a diffusion model, whereas ILVR is a sampling method that does not require training.

Importance of 3D Molecule Generation (Q1): As discussed in our paper, generating 3D molecular structures provides more detailed and precise information about the molecule's spatial arrangement and its potential chemical properties compared to 2D molecular graphs. This is particularly useful in numerous applications such as drug discovery and materials science.

Advantages of Textual Conditioning over Numerical Conditioning (Q2, Q3, Q5): The superiority of textual conditioning over numerical conditioning lies not in the additional information provided by specific numerical values but in its ability to capture intricate conditions that may not be sufficiently represented by a single or a handful of target properties. For instance, "small HOMO-LUMO gap" as mentioned in Fig 1 is an example where the text provides a more nuanced condition than could be captured by numeric values alone.

Description of Multi-modal Conversion Module & Reference Geometry (W2, Q4, Q5): Our multi-modal conversion module $\Gamma$ comprises a GIN molecular graph encoder and a language encoder-decoder extended from BERT which we pretrained on a randomly selected subset of our PubChem dataset for converting textual prompts to reference geometry $c_\text{P}$. We believe this approach does not restrict the conditioned diffusion model as we do not aim for deterministic output given certain conditions; instead, we strive for exploring the molecular space which aligns with our motivation.

Evaluation Concerns (W6, Q8, W3, W4): The evaluation method we used for assessing the quality of the generated molecules with respect to their conditioned property is a common approach adopted by mainstream methods such as EGNN and EEGSDE. It focuses on the quality of the generated molecules in terms of their conditioned properties rather than the quality of the generated 3D molecular structures. We used the same random splitting method of the dataset and trained the network $\phi$, so the answer for Q7 is no. As for the GEOM-Drugs dataset, its complexity makes it challenging to design specific conditional generation experiments; therefore, we did not include it in our study. We acknowledge that error bars are missing due to dense data tables and numerous experiments. However, improvements in Novelty and Stability are indeed challenging aspects in diffusion-based 3D molecular generation.

Other Considerations (W7, Q6, Q7): Regarding your question about matching $c_\text{P}$ with atoms generated by the diffusion model (Q6), we don't require an exact match in terms of atom count – for cases with fewer atoms, we apply zero-padding for non-corresponding atoms. We have rectified spelling mistakes and unscientific language in our manuscript; a revised version will be uploaded shortly.

We hope these responses address your concerns adequately and look forward to further feedback.

Dear Reviewer ytEe,

Thank you for your insightful comments and questions. We appreciate the time and effort you have invested in reviewing our paper. Please find our responses below:

W1: We understand your concern about the reproducibility of our results due to the unavailability of the text dataset used in Section 4. We have prepared the dataset for open-source release. However, in line with the double-blind review process, we will fully release the dataset upon the official publication of the paper. We believe this will ensure that subsequent research can reproduce and build upon our results. Furthermore, we are continuously maintaining this dataset from the PubChem database which is also constantly updated. This will ensure that the dataset remains relevant and useful for future research.

W2: We apologize for any confusion caused, and we appreciate your feedback. As we mentioned in the main text, valid and stable 3D molecules can hardly be obtained directly from $c_P$, which makes it difficult to evaluate it with a metric suitable for molecular generation. We ubderstand your concerns regarding the complexity of our novel approach of incorporating text guidance in diffusion-based 3D molecular generation. To substantiate the efficacy of our model design, we have performed an array of experiments, detailed in the experimental section. We concur that refining and simplifying the model structure are worthwhile pursuits, and we are diligently working towards this end. It is our hope to furnish more corroborative evidence in our forthcoming work.

Q1: We appreciate your attention to detail. Indeed, there was an error in our reference to EEGSDE. It was incorrectly cited as (Igashov et al., 2022). We have corrected this in our revised paper and the reference now points to the correct source. We apologize for this oversight and thank you for bringing it to our attention. The revised paper will be uploaded shortly, and we kindly request you to review it at your convenience.

Once again, we thank you for your valuable comments and queries which have helped improve our work. We look forward to your further feedback.

Dear Reviewer Ced2,

Thank you for your insightful comments. Here are our responses:

Q1: We appreciate your interest in our dataset. We have curated a text-molecule dataset from PubChem, one of the most comprehensive databases for molecular descriptions. It collects annotations from various sources including ChEBI, LOTUS, and T3DB, each focusing on different molecular properties. Our dataset comprises 330K text-molecule pairs, which we believe sufficiently represent a wide diversity of molecular structure and properties. The dataset was assembled as follows:

1. We downloaded textual molecule descriptions using PUG View, a REST-style web service provided by PubChem. This yielded over 382,000 annotations.
2. The downloaded record descriptions were sorted according to the isomeric SMILES of molecules, resulting in 330,212 unique molecules with descriptions from multiple data sources within the PubChem database.
3. To standardize the dataset, we replaced IUPAC names or synonyms at the beginning of each description with "The molecule ...".
4. Finally, we used RDKit to convert the SMILES into graph-structure data paired with their textual descriptions.

Q2: For pre-training our multi-modal module, we utilized 300K text-molecule pairs from the Pubchem dataset following a methodology similar to CLIP's architecture. The Graph Encoder is based on GIN and KV-PLM (based on BERT) is used as the text encoder. The alignment between graph representation and text representation is achieved via InfoNCE loss. Subsequently, we trained a multi-modal decoder using over 10K text-molecule pairs to generate molecular graph structures based on corresponding texts. For generating reference geometry information, we identified functional groups using RDKit toolkit and created 3D conformers accordingly.

Q3: We adopted the experimental approach from EEGSDE and compared our method with it under multiple conditions in Table 3. Due to the different numerical values and scales of various properties, it's challenging to evaluate multiple numerical properties simultaneously. Moreover, due to the computational cost of training diffusion models, reproducing all baselines on a new metric is difficult. We acknowledge this limitation and plan to address it in future work.

Dear Reviewer wYzV,

Thank you for your insightful comments and questions. We welcome this opportunity to provide further clarification on these points.

(a) In our model, we do not sample the number of atoms separately. For cases where the number of atoms is relatively small, we resort to padding with a value of 0, which does not correspond to any atom.

(b) Our multi-modal module is pretrained using 300,000 text-molecule pairs from the PubChem dataset. Specifically, we adopt an architecture akin to CLIP, utilizing GIN as our Graph Encoder and KV-PLM (based on BERT) as our text encoders. We employ $L$ transformer layers equipped with self-attention (SA) and feed-forward network (FFN) blocks to generate hidden states and pool them into representations. Subsequently, we align the graph representation and text representation using an InfoNCE loss function. After this step, we train a multi-modal decoder on more than 10,000 text-molecule pairs to generate molecular graphs based on corresponding texts. Lastly, by leveraging RDKit's functionality for identifying functional groups and generating 3D conformers, we create a reference geometry.

(c)&(d) We apologize for any confusion caused by Figure 2 --- due to computational constraints, we utilized a pretrained EDM model (not involving text), which was not depicted in Figure 2's illustration of training stages. During training, only $\theta$ is further trained. It's crucial to note that our objective was to propose a framework for a text-guided diffusion model for 3D molecule generation; in future work, we aim for more comprehensive training leveraging richer datasets and computational resources. Indeed, the number of text-molecule pairs (330K) that we collected from PubChem is relatively limited and only a portion of these data can yield corresponding 3D conformations.

Concerning Table 2, we concur that the differences among results are minor and could benefit from a significance test. In fact, it is reasonable to expect some variability in the novelty displayed by the diffusion model under different conditions. However, stability, calculated through atomic valence, inherently exhibits considerable fluctuations in experiments. We adhered to the settings of prior work when reporting experimental results and aspire to propose more persuasive evaluation methods in future work.

Lastly, we greatly appreciate your careful reading of our article and the appendix. We have addressed all nitpicks and revised some content in the limitations section. Please refer to these updates upon the upload of our revised paper.

Thank you once again for your constructive feedback.