Response to Reviewer 89Qo
We sincerely appreciate your thoughtful response to our rebuttal. We respond to each additional comment in the following content.
---
**In the generation of the Hi-Mol method, are active molecules used as references? If so, have other generative models, including large text-to-molecule models, been fine-tuned based on active molecules?**
For clarification, our proposed hierarchical textual inversion framework and all other baseline generative models in Table 1 of our manuscript are trained using active molecules. On the other hand, for large text-to-molecule models in the initial rebuttal response, we prompt the task description into the model without such fine-tuning, due to the maximum token length of the input prompt. Here, we did not try fine-tuning large text-to-molecule models because fine-tuning such large-scale generative models with low-shot data is known to be difficult to perform well, e.g., see [1, 2, 3]. To address the issue, the textual inversion technique has successfully adopted low-shot data to large-scale generative models in the image domain. Inspired by this recent breakthrough, we carefully design a molecule-specific textual inversion framework for low-shot molecular generation.
Nevertheless, for your information, we fine-tuned large text-to-molecule models with active molecules as reported in the table below, where HI-Mol significantly outperforms them. We added these respects of discussion and the experimental details in Appendix F of our revised manuscript.
\begin{array}{l|cccccc}
\hline
\text{HIV} & \text{Active. $\uparrow$} & \text{FCD $\downarrow$} & \text{NSPDK $\downarrow$} & \text{Valid. $\uparrow$} & \text{Unique. $\uparrow$} & \text{Novelty $\uparrow$} \newline
\hline
\text{MolT5-Large-Caption2Smiles (fine-tuned)} & 6.0 & 23.3 & 0.024 & \textbf{100} & \textbf{96.0} & \textbf{100} \newline
\text{Text+Chem T5-augm-base (fine-tuned)} & 5.8 & 22.4 & 0.023 & \textbf{100} & 95.4 & \textbf{100} \newline
\hline
\textbf{HI-Mol (Ours)} & \textbf{11.4} & \textbf{16.6} & \textbf{0.019} & \textbf{100} & {95.6} & \textbf{100} \newline
\hline
\end{array}
---
**I still think SMILES alone is not enough. Many graph-based generative models can also perform molecular optimization tasks. Examples of such models include Graph-GA, DST, RetMol, etc.**
Molecular graphs are often more preferable at modeling molecular structures because they explicitly consider connectivity between atoms. However, in the recent literature on molecular generation, SMILES-based methods are known to outperform graph-based methods, e.g., as observed and discussed in [4]. This is because (a) there is “enough” information in a SMILES string to define a molecule (i.e., a SMILES string corresponds to a single molecule), and (b) the modeling dimension of SMILES strings (1D) is smaller compared to graphs (2D). In our experiments, we also observed that SMILES-based baselines (e.g., STGG) outperform graph-based methods.
For molecular optimization tasks (as reported in Table 4 of our manuscript), we kindly remind you that we follow the evaluation settings of the most competitive baseline, STGG, to show the wide applicability of HI-Mol. Here, our method still outperforms several graph-based methods (e.g., MHG-VAE, GraphAF, and GraphDF). Also, the mentioned Graph-GA is reported to perform worse than our baseline, PS-VAE. Nevertheless, we will include a comparison with the mentioned molecular optimization works in the final draft.
---
[1] Mo et al., Freeze the Discriminator: a Simple Baseline for Fine-tuning GANs, arXiv 2020\
[2] Zhao et al., On Leveraging Pretrained GANs for Generation with Limited Data, ICML 2020\
[3] Moon et al., Fine-tuning Diffusion Models with Limited Data, NeurIPSW 2022\
[4] Vignac et al., DiGress: Discrete Denoising Diffusion for Graph Generation, ICLR 2023