GenMol: A Drug Discovery Generalist with Discrete Diffusion

We sincerely appreciate your comments that our contribution could aid drug discovery and GenMol is a generalist. We address your concerns below.

---

One baseline in de novo and fragment-based generation.

We included JT-VAE (Jin et al., 2018) in Table 9, and have provided the results of DiGress in our first response to 5s8Z.

---

No new developments in discrete diffusion. Graph discrete diffusion should be discussed.

GenMol is the first discrete diffusion framework on small molecule sequences, which is superior to graph discrete diffusion that suffers from low validity and efficiency (GenMol vs. DiGress above). We also showed confidence sampling achieves better quality and efficiency in molecular sequence generation. Moreover, we proposed MCG tailored for discrete sequence diffusion.

---

It is unclear how many diffusion timesteps are used.

GenMol uses confidence sampling with $N$ denoting the number of tokens to unmask at each step (line 323). GenMol with $N=1$ uses the same number of steps with AR models, and higher $N$ improves efficiency. In Table 1, GenMol with $N=3$ shows higher quality than the AR model with 2.5x faster sampling.

---

Could the fragment remasking be adapted to AR models?

Yes, but it cannot be applied without relying on a specific generation order in AR models, leading to inferior performance (Table 5).

---

Why are fragments not used as tokens?

Using fragments as tokens is challenging due to the large vocabulary, limited generalizability, and the difficulty of tokenizing attachment points. Applying BRICS to ZINC250k yields 39034 fragments, resulting in an imbalance with sequence lengths of 3~5. It also does not allow the model to generate novel fragments. We have introduced a simpler yet effective approach that generates tokens at atom level and performs the remasking at fragment level.

---

A reason for using SAFE instead of SMILES? Could the approach be adapted to SMILES generation?

SAFE makes fragment remasking simpler and more effective, and adapting the approach to SMILES is non-trivial. As tokens that belong to the same fragment are not contiguous in SMILES, SMILES models need to classify each token into fragments before every remasking and cannot guarantee the re-predicted tokens belong to the same fragment. On the other hand, SAFE allows GenMol to easily identify fragments in a sequence and generate new fragments.

---

Tokenization details are unclear.

We adopt the SAFE tokenizer (Noutahi et al., 2024) (lines 725-726), which is a BPE tokenizer with the SMILES regular expression. It covers 1873 different atoms, bonds, and ions, and thus can be universally used for molecule-related tasks.

---

Do Sections 5.1 and 5.2 utilize the framework in Figure 3a? It is unclear how the iterations in Figure 3 contribute to the quality and efficiency of generation.

No, Sections 5.1 and 5.2 use the process of Figure 2(c1-c5) (line 173, Figure 3(a) caption). The generic quality and efficiency stem from the discrete diffusion process, and the iterations in Figure 3(a) are multiple rounds of the process, not the process itself.

---

What if the number of calls is low or the oracle is incorrect?

We followed Gao et al., 2022 in Section 5.3, where the 10k oracle calls were set for all baselines. Nevertheless, we observed that GenMol consistently performed better in both low and high oracle call regimes, so we hypothesize that GenMol would perform better with fewer calls or an incorrect oracle, and will include this in the revision.

---

Only two baselines in lead optimization. Missing values in Table 4.

In Table 4, baseline selection and the use of the ‘-’ symbol to indicate failure (no successful leads generated) followed Wang et al., 2023 (lines 412-413).

---

The same name was used for the model and the framework.

GenMol is the framework that uses discrete diffusion in versatile ways for diverse drug discovery tasks. GenMol adopts the sampling of Figure 2(c1-c5) for de novo and fragment-constrained generation and adds fragment remasking of Figure 3(a) for goal-directed optimization.

---

How is the molecular context defined?

See our second response to 5s8Z.

---

The fragment vocabulary is not sufficiently discussed.

The training SAFEs are based on BRICS following Noutahi et al., 2024. During inference, we used two other rules (lines 234-236, 756-762): $R_{vocab}$, that cuts one non-ring single bond, to construct the vocabulary, and $R_{remask}$, that cuts all non-ring single bonds, to determine which fragments the remasking will operate on.

---

Data statistics should be provided.

We used the dataset of Noutahi et al., 2024, and will include the statistics in the revision.

---

We hope our response could address your questions and clarify any confusion. If our response is satisfactory, we would like to kindly ask you to consider raising your score. Otherwise, we will be happy to further discuss and update the paper.

We sincerely appreciate your comments that all our claims are supported by extensive evaluations, that our proposed fragment-based generation approach aligns well with industry needs, and that our paper demonstrates strong methodological innovation. We address your questions below.

---

Additional baselines, such as reinforcement learning-based optimization, could provide a more comprehensive validation. (Also related to Reviewer 9oFx’s comment: In de novo generation and fragment-based generation, only one baseline is included.)

We used the PMO benchmark (Gao et al., 2022) for the goal-directed hit generation task (Tables 3, 11-13), and compared our GenMol with a total of 28 baselines. These baselines span a broad class of molecular optimization methods, including reinforcement learning methods, genetic algorithm methods, and Bayesian optimization methods. The proposed GenMol achieves the best performance against all of these baselines and its ablated baselines (Table 5).

Furthermore, here we compare GenMol with DiGress (Vignac et al., 2023), a graph discrete diffusion method on the de novo and fragment-constrained generation tasks, where GenMol significantly outperforms DiGress in both quality and sampling efficiency.

Table: De novo generation results.

Table: Fragment-constrained generation results.

---

How molecular context guidance (MCG) is integrated into the diffusion process could be more explicitly explained. (Also related to Reviewer 9oFx’s comment: How is the molecular context defined in Lines 106-107?)

MCG is Autoguidance (Karras et al., 2024) specifically designed for the discrete diffusion of GenMol to fully utilize the given molecular context information. This is done by comparing two outputs from a single model, with good (i.e., less masked) and poor (i.e., more masked) input, respectively. Specifically, the good input $z_t$ is a given partially masked sequence, which is further masked by $100 \cdot \gamma$ % to yield poor input $\tilde{z}_t$, and the two resulting logits are compared and calibrated as Eq. (7). Intuitively, the logits with the poor input exaggerate errors in the logits with the intact input, and $w>1$ removes the errors from the good logits.

Following the use cases of context in LLM literature, molecular context in MCG denotes the given information in the form of the input sequence, i.e., given fragments in fragment-constrained generation and partially masked sequences during the fragment remasking in goal-directed generation.

We apologize for the brief description due to space limitations and will clarify this in a future revision.

---

What would be the cost of training this model on a broader space like the Enamine 1B dataset to fully explore drug space by leveraging this model's strengths?

The cost of training GenMol on a new SMILES-formatted dataset comes from two sources: (1) preprocessing SMILES -> SAFE, and (2) training with new SAFE strings. SMILES to SAFE conversion of 249,455 ZINC250k molecules takes ~6 minutes. For GenMol to go through 1B training molecules with the same setting in the paper would take ~60 hours.

---

Is there any possibility of conditioning this model to guide the generation towards a specific biological target?

GenMol can be easily extended to be conditioned on specific biochemical conditions by adopting any off-the-shelf conditioning methods. For example, protein pocket embeddings obtained from a pretrained protein model can be incorporated into GenMol via cross-attention to generate corresponding ligands [A, B].

---

References

[A] Fu et al., Fragment and geometry aware tokenization of molecules for structure-based drug design using language models, ICLR, 2025.

[B] Wang et al., Token-Mol 1.0: tokenized drug design with large language model, arXiv, 2024.

We sincerely thank you for your comments. We appreciate your positive comments that our paper introduces a versatile molecular generative model that can address various drug discovery tasks using a single backbone model, that all the claims are supported well by the experiments, and that the experiments are carefully designed. We address your concerns and questions below.

---

The paper can be improved by comparing the proposed guidance method to existing methods in the literature like Nisonoff et al. (2024). In MCG (Eq. (7)), how is the goal value $y$ involved?

We appreciate the suggestion. MCG and Discrete Guidance (Nisonoff et al., 2024) are not directly comparable: MCG does not rely on the target property conditions for guidance while Discrete Guidance does. To use Discrete Guidance, one has to train a conditional generative model that is conditioned on the target properties, which cannot be easily transferred to other tasks. Instead, MCG is specifically designed for the discrete diffusion of GenMol that is property-unconditional and general-purposed. For example, MCG can be used for fragment-constrained generation as well as goal-directed generation, while Discrete Guidance cannot. Therefore, the two methods have different objectives and can be used orthogonally.

The score $y$ (Eq. (5)) is used for fragment vocabulary selection which allows GenMol to effectively perform goal-directed generation without any target property-based diffusion guidance, and is not related to MCG.

---

$x^l_\theta(z_t, t)$ is already a probability distribution over the token vocabulary, and you could introduce the temperature in the softmax used in itself. Is there a specific reason for applying the softmax(log(x)) transformation again?

To avoid introducing additional notation, we used $\log(x^l_\theta(z_t, t))$ to denote the logits generated by the model, and in the actual implementation, we applied the temperature $\tau$ to the generated logits during inference (Eq. (4)). We will clarify this in the revised version.

---

Why does the diversity increase with $N$ (in Table 1)? Is the diversity increasing solely at the expense of validity?

A larger $N$ means the model predicts more tokens at once at each time step, which means the model is given fewer tokens to complete the sequence on average. The less conditional information results in more freedom in the generation, leading to an increase in diversity. On the other hand, it also puts more burden on the model capacity, which may lead to a drop in molecular validity (Table 1 and Figure 4).