TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models

Summary

This paper presents TurboHopp, an accelerated pocket-conditioned 3D scaffold hopping model designed to enhance the efficiency and speed of drug discovery. It addresses the slow processing speeds of 3D-SBDD generative models by offering up to 30 times faster inference speed while maintaining or improving on key metrics like drug-likeness, synthesizability, connectivity, and binding affinity. Additionally, it incorporates reinforcement learning to further optimize molecule designs, demonstrating its potential in various drug discovery scenarios.

Strengths

Accelerated Generation: TurboHopp's inference speed is 5-30 times faster than that of DDPM-based models, greatly improving the efficiency of drug discovery. Combination with Reinforcement Learning: By leveraging the fast inference speed of consistency models, TurboHopp applies reinforcement learning to 3D-SBDD-DMs, enabling fine-tuning of generative models based on specific objectives, such as improving binding affinity or reducing steric clashes, for more refined molecule design.

Weaknesses

The consistency model and reinforcement learning are well-established techniques, each with a robust body of research. The integration of reinforcement learning into diffusion models has been explored in the literature. In this paper, the authors concatenate these two methodologies for the purpose of scaffold hopping, without explicitly detailing any novel strategies or innovations in their implementations. The empirical comparison can be done in a more thoroughly by comparing with other latest state-of-the-art drug discovery algorithms.

Questions

1. Could the authors provide a more detailed analysis of the trade-off between inference speed and the quality of the generated molecules? 2. How does the novelty exceed 1 at Table 1? 3. The part that reduces time is mainly due to the consistency model. How did the author transfer the consistency model? What improvements have been made after the transfer? 4. What are the advantages of this model over SBDD models, such as Targetdiff [1] and Decomdiff [2]? 5. In terms of performance comparison, the enhancement provided by reinforcement learning to the model is significant. Can it be understood that surpassing DiffHopp primarily relies on reinforcement learning? Is it reasonable to use evaluation metrics directly for optimization? Can you conduct an ablation study focused solely on the integration of reinforcement learning? [1] Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. arXiv, 2023. [2] Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, and Quanquan Gu. Decompdiff: diffusion models with decomposed priors for structure-based drug design. arXiv, 2024.

Rating

3

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors state that they have improved generation efficiency, but they do not compare their approach with methods that generate molecules from scratch.

Summary

This paper proposed a pocket-conditioned 3D molecular scaffold hopping model based on the well-established consistency models. The framework is superior in terms of inference speed. Besides, the authors also proposed a corresponding RL method to fine-tune the model towards generating molecules with desirable properties. The experimental results show the effectiveness of the proposed approach compared with DiffHopp.

Strengths

- This work first introduced consistency models to molecular scaffold hopping and achieved promising results. - The evaluation was done from various perspectives, including connectivity, QED, SA, Vina, etc. - Introducing RL for optimizing the generated molecules towards desired properties is useful in practice.

Weaknesses

- More baselines are needed. There are also some other methods for scaffold hopping, such as [1,2,3], etc. Comparison with these methods is necessary to show the significance of the proposed method in the practice of drug discovery. - Some related works are missing. For example, this work introduced Reinforcement Learning for Consistency Models to improve the properties of generated molecules. There are related works in the field of diffusion models for molecular science that utilize the similar idea. For example, [4] use a RL method (e.g., actor-critic) to fine-tune the diffusion model to generate molecules with higher binding affinity, and [5] also uses RL-like methods to improve the quality of sampled docking poses in the protein-ligand docking task. - Lack of some ablation studies. The effectiveness of some proposed modules is not clear, e.g., the metric-based sampling methods. - Though many evaluation metrics are utilized in this work. some are still needed. For example, the geometric properties (e.g., bond lengths, bond angles, and torsion angles) are needed to be checked. References: [1] Hu, Chao, Song Li, Chenxing Yang, Jun Chen, Yi Xiong, Guisheng Fan, Hao Liu, and Liang Hong. "ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks." Journal of Cheminformatics 15, no. 1 (2023): 91. [2] Yu, Yang, Tingyang Xu, Jiawen Li, Yaping Qiu, Yu Rong, Zhen Gong, Xuemin Cheng et al. "A novel scalarized scaffold hopping algorithm with graph-based variational autoencoder for discovery of JAK1 inhibitors." ACS omega 6, no. 35 (2021): 22945-22954. [3] Zhou, Xiangxin, Xiwei Cheng, Yuwei Yang, Yu Bao, Liang Wang, and Quanquan Gu. "DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization." arXiv preprint arXiv:2403.13829 (2024). [4] Zhou, Xiangxin, Liang Wang, and Yichi Zhou. "Stabilizing Policy Gradients for Stochastic Differential Equations via Consistency with Perturbation Process." arXiv preprint arXiv:2403.04154 (2024). [4] Corso, Gabriele, Arthur Deng, Benjamin Fry, Nicholas Polizzi, Regina Barzilay, and Tommi Jaakkola. "Deep confident steps to new pockets: Strategies for docking generalization." ArXiv (2024).

Questions

See the weaknesses.

Rating

4

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

The author have discussed the limitations.

Summary

Given a protein pocket and a reference ligand, the authors suggest a method to generate different scaffolds to be able to eventually come up with new ligands with similar or even improved properties. Precisely, the authors learn a consistency function which maps noise to a scaffold (created as a 3D-conformation) while being aware of the protein pocket and the functional groups of the reference ligand. This created scaffold together with the functional groups build a new potential ligand. By combining the consistency-based model with RL, the generation process can be biased towards scaffolds which build together with the functional groups ligands with optimized properties, e.g. binding affinity or protein steric clashes. The proposed method is compared with a baseline. Also the impact of adding RL to the approach is evaluated.

Strengths

**Originality**: - **(S-O)**: As far as I know, scaffold hopping with consistency models and combining them with goal-directed RL has not been done before. Therefore applying consistency models to scaffold hopping and combining them with RL to optimize chemical properties is novel. **Quality**: - **(S-Q1)**: In terms of clarity and writing style the quality is very high (see clarity section) - **(S-Q2)**: The results section shows that the proposed method is promising and helpful for scaffold hopping. **Clarity**: - **(S-C1)** The paper is written very well. This together with a good paper structure, a good introduction and nice figures achieves that key take aways get very clear. - **(S-C2)** The introduction and the related work section set the stage very well for the proposed method. The authors give a good overview about recent work along with their strengths and weaknesses. Figure 1 highlights nicely the idea of the proposed method and Figure 2 shows its effectiveness. - **(S-C3)** The generation approach is described very well. Also Figure 3 is good and helpful. The auhors clearly describe which components of their pipeline are learnable. **Significance**: - **(S-S1)**: The results are significant because the suggested approach outperforms the compared baseline. - **(S-S2)**: Error bars are reported.

Weaknesses

**Quality and Clarity**: - **(W-QC)**: The mathematical notation / the formulas sometimes seem a bit cluttered and inconsistent: * In formula (8), $f_\theta^{n+1, x}$ is used for $f_\theta(Z_{n+1}, t_{n+1}|u)$ without having mentioned that one is the shorthand for the other. * Equation (5) indicates that $F$ has two outputs. The way this equation is written is rather code-style notation than a well defined mathematical expression since $x^\prime_t, h^\prime_t$ is from a mathematical viewpoint not well defined. This is why this equation style should be avoided. * $\sigma_\text{data}$ is not introduced but used in (7). **Significance**: - **(W-S)**: Because of (Q) the relevance for the RL-based scaffold generation might be limited for real-world scenarios.

Questions

- **(Q)**: For SMILES based goal-directed optimization [1] show that there is a risk that the generator learns to find blind spots in the reward function and rather learns to trick the scoring function than to optimize real-world properties. Do the authors think this might be an issue also for their RL-based approach? [1] Renz, Philipp, et al. "On failure modes in molecule generation and optimization." Drug Discovery Today: Technologies 32 (2019): 55-63.

Rating

7

Confidence

2

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors describe potential areas for improvement in the conclusion section.

Additional results including TurboHoppRL-50_metric evaluated on CrossDocked

For all reviewers, thank you for your time and effort for having interest in our work. We additionally add the results of TurboHoppRL-50 trained on CrossDocked. Through RLCM(RL for consistency models) training, we were able to increase all metrics (Validity, Connectivity, Diversity, Novelty, QED, Docking Scores) except synthesizability exceeding those of the reference set especially for docking scores. Please refer to the novelty/diversity of generated molecules regarding model overfitting. | Method | Validity (↑) | Connectivity (↑) | Diversity (↑) | Novelty (↑) | QED (↑) | SA (↑) | QVina (↓) | Time | |--------------------------|--------------|------------------|---------------|-------------|---------|-------|-----------|---------| | TargetDiff_inpainting | 0.927 | 0.826 | 0.841 | 0.914 | 0.424 | 0.661 | -5.896 | 740.33 | | DecompDiff_inpainting | 0.876 | 0.722 | 0.856 | 0.895 | 0.420 | 0.648 | -6.225 | 1263.72 | | DecompOpt_inpainting | - | - | - | - | 0.490 | 0.710 | -7.280 | - | | TurboHopp-100 | 0.990 | 0.853 | 0.484 | 0.936 | 0.488 | 0.702 | -7.051 | 6.17 | | TurboHopp-100_metric | 0.993 | 0.906 | 0.486 | 0.935 | 0.502 | 0.710 | -7.204 | 8.18 | | TurboHopp-50RL_metric | **0.997** | **0.951** | **0.800** | **0.952** | **0.524** | 0.674 | **-8.798** | **3.51** | | CrossDocked Test | 1.000 | - | 1.000 | 0.599 | 0.476 | 0.727 | -7.510 | - | If you have any issues to discuss about please leave any comments!