Navigating Chemical Space with Latent Flows

Recent progress of deep generative models in the vision and language domain has stimulated significant interest in more structured data generation such as molecules. However, beyond generating new random molecules, efficient exploration and a comprehensive understanding of the vast chemical space are of great importance to molecular science and applications in drug design and materials discovery. In this paper, we propose a new framework, ChemFlow, to traverse chemical space through navigating the latent space learned by molecule generative models through flows. We introduce a dynamical system perspective that formulates the problem as learning a vector field that transports the mass of the molecular distribution to the region with desired molecular properties or structure diversity. Under this framework, we unify previous approaches on molecule latent space traversal and optimization and propose alternative competing methods incorporating different physical priors. We validate the efficacy of ChemFlow on molecule manipulation and single- and multi-objective molecule optimization tasks under both supervised and unsupervised molecular discovery settings. Codes and demos are publicly available on GitHub at https://github.com/garywei944/ChemFlow.

Paper

Similar papers

Peer review

Reviewer D6r56/10 · confidence 3/52024-07-03

Summary

The authors built a general latent flow-based framework unifies traversal and optimization in the molecular latent space. The flow is trained by utilizing energy functions so that the vector field aligns with the gradients, with regularization imposed by an auxiliary classifier that tries to differentiate each distinct flow. Under multiple evaluation settings, ChemFlow outperforms or is generally on par with previous SOTA.

Strengths

- This paper is generally well-written and easy to follow. - It is a novel contribution to formulate the manipulation and optimization of molecules in latent space as learning the vector fields toward optimal distribution.

Weaknesses

- Typo: L747 "we first verify if the learned variational poster also follows a Gaussian distribution and we find that it does learn so", poster -> posterior - For Table 1, it risks not fully revealing the optimization ability if only TOP3 results are reported. Please consider adding mean and median values, too.

Questions

- In Table 2 and 3, why did ChemFlow perform best under mild similarity constraint, while suboptimal when the similarity threshold gets higher than 0.4? I think the similarity constraint is important, since usually in practical drug design people would expect to develop new therapeutics based on some drugs whose effects are already known, and the newly designed drug molecules are preferred to be similar so as to keep the effect. - For Figure 4, why is the predictor deviating so much from ground truth? LogP, QED and SA don't seem very hard to learn as far as I know. - Please consider elaborating on why "the learned variational poster[ior] also follows a Gaussian distribution" and "a strong correlation between almost all molecular properties and their latent norms" would contribute to the observed result that "a random latent vector taking a random direction will change the molecular property smoothly and monotonically".

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

This paper does not include a discussion on efficiency as compared with ChemSpace.

Reviewer BiGQ6/10 · confidence 2/52024-07-09

Summary

Designing new functional molecules within the vast chemical space is challenging, which necessitates efficient exploration and understanding of this space. The paper introduces a new framework called ChemFlow, which leverages latent space learned by molecule generative models and navigates it using flows. ChemFlow formulates the problem as a vector field that guides the molecular distribution to regions with desired properties or structure diversity. The paper conducts extensive empirical studies and justifies the effectiveness of the proposed method.

Strengths

- The proposed ChemFlow unifies the previous approaches via the vector field, which is novel and effective for learning a latent space with rich nonlinearity information. This can benefit various downstream tasks, including drug-related properties and protein-ligand binding. - Extensive experiments have been conducted to provide a good insight into the components of the proposed method. ChemFlow achieves faster empirical convergence and higher success rates, especially using Langevin dynamics. - The paper is generally well-written, with clear illustrations and tables.

Weaknesses

- Despite the novelty of the proposed method, the ChemFlow mainly focuses on small molecules. This might hinder the border impact of the learned latent space for macromolecular tasks like protein. - ChemFlow employs multiple approaches to learning different latent flows. However, the experiment's results show that different methods have different specialties, and the paper does not discuss the connection between flow learning and downstream tasks. - The paper mentioned the out-of-distribution generation problem in Appendix D.7 and Sec. 4.2. ChemFlow has encountered such a problem in an unsupervised manner. This could hinder the utility of the learned latent space for scenarios with distribution shifts.

Questions

- Is it possible to visualize the vector field using a tool like t-SNE or UMAP to provide further insight into the entanglement of the molecular properties? - Could you discuss the connection between flow learning and downstream tasks? - Could you discuss the extension of ChemFlow to macromolecular tasks like protein?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

The paper could provide further insights into the connection between flow learning and downstream tasks. Additionally, the author can consider discussing the extension of ChemFlow to macromolecular tasks like protein, which could make the proposed framework have a broad impact.

Reviewer piJE6/10 · confidence 3/52024-07-12

Summary

The authors propose a new method called ChemFlow, which navigates molecular distributions in chemical space through flow.

Strengths

1. The method demonstrates high generality, applicable to various molecular optimization tasks. 2. Based on the experimental results presented, the method shows significant optimization of molecular properties.

Weaknesses

1. The description of the experimental section lacks detail, such as which software was used to measure the docking scores? 2. Table 1 displays properties of several indicators, but showcasing only the top 3 among numerous sampled molecules may lack sufficient persuasiveness. It would be better to include a broader distribution, such as the mean and so on.

Questions

1. Why were only two specific targets evaluated in the docking score experiment? 2. I would like to get some intuitive understanding : What problems arise if training a molecular property predictor in latent space and updating it directly based on its gradient? How is this issue typically addressed in the proposed method?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

This work is a preliminary study.

Authorsrebuttal2024-08-07

**C4: Table 1 displays properties of several indicators, but showcasing only the top 3 among numerous sampled molecules may lack sufficient persuasiveness. It would be better to include a broader distribution, such as the mean and so on.** Thank you for pointing it out. We have included a new table with mean, median, and standard deviation in the general response.

Reviewer kzEx7/10 · confidence 3/52024-07-13

Summary

This paper presents a novel gradient flow-based method to traverse the latent space of molecular generation models, known as ChemFlow. The authors instantiate their framework with a number of different flows inspired by dynamical systems. They also investigate the use of supervised and unsupervised guidance for the flow methods with the goal of optimising molecular properties of interest. The authors perform many experiments focused on molecular optimisation in order to evaluate their method. Specifically, they investigate unconstrained optimisation, similarity-constrained optimisation and multi-property optimisation.

Strengths

- The authors present a novel formulation for exploring chemical latent spaces and optimising molecular properties which is an important and relevant task within the pharmaceutical and molecular design domains. They present a few different versions of their method, including an implementation which can use a surrogate property prediction model to guide the flow to optimised chemical space, as well as an unsupervised implementation which aims to maximise structural changes to the molecule. - The authors perform an extensive evaluation on molecule optimisation related tasks, comparing a number of different instantiations of their framework. They also benchmark against a previously introduced model for latent space traversal and a random traversal strategy. - ChemFlow methods show very promising results in comparison to baselines, especially when only looking at top performing molecules or applying similarity-constrained optimisation. - The authors also present a useful analysis of the latent changes under the random traversal strategy and an explanation for why random traversal can work reasonably well. - The paper is mostly very well written and, the evaluations in particular, are very clearly presented and easy to follow.

Weaknesses

- I find the methodology section on its own quite unclear since it's not clear how to actually use the objective functions that are outlined. Appendix sections D4 and D5 are helpful but ideally it would be possible to follow the main text on its own. Particularly, I think the text would benefit from showing the full loss function in the methodology and including a short outline of the training and sampling procedures and referring to the appendix. - The baselines for some tasks are a bit weak, particularly for the unconstrained molecular optimisation. For this task techniques like evolutionary algorithms and reinforcement learning fine-tuning have been proposed and widely used before. It would be very useful to see a comparison of ChemFlow with methods like these, as well as an evaluation of the training and sampling time for each.

Questions

- For the unsupervised guidance, when you match flows with properties, what the correlations are computed between? Does this require you to have an existing dataset of molecule-property pairs? - For the unsupervised cases did you experiment with different values of k? Which values of k were used? It seems to me that k might need to be very large in general in order to find a flow which matches with an arbitrary property. Other suggestions: I assume equation 8 should have $\phi^k$ instead of $\phi$? Additionally, the ordering of t and z in $\phi(.,.)$ or $\phi^k(.,.)$ is inconsistent - compare figure 1, the caption for figure 1 and equation 11 with all the other places were $\phi$ is used.

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

- While the authors evaluate their approach on optimising two properties simultaneously it's unclear how well this would work with a larger number, since their approach relies on simply adding the guidance terms together. - In this study the authors used a VAE with a fixed input size. It's unclear how easy it would be to apply a similar approach to a model with an arbitrary number of elements in an input or latent sequence, such as chemical language models or many diffusion-based models, which are much more commonly used in practice for molecular generation. - As far as I can tell the method as it stands requires a latent space optimisation to be done for every sample. This could lead to much longer sampling times than other methods such as RL fine-tuning which allows samples to be generated as normal but from an optimised model.

Authorsrebuttal2024-08-07

**C7: I find the methodology section on its own quite unclear since it's not clear how to actually use the objective functions that are outlined. Appendix sections D4 and D5 are helpful but ideally it would be possible to follow the main text on its own. Particularly, I think the text would benefit from showing the full loss function in the methodology and including a short outline of the training and sampling procedures and referring to the appendix.** A: Thanks for the suggestion. We will add the training objective $\mathcal{L} = \mathcal{L_r} + \mathcal{L_\phi} + \mathcal{L_\mathcal{P}}$ for supervised scenario and $\mathcal{L} = \mathcal{L_r} + \mathcal{L_\phi} + \mathcal{L_\mathcal{J}} + \mathcal{L_k}$ for unsupervised scenario to Sec 3.1 in the revised manuscript. We will also briefly discuss them and link them to the pseudocodes in the appendix. **C8: Inconsistency in Section 3 and Figure 1** A: Thanks for pointing out the mistakes. We fixed the notation in Section 3 and Figure 1 to make everything consistent with $\phi^k$.

Authorsrebuttal2024-08-07

**C4: Despite the novelty of the proposed method, the ChemFlow mainly focuses on small molecules. This might hinder the border impact of the learned latent space for macromolecular tasks like protein. Could you discuss the extension of ChemFlow to macromolecular tasks like protein?** Thank you for the suggestion. We have included a discussion of the broader application of our approach in the general response.

Authorsrebuttal2024-08-07

**C6: For Table 1, it risks not fully revealing the optimization ability if only TOP3 results are reported. Please consider adding mean and median values, too.** Thank you for the suggestion. We have included a new table with mean, median, and standard deviation in the general response.

Reviewer D6r52024-08-13

Thank you for the response. My concerns are addressed and I've raised my score to 6.

Authorsrebuttal2024-08-09

Thanks for the reply

We thank the reviewer again for the time spent and we are glad our revisions addressed your concerns. If the reviewer has any further questions or concerns, please don't hesitate to let us know!

Reviewer kzEx2024-08-13

Thank you for your thorough response and for conducting extra experiments. Most of my concerns have been addressed, however, if my understanding of the method is correct, I think the following two limitations still remain: 1. I agree that RL fine-tuning methods (eg. REINVENT as a representative example) don't optimise within the latent space but they are attempting to solve the same problem as ChemFlow - sampling molecules which optimise some scoring function. Of course these methods have their own strengths and weaknesses compared to latent space methods but I still think a performance (and possibly evaluation time) comparison would be beneficial here. 2. The study doesn't address many scenarios that are likely to encountered in practice for molecular design, such as optimising many properties simultaneously and using larger generative models such as chemical language models. I believe VAEs are not really much in practice for molecular generation because single-step generation is too weak. I would still like to thank the authors for the very interesting ideas presented and I am happy to increase my score to 7.

Authorsrebuttal2024-08-13

Thank you for your comment

We appreciate again the reviewers' efforts in providing useful comments that greatly helped us improve the manuscript. Given the limited rebuttal period, we cannot finish the additional experiments, but we will add them to the camera-ready version.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC