Generalized Protein Pocket Generation with Prior-Informed Flow Matching

Designing ligand-binding proteins, such as enzymes and biosensors, is essential in bioengineering and protein biology. One critical step in this process involves designing protein pockets, the protein interface binding with the ligand. Current approaches to pocket generation often suffer from time-intensive physical computations or template-based methods, as well as compromised generation quality due to the overlooking of domain knowledge. To tackle these challenges, we propose PocketFlow, a generative model that incorporates protein-ligand interaction priors based on flow matching. During training, PocketFlow learns to model key types of protein-ligand interactions, such as hydrogen bonds. In the sampling, PocketFlow leverages multi-granularity guidance (overall binding affinity and interaction geometry constraints) to facilitate generating high-affinity and valid pockets. Extensive experiments show that PocketFlow outperforms baselines on multiple benchmarks, e.g., achieving an average improvement of 1.29 in Vina Score and 0.05 in scRMSD. Moreover, modeling interactions make PocketFlow a generalized generative model across multiple ligand modalities, including small molecules, peptides, and RNA.

Paper

References (94)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer bfsA8/10 · confidence 4/52024-07-06

Summary

In this paper, the authors proposed PocketFlow, a protein-ligand interaction prior-informed flow matching model for protein pocket generation. The flow matching for backbone frames, sidechain torsion angles, and residue/interaction types are appropriately defined. To enhance the structural validity and binding affinity of the generated pockets, the authors proposed to leverage affinity and geometry guidance in the sampling process. Extensive experiments show that PocketFlow provides a generalizable model for pocket generation covering various ligand modalities such as small molecules, nucleic acids, and peptides.

Strengths

1. PocketFlow leverages domain knowledge including protein-ligand interactions and geometric constraints to enhance pocket generation quality, which is quite novel. 2. PocketFlow extends the pocket generation task to broad ligand modalities such as small molecules, RNA, and peptides. Experiments show that the prior guidance can effectively improve the generalization capability. 3. In experiments, the advantage of PocketFlow over state-of-the-art baselines such as RFDiffuisionAA is obvious, achieving an average improvement of 1.29 in Vina score and 0.05 in scRMSD in Table 1. 4. The description of the PocketFlow algorithm is quite clear, and a preliminary version of code is provided for reproducibility. 5. The authors conducted comprehensive ablation studies to show the contribution of affinity/geometry guidance and protein-ligand interaction learning.

Weaknesses

1. The backbone model of PocketFlow is modified from existing work such as FrameDiff and is less novel. 2. In lines 252-253, the authors said PocketFlow is only pretrained on protein-small molecule datasets, i.e., CrossDocked and Binding MOAD. Is it possible to train PocketFlow on the combination of protein-small molecule/peptide/RNA datasets for better performance? 3. In experiments on protein-peptide and RNA, PocketFlow represents peptide/RNA ligands as molecules. Could the frame representation for the protein residues also applied to peptide/RNA ligands?

Questions

Please see the Weaknesses.

Rating

8

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

The limitations and broader impacts are well discussed in Section 5.5 of the main paper.

Reviewer bfsA2024-08-13

Response to Author

Hi, Thanks for the response, my concerns have been successfully addressed! Regards,

Reviewer e3xX7/10 · confidence 4/52024-07-08

Summary

This paper studies the task of generalized ligand-binding protein pocket generation. To tackle the challenges of existing works, the authors proposed PocketFlow, a generative model that incorporates protein-ligand interaction priors based on flow matching. PocketFlow explicitly learns the protein-ligand interactions during training and leverages multi-granularity guidance to generate high-quality pockets during sampling. Experiments show that PocketFlow is a generalized generative model across various ligand modalities, including small molecules, peptides, and RNA.

Strengths

1. The paper is well-written and easy to follow. The illustration of Figure 1 is attractive. The technical details and the code are provided for better reproducibility. 2. PocketFlow represents the first few works to broaden the scope of pocket generation tasks to include various ligand modalities, such as small molecules, nucleic acids, and peptides. 3. PocketFlow effectively combines the latest flow-matching models with prior knowledge (affinity guidance and interaction geometry guidance) to generate protein pockets with enhanced affinity and validity. 4. The proposed algorithms overall sound valid to me. The tasks and evaluation metrics of peptide/RNA-binding pocket design are also well formulated. 5. Experiments show that PocketFlow achieves an average improvement of 1.29 in Vina score and 0.05 in scRMSD. The authors also leverage PoseCheck to evaluate the protein-ligand interactions and steric clashes, which makes the evaluation comprehensive.

Weaknesses

1. What is the time cost of PocketFlow and baseline methods in generating protein pockets? Would the prior knowledge-based guidance bring an extra burden to pocket generation? 2. In line 222, the authors mentioned sidechain ensemble for the interaction geometry calculation, which is too concise. The authors are recommended to elaborate more on this in the main pages. 3. In Table 2, some metrics of PocketFlow are not the best and RFDiffusionAA has better results.

Questions

1. Why didn’t the authors apply discrete flow matching methods for residue/interaction type prediction, e.g., [1]? [1] Dirichlet flow matching with applications to dna sequence design, arXiv preprint arXiv:2402.05841, 2024

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors discussed the limitations and broader impacts in Section 5.5.

Reviewer gVL47/10 · confidence 4/52024-07-10

Summary

The paper explores methods for generating protein pockets given a ligand using a flow matching generative approach. Unlike previous methods, the proposed approach integrates additional constraints into the flow matching learning process to guide the search for relevant pockets. Two types of constraints are considered: 1. Affinity score prediction, which utilizes an oracle predictive model to guide the generation process. 2. Geometric constraints, which ensure that the distances between atoms involved in specific types of bonds remain below certain thresholds. The authors examine the effectiveness of these constraints, incorporating both domain knowledge and transferred knowledge from an affinity prediction model, in the generation of protein pockets. They benchmark their approach against state-of-the-art methods that do not use such constraints. The results demonstrate that the new pockets generated have improved RMSD and Vina scores.

Strengths

The concept of incorporating additional constraints into Flow Matching is interesting. While conditional flow matching exists, determining the appropriate constraints for pocket generation requires substantial domain knowledge and engineering effort to be effective. The results are promising, as the authors conducted numerous experiments and comparisons with state-of-the-art methods. It is evident that considerable time was invested in gathering results from these methods. The paper is well-written and enjoyable to read, even for those with limited knowledge of pocket generation problems.

Weaknesses

I have a concern about the inclusion of the affinity oracle predictor, which was trained on separate datasets. The authors did not specify which dataset was used to train their predictor or address the potential for information leakage between the training dataset and the test set used for pocket finding. Since the predictor provides the flow matching with a prior on which pocket might be relevant for a given ligand, any leakage would give a clear advantage. I looked for more detailed information about the affinity predictor in the appendix but found none regarding the training dataset. I also question the novelty of the work in terms of its technical contribution to the machine learning community. While conditional flow matching itself is not new, the novel aspect of this work is the incorporation of additional domain constraints into the framework. Therefore, it is unlikely to be of significant interest to the machine learning research community from a methodological development perspective. Additionally, since the authors retrained the baseline models with additional data, it is important that they report how hyperparameters were selected and tuned for these baselines.

Questions

1. Could you please address the potential leakage between the dataset used for training the affinity predictor and the datasets used for assessing the pocket generation? 2. Could you please clarify the impacts of HPO of the baselines and the proposed method in this work?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have addressed relevant limitations of the work in the paper.

Reviewer gNH76/10 · confidence 4/52024-07-23

Summary

The paper proposed PocketFlow, a generative model for designing protein pockets that bind with ligands. It aims to overcome limitations in existing methods by incorporating protein-ligand interaction priors and utilizing flow matching. PocketFlow is designed to handle multiple ligand modalities and demonstrates superior performance on various benchmarks.

Strengths

- The paper provides detailed methodology and includes anonymous code to reproduce the results. - The proposed method generalizes across various ligand modalities, including small molecules, peptides, and RNA. - The model outperforms existing methods on multiple benchmarks, demonstrating significant improvements in Vina scores and scRMSD. - The model explicitly models key protein-ligand interactions, enhancing binding stability and affinity.

Weaknesses

- The method only considers interactions between protein and ligand, potentially neglecting interactions between protein sidechains within the pocket region. - For some residue types, there might be π instead of 2π symmetry in the sidechain structures, which the proposed method seems to simplify. - The use of flow matching and multiple guidance mechanisms could result in higher computational costs compared to simpler models.

Questions

1. Why were these four types of interactions (hydrogen bond, salt bridge, hydrophobic, π-π stacking) chosen? Are there other interactions existing in protein-ligand binding that should be considered? 2. Can the proposed method be extended to protein-ligand docking (fixing the pocket type and structure)? 3. What is the efficiency of calculating different interaction types? 4. Is there any metric to evaluate the correctness of the interactions generated? 5. The original IPA relies on the rotational equivariance of frame orientation to achieve model’s invariance. However, the proposed method additionally treats the ligand atom as a residue and uses an invariant identity matrix to represent its orientation. Will the proposed IPA still output invariant embeddings? 6. How are the guidance coefficients for different guidance mechanisms determined? 7. Is RFDiffusionAA trained on the same dataset as the proposed model?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Authorsrebuttal2024-08-05

Further Response to Reviewer gNH7

**Comment 6**: What is the efficiency of calculating different interaction types? **Response 6**: Thanks for the question! In Figure 7 included in the Appendix, we compared the average generation time of PocketFlow and its variants such as PocketFlow without different guidance terms. We can observe that calculating the protein-ligand interaction for guidance will not bring much overhead (~28% of the total generation time). To detect the different interaction types in the generated pockets, we leverage PLIP and posecheck. They are also efficient tools and it totally takes around 20 seconds to process 100 generated pockets. **Comment 7**: Is there any metric to evaluate the correctness of the interactions generated? **Response 7**: Thanks for the question! In PocketFlow, we used the protein-ligand interaction profiler (PLIP) [69] to detect and annotate the protein-ligand interactions for each residue by analyzing their binding structure (Table 3 and lines 304-314 of the paper). Generally, PLIP is based on physical/chemical rules and employs four steps including structure preparation, functional characterization, rule-based matching, and filtering to detect the generated interactions. In the PLIP paper [69], the authors compared the detected/ground truth interactions of 30 literature-validated examples and achieved good consistency. Therefore, we can regard the detected interactions by PLIP as correct. **Comment 8**: The original IPA relies on the rotational equivariance of frame orientation to achieve model’s invariance. However, the proposed method additionally treats the ligand atom as a residue and uses an invariant identity matrix to represent its orientation. Will the proposed IPA still output invariant embeddings? **Response 8**: Thanks for the insightful question! We agree that IPA relies on the rotational equivariance of frame orientation to achieve the model’s invariance. In our case, there is no canonical orientation for the ligand atoms and we set them as an identity matrix for simplicity. To achieve invariant embeddings, we can initialize the protein scaffold by aligning the protein principal axes with the coordinate axes. This can be achieved by subtracting the center of mass (COM), computing the inertia matrix, diagonalizing the inertia matrix, and aligning the protein to principal axes. In experiments below, we compared such an initialization strategy with the default setting (only subtracting COM). The experimental results are comparable and show no significant differences. Therefore, we use the default setting for simplicity. We will include more discussions in our revised paper. | Model | CrossDocked | | | Binding MOAD | | | |---------------|-----------------------|-------------------|------------------|----------------------|-------------------|------------------| | | AAR (↑) | scRMSD (↓) | Vina (↓) | AAR (↑) | scRMSD (↓) | Vina (↓) | | PocketFlow | 52.19±1.34% | 0.67±0.04 | -8.236±0.16 | 54.30±1.70% | 0.68±0.03 | -9.370±0.24 | | w/ principal axes aligning | 52.12±1.29% | 0.67±0.03 | -8.227±0.18 | 54.47±1.73% | 0.69±0.03 | -9.362±0.28 | **Comment 9**: How are the guidance coefficients for different guidance mechanisms determined? **Response 9**: In the default setting, the guidance coefficients of PocketFlow including $\gamma, \xi_1, \xi_2,$ and $\xi_3$ are set as 1 and achieve good results. We also explore the influence of guidance coefficients in the appendix. For example, in Figure. 8 of the submitted paper, we explore the impact of Affinity Guidance Strength ($\gamma$) on various generation metrics. As $\gamma$ is scaled up, the Vina Score significantly improves and quickly stabilizes; AAR initially increases before gradually decreasing; scRMSD, on the other hand, increases with higher $\gamma$. These observations underscore the importance of selecting an appropriate $\gamma$ to effectively balance the guidance and unconditional terms. While Affinity Guidance promotes the generation of high-affinity pockets, an excessively high $\gamma$ can result in less valid pocket sequences or structures. In the default configuration, $\gamma$ is set to 1 for simplicity. **Comment 10**: Is RFDiffusionAA trained on the same dataset as the proposed model? **Response 10**: As indicated in lines 822-823, we used the provided checkpoints of RFDiffusionAA for all the experiments because the training code is not available (https://github.com/baker-laboratory/rf_diffusion_all_atom). The original training data of RFDiffusionAA contains 121,800 protein-small molecule structures, 112,546 protein-metal complexes, and 12,689 structures with covalently modified amino acids, which represent a broad set of protein-ligand complex structures.

Reviewer e3xX2024-08-09

Response to author

Thanks for the authors response. After reading the rebuttal, my concerns are resolved and I decided to rise my score.

Reviewer gVL42024-08-13

Thank you for your response to my concern, I read the paragraph in the appendix: "To train the binding affinity predictor, we first annotate the data points in the corresponding training set: data points are annotated 1 if their affinity is higher than the average score of the dataset, otherwise 0." Could you please clarify more on "higher than the average score of the dataset", the average score you mentioned in this sentence is calculated on the training data?

Authorsrebuttal2024-08-13

Thanks for the response!

Yes, the average score (Vina score) is calculated on the training data. We annotate the data point as 1 if its calculated Vina score is lower than the average (higher affinity); we annotate the data point as 0 if its calculated Vina score is higher than the average (lower affinity). Therefore, all the calculations are based on the training data and there are no data leakage risks. We will make the statements in the paper clearer in the revised version. Thanks for the comments! Bests, Authors

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC