DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking

Summary

This paper proposed a model named DeltaDock for molecular docking. DeltaDock integrates protein pocket prediction and protein-ligand binding refinement. The pocket prediction is modeled as a pocket-ligand alignment problem with candidates of pockets given by other pocket prediction methods. The docking is formulated as a refinement problem given the initial docking conformation. Comparison with computational chemistry docking tools and learning-based methods show the advantages of DeltaDock. DeltaDock runs much faster than most of the baselines.

Strengths

1. The model is faster than the diffusion-based method DiffDock. 2. The experimental results show that DeltaDock has a good performance in blind docking.

Weaknesses

1. The testing set is not clearly described. Training and test proteins should have a maximum sequence identity of 40% to test the generalization. 2. in figure2, the compared methods are not consistent. For example, vina is in Figure 2 (a) but not in Figure 2 (b). This may be sometimes misleading. 3. in Line 60, "a GPU-accelerated pose sampling algorithm generating high-quality initial structure" is just DSDP, which limits the novelty of the paper. 4. what if you combine the pocket prediction of DeltaDock and autodock vina? Whether the improvement is due to the pocket prediction or the refinement is not clear.

Questions

what is the distance threshold used for graph construction?

Rating

3

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes

Summary

This manuscript introduces DeltaDock, a novel two-stage framework for molecular docking. Similar to previous works that use geometric deep learning methods, this work also takes the geometric dl network to do the modeling and the prediction is based on a regression problem. The main contribution is a contrastive based learning method for pocket prediction and also a bi-level update of the ligand pose prediction. The experiments are conducted on common datasets and also many studies are performed.

Strengths

1. The overall framework makes sense and is valid and useful for blind molecular docking. 2. The proposed contrastive learning approach for pocket prediction and the bi-level refinement module are novel contributions to the field. CPLA, in particular, reframes pocket prediction as a selection problem from a candidate set. 3. Lots of small tricks designed in the framework are beneficial to hold the correct structure and also help the final prediction. The detailed ablation studies, generalization analysis, and assessment of pose validity highlight the contribution of each component. 4. The experiments results are good compared to previous works.

Weaknesses

1. Though the process and the framework are reasonable, the overview of the work is quite similar to FABind, which implements a pocket prediction and a docking prediction molecules, The authors should put more comparison and describe more differences between these two works. 2. As the prediction module takes other external methods/tools, this framework is not a fully end-to-end one, though this is not a thing that we would purse must, I would like to hear my thoughts on this point. Why not do in an end-to-end way? The external tools indeed cost more and slow down the docking process. 3. The manuscript focuses on rigid protein docking, neglecting the inherent flexibility of protein side chains. Can the method get rid of the assumption that the protein side is rigid, modeling the side chain flexibility like previous works? [1, 2] [1] Qiao, Zhuoran, et al. "State-specific protein–ligand complex structure prediction with a multiscale deep generative model." Nature Machine Intelligence 6.2 (2024): 195-208. [2] Plainer, Michael, et al. "Diffdock-pocket: Diffusion for pocket-level docking with sidechain flexibility." (2023).

Questions

Please refer to Weakness Section

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Please refer to Weakness Section

Summary

A new method for molecular docking based on neural networks, called DeltaDock, is introduced. DeltaDock uses a two-step procedure. The first step is finding the binding pocket for a given ligand, which is implemented by aligning the molecular structure with the pocket embedding. The alignment is conducted with the use of contrastive learning. The second step is the placement of the ligand in the binding pocket, which is defined as a regression task on atom positions. This step includes predicting interactions on a coarse and fine level of protein representation. A fast pose correction is implemented using torsion alignment and SMINA-based energy minimization. The experimental section shows that the proposed method outperforms other state-of-the-art methods in terms of the RMSD of generated poses. Moreover, these structures are more realistic than those produced by other deep learning approaches in terms of the PoseBuster filters.

Strengths

Originality: - The idea of pocket alignment using contrastive learning is interesting and novel in this context. This approach is usually used to predict drug-target interactions. It is interesting to see it applied to find binding pockets based on the ligand input. - The torsion alignment proposed in this paper is a new and effective method of correcting conformations produced by regression models. Quality: - The experimental section contains results on two data splits and compares both classical and neural-network-based molecular docking models, which places the DeltaDock among the state-of-the-art methods. - The average time of inference is provided for each method, which demonstrates the difference in execution time between DeltaDock and other methods. Clarity: - Overall, the paper is written clearly and is easy to follow. Significance: - The experiments show the proposed method's strong performance, outperforming other approaches in terms of pose quality and alignment with the ground-truth pose. - The full docking protocol includes some constraints that ensure there are no clashes or torsional strain. This is important due to the recent criticism of deep-learning docking methods. - At the same time, DeltaDock is faster than diffusion methods like DiffDock, which can further accelerate virtual screening.

Weaknesses

Originality: - The Bi-EGMN is based on a few architectural decisions that may have influenced the model's strong performance. For example, in Equations 7 and 8, the messages are weighted by the interatomic distances. In Equations 9 and 10, skip connections are used. Have you compared this architecture with one that uses other message weights or does not use skip connections? It would be interesting to see it in an ablation study. - FABind+ should also be described in the related work as it follows a similar approach, where first the binding pocket is predicted, and then the molecule is docked inside this binding pocket. Interestingly, the size of the binding pocket is also predicted in this improved version of FABind, which speaks to the claim that current GDL methods have “difficulties in handling large binding pockets.” Also, FABind+ shows very promising results in their paper. Now, the code of this method has been published, so it would be interesting to see this model in the comparison if possible in such a short discussion period. Quality: - I am wondering why the average (or median) RMSD is not provided in Table 1. Other papers, including FABind and EquiBind also provided percentiles. Clarity: - Figure 1b could be improved to better depict the coarse level and atom level pocket representation. I am unsure if the model block shown in between refinement steps helps in the comprehension of this figure. - The sentence in line 31 is imprecise. These methods use binding pockets in some sense because they encode the whole protein and use it to condition the generative or predictive model.

Questions

1. Is it possible to generate diverse poses with DeltaDock, e.g. by sampling different initial conformations? How diverse can these poses be, given that the networks are E(3)-equivariant? 2. Do you have any mechanism to estimate the predicted pose confidence? For example, some docking models can also provide affinity prediction or likelihood of the predicted poses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

4

Limitations

The limitations have been described.

Discussion

Dear Reviewers, Please kindly review the authors' response with care, engage in discussions to clarify any uncertainties, and, if necessary, revise your comments accordingly. Thanks, AC

Follow-up Response

I appreciate the authors' hard work in the rebuttal. It has addressed all my concerns. Although the improvements are somewhat trivial, the authors have successfully demonstrated that this is a solid pipeline. I will raise my score.