Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization

Summary

This paper uses a pre-trained conditional diffusion model for antibody design. This diffusion model is fine-tuned using a direct energy-based preference optimization method, focusing on optimizing residue-level energy preferences to enhance the generation of antibodies with desirable structures and high binding affinities. The authors also compared their method with SOTA baselines on 55 cases from RAbD benchmarks and showed good results.

Strengths

1. The proposed algorithm is novel and interesting. The authors combine diffusion models and DPO to solve the specific problem of antibody design. 2. The algorithm is clear. The authors both define the diffusion process and DPO formation with very clear definitions. Meanwhile, all the figures with protein structure are very clear and informative. 3. The results are convincing. AbDPO archives good results on CDR total energy compared to other methods. 4. The experiments part is very detailed. The authors conduct experiments on 55 of 60 antibodies on both their methods and the compared baselines.

Weaknesses

Lack of explanation of SE(3)-equivariant neural network. The author uses the diffusion model with such an equivariant neural network from Luo et al. Lacking such an explanation may hurt the understanding of the whole method. **Minor:** - There are some writings inconsistent. For example, equations under line 130 are not marked with numbers, but equations under line 138 have numbers. The ABDPO in everywhere this paper is written as \textsc{AbDPO}, but in line 242 is "ABDPO". In line 235, the first letter in pyRosetta should be capitalized since it is a proper noun. Checking these format issues will improve the consistency in the future published version. Reference: ---------------------------------- Luo et al: Shitong Luo, Yufeng Su, Xingang Peng, Sheng Wang, Jian Peng, and Jianzhu Ma. 2022. Antigen-Specific Antibody Design and Optimization with Diffusion-Based Generative Models

Questions

1. Why we need to use SE(3)-equivariant neural network? 2. What is the inference time for generating one sample or one batch?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are adequately addressed.

Summary

This paper proposes a new perspective for antibody design-- incorporating the energy factors aiming to minimize the overall energy of designed sequence and structure. It involves diffusion to maintain the sequence-structure co-design. Towards the variance of energy factors, the paper proposes the idea of gradient surgery to simplify it. The paper conducted extensive experiments, showing its performance towards other baselines. The paper is easy to follow.

Strengths

* The paper proposes an energy-based preference optimization approach to achieve better rationality and binding affinity, which is inspiring in the field of antibody design. * The paper decomposites the energy factors, simplifying the gradient calculation process. * The paper achieved significantly lower energy than other approaches, reflecting the effectiveness of the model design. * The paper clearly states why it chooses energy as the main evaluation metric with strong statistical evidence. * The paper provides a detailed sample-level comparison of existing baseline approaches, which is a solid solution and beneficial to successive works.

Weaknesses

* Although the authors stated why they chose energy as the main metric, the AAR is about 10% lower than dyMEAN. The authors did not provide a solid reason for what caused the result (Are all the lower results considered hacked or biased? Maybe a sample-level analysis would be more convincing.) * The derivation of the final loss for fine-tuning could be simplified. * The adaptation of RLHF into this approach should be further clarified.

Questions

* This paper relies heavily on energy function and sidechain packing methods in pyrosetta. Would it be helpful to directly train the diffusion model from scratch by optimizing the loss function using pyrosetta? * There are several other tasks except for CDR-H3 design, as the paper aims to optimize the energy of the complex, would it be helpful to test the method with other downstream tasks that may be closely related to the energy(refer to dyMEAN)?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

As stated in the article.

Summary

The paper proposes an approach for fine-tuning diffusion models for the design of antibodies. The core diffusion-based generative model comes from Luo et al. [36] and to my understanding there are no technical changes to it. The second component is direct preference based optimization, inspired by fine-tuning of large language models. This is the main technical contribution and builds entirely on the works by Rafailov et al [41] and Wallace et al [46]. The reward signal comes from binding free energy that is decomposed at the residue level. Different components include attraction and repulsion forces that can be linked to antibody function. To overcome possibly diverging gradients associated with different energy components, the authors propose to leverage gradient surgery from Yu et al [51] that essentially increases cosine similarity between gradients for different tasks/energy components. Empirical evaluation is focused on qualitative aspects and whether the approach is able to discover better binders than initially given complexes, measured using binding free energy.

Strengths

I think it is an interesting approach to merry molecular simulations and physics based energy calculations with diffusion and generative models. Especially given the small number of available crystal structures in the SAbDAB database. Adding energy-based signal via direct preference optimization is an interesting re-purposing of that method. Empirical evaluation goes beyond amino-acid recovery rate and RMSD metrics. The “success” at generating better binders quantified via improvement in binding free energy relative to the initial complex is an interesting metric.

Weaknesses

Table 1 indicates that the approach is able to design better binders, measured via binding free energy relative to the initial complex. However, this comes at the expense of increasing the number of hydrophobic residues which is typically associated with non-specific binding. This would in all likelihood be useless binders and from the perspective of function no better than baselines. It would be interesting to see the results of a baseline that “fine-tunes” relative to ddG score directly. In my understanding, the results in Table 1 are vanilla baselines. None of them (e.g., MEAN, dyMEAN, HERN) has for instance been used in combination with iterative improvement algorithm and some physics based simulator. How would this compare to fine-tuning relative to the directed preference optimization? It would be interesting to see the results relative to different physics-based simulators. I’m not sure how many different simulators were used to generate “fine-tuning” signal? Would it be possible to include additional metrics such as lDDT, TM, and some metrics characterizing the fit on angles?

Questions

Could you also list how many mutations from the original CDR sequence are the improved binders (Table 1)? Can you share the structural similarity scores between epitopes in the train and test folds? If you limit the fraction of hydrophobic residues (PHR) to the value for the original complex and accept only designs with lower score, how does Table 1 look? Can you list relative to how many test fold complexes each method achieves better binding free energy (e.g., MEAN wins on N, AbDPO wins on 20, etc.)? How does Table 1 look like if you apply ITA to the vanilla baselines?

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

NA

Summary

This paper applies direct preference optimization to antibody design. Specifically, it uses Rosetta binding energy to guide a pre-trained diffusion model to generate antibody CDR structures with low binding energy.

Strengths

* Optimizing antibody binding energy is an important problem. * The proposed gradient surgery procedure is technically interesting.

Weaknesses

* To evaluate binding energy using Rosetta, it is necessary to run side-chain packing and energy minimization to clean up the predicted structure. Therefore, it usually takes couple of minutes to evaluate binding energy using Rosetta for just one structure. In other words, it is computationally expensive to guide diffusion models using Rosetta. * Rosetta side-chain packing is stochastic and non-deterministic. Therefore, if we relax the structure generated by diffusion model multiple times, the calculated Rosetta energy will be very different, and the standard deviation can be very high (sometimes it can be twice or three times higher than the mean). In other words, it is very tricky to construct a preference dataset because you need to compare the binding energy distribution between two CDR sequences, and their standard deviation is very high. * Despite the high standard deviation, the reported binding energy in this paper does not have standard deviation. It seems like the authors only calculate Rosetta binding energy once for each structure instead of running side-chain packing or energy minimization multiple times and take the average. Therefore, the reported results may not be statistically significant.

Questions

* Can you report your model performance by running Rosetta relaxation multiple times with different random seeds and report standard deviation of the reported binding energy for each method?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Response to the rebuttal

I’m satisfied with the response and will be increasing my score. Well done!

Thank you

Thank you for your response. I think the general methodology is applicable to properties other than rosetta. The new experiment shows the standard deviation and I encourage the authors to put them into the final paper. I have increased my score to 5.