Thank you very much for your thorough review and feedback! We appreciate it a lot. We are glad that you thought that applying opponent shaping to antibody design is "a good contribution that extends beyond traditional ML-based antibody optimisation"!

Answering Questions

1. Evaluating ADIOS on Other Viruses

We agree that testing ADIOS against a wider range of viruses would help further validate our claims. To address this, we have conducted additional shaping experiments with three more viral antigens: the flu, MERS, and the West Nile virus. For all three, our experiments show that the antibody shapers generated by ADIOS successfully shape the new viruses, limiting viral escape. Interestingly, the computational trade-offs of shaping horizons are dependent on the virus but follow the same main trends. We have added these new results to the paper.

2. Sensitivity of ADIOS to Hyperparameters

Based on our experiments ADIOS is not very sensitive to choices of different evolutionary parameters. As an example, we experimented with different mutation rates of the virus and found qualitatively similar results. To test the transferability of our results we also designed the external pressure experiment, the results of the experiment are in Figure 4. This result demonstrates the robustness of ADIOS to a domain shift. Even with an added pressure of an external antibody we can still see a shaping effect induced on the virus by the long horizon shapers. Additionally, the new results of the extra experiments on 3 other viruses show that ADIOS achieves shaping and limits viral escape for all the selected viruses, not only Dengue. Overall, this demonstrates that ADIOS is not sensitive to the choice of evolutionary parameters.

3. Incorporating Dynamics

Absolut! already allows us to incorporate an aspect of ‘dynamics’, as it tests binding over a larger number of potential poses, and the antibody structure is permitted to change between poses. The minimum energy poses change dramatically throughout the evolution of antibody/antigen, and we study the influence of shapers on binding poses in Appendix B.

That said, you are completely correct about the static structure of the antigen which is used to generate these binding poses in Absolut!. In future work, we plan to directly account for the changes in the antigen structure by using structure prediction models like AlphaFold3 (AF3), which can now accurately predict the structure of an antibody-antigen complex. We are building a binding simulator that combines the AF3 structure with a protein-protein docking score to achieve a more accurate binding affinity score. We plan to use this new binding simulator which accounts for antigen structure changes, instead of Absolut!, to further evaluate the ADIOS framework and demonstrate its real-world applicability.

4. Real-World Validation

We want to conduct a real-world evaluation of ADIOS in the new simulator mentioned above by using data from past COVID variants. All SARS-CoV-2 pandemic strain sequences are available through GISAID (https://gisaid.org/). Additionally, a dataset of COVID antibodies is also available through CoV-AbDab (https://doi.org/10.1093/bioinformatics/btaa739). We could retrospectively evaluate the ‘Myopic’ case by using these antibodies. I.e. we can show that by setting up our simulator to a state equivalent to the beginning of the pandemic and simulating viral escape as a response to some of these real-world antibodies, the real viral strains observed later in the pandemic are ‘in distribution’ of our simulator’s outputs. That would show the simulator is ‘trustworthy’, and successfully running ADIOS in it would indicate real-world applicability.

Going beyond simulation, we can validate ADIOS using bacteria phages, i.e. viruses that infect bacteria, and run shaping experiments in a wet lab with mutating bacteria and bacteria phages. Co-evolving bacteria phages (https://doi.org/10.1073/pnas.2104592118) would be a good real-world test ground for ADIOS, that is experimentally tractable and low-risk. ADIOS’s inner loop would correspond to bacteria phage evolving while the outer loop would be evolving the actual bacteria.

5. ADIOS for Cancer Therapy

Monoclonal antibodies (mAb) are a common therapy used for cancer treatment. The way to adapt ADIOS to cancer therapy would be to still optimise antibodies in the outer loop, this time these would correspond to mAb antibodies. Instead of viral escape simulation in the inner loop, we would simulate the evolution of cancer cell’s growth factor receptors. The goal would be to design antibodies that shape cancer cells into cells that don’t proliferate well. We have added a short phrase to our manuscript suggesting this as a potential domain of application.

Again, thank you for your detailed feedback! Hopefully, we have addressed all your concerns. If we did, could you consider updating your score for our paper?

Thank you for your review!

As we understand, you have three primary concerns:

1. Lack of discussion of antibody design methods
2. Unclear how Myopic Antibodies are generated
3. No comparison to other reinforcement learning algorithms

1. Discussion of Antibody Design Methods

You are completely correct, and we have updated the paper to include references to more antibody design methods. The reason we didn’t include such references originally is because ADIOS is somewhat agnostic to the choice of the antibody optimization algorithm used, the primary difference is that ADIOS emphasises the importance of accounting for the adaptation of the virus using simulated evolution. However, it is important to discuss the breadth of different antibody design methods in our related work section, which we have now rectified with additional references. In particular, we now reference energy-based antibody optimization methods (https://doi.org/10.1371/journal.pone.0105954, https://doi.org/10.1371/journal.pcbi.1006112), including optimization methods based on GFlowNets (https://doi.org/10.48550/arXiv.2411.13390), sequence-based language models (https://doi.org/10.1093/bioinformatics/btz895, https://doi.org/10.1038/s41598-021-85274-7) and structure-based approaches relying on GNNs (https://doi.org/10.48550/arXiv.2207.06616) and diffusion models (https://doi.org/10.48550/arXiv.2308.05027).

2. Generation of Myopic Antibodies

The Myopic Antibodies are generated in the exact same way as the other shaper antibodies, but with a horizon of 0 in the objective. We now recognise that our explanation of this was unclear, as we didn’t explicitly state that we also optimize myopic antibodies in the same way we optimize shapers. We have now changed line 233 from “To optimise antibody shapers, we …” to “To optimise both shapers and myopic antibodies, we …” to rectify this. Thank you for bringing this to our attention!

3. Comparison to RL Baselines

We don’t compare against several baselines because the core question of our current work is whether it is possible to apply opponent shaping to viruses and design effective antibody shapers, which we have now successfully demonstrated.

However, your suggestion is a natural follow-up question - given that we can shape viral evolution, what algorithms (PPO, Rainbow DQN, etc.) allow us to create the optimal antibody shapers in the outer loop? This becomes significantly more viable to test due to our enormous speed up of Absolut! (x10,000), making fast, full training rollouts on the GPU possible. Still, there are some challenges in adapting a few of the works you mentioned to our problem setting. In particular, a lot of opponent shaping literature, such as LOLA and AAA, rely on assumptions suitable for RL-trained/humanlike agents, but less suitable to biological adaptive agents, such as viruses, which do not perform value iteration but evolve instead.

Back-Testing on Historical Data

Thank you for this suggestion. We fully agree with you! We are already planning on doing the back-testing, but it requires a much higher fidelity simulation of both binding and viral escape. As a result, we are working on extensions to this work, where instead of using Absolut! as our binding simulator we use more complex models. These include AlphaFold3 (and similar) to predict the changing structure of antibodies and viruses, as well as protein-protein docking models to estimate the binding strength based on the structure. Using other predictive models like EVEscape (https://doi.org/10.1038/s41586-023-06617-0), we also account for additional factors that influence viral escape. We can retrospectively evaluate the ‘Myopic’ case by using historical antibodies. I.e. we can show that by setting up our simulator to a state equivalent to the beginning of the pandemic and simulating viral escape as a response to some of these real-world antibodies, the real viral strains observed later in the pandemic are ‘in distribution’ of our simulator’s outputs. That would show the simulator is ‘trustworthy’, and successfully running ADIOS in it would indicate real-world applicability.

Additional Experiments

To better evaluate ADIOS we have now also conducted extra experiments with 3 additional viruses: flu, MERS and the West Nile virus. In all of these cases, described in the updated manuscript, our experiments show that the antibody shapers generated by ADIOS can successfully shape the viruses. The results of these experiments are now added to the paper as well.

Conclusion

Hopefully, we have addressed all your concerns! If we did, would you consider raising your score for our paper? If not, please let us know any further questions or suggestions. We are very focused on making this paper as high quality as possible, so we really appreciate your feedback.

Thank you for your detailed review! We address everything possible within the character limit:

Accuracy + Real-World Utility of Absolut!

To address the real-world utility of our simulator we quote the original Absolut! paper: “Of note, Absolut! is neither suited nor designed to directly predict if and where an antibody sequence binds to an antigen in the real world. Rather, Absolut! has been developed with the premise that a successful ML antibody-binding strategy for experimental (real world) datasets should also perform well on synthetic datasets (and vice versa)”. Hence Absolut! is not designed for direct real-world predictions, but it has been shown that ML models which perform better according to Absolut! also perform well on real-world datasets. This makes Absolut! ideal as a testing ground for ADIOS. We reference this in Appendix Section C, where we explain Absolut! in more detail. To make this more clear, we can incorporate an additional explanation in the background section?

Notation and MDP Definition

We note that the functions $R_v$ and $F_v^H$ were not defined in the algorithms; we have therefore added their definitions. As for the formal definition of the MDP, we agree and have added a section detailing the MDP to the methods section.

Answering Questions

1. Simulating viral evolution is generally challenging and computationally involved. However, recent works, such as EVEscape (https://doi.org/10.1038/s41586-023-06617-0) have had success forecasting viral mutations, so it is possible to have some predictive power, which is likely to increase as models and data improve. In our viral escape simulation, the virus responds to two main factors, the antibody and the viral target. In more accurate simulators you might want to account for other factors too, for example in real life the evolution will depend on: how chronic the virus is, what is the percentage of people that will be infected, and the mortality. To keep our approach computationally feasible we omit some of these factors.
2. See above.
3. The key consideration in antibody design is the specificity of the antibody, i.e. does it bind to the correct antigen and nothing else (https://doi.org/10.1186/s12929-019-0592-z). In ADIOS we just consider a region of the antibody sequence which is responsible for the specificity, the CDRH3 region. Designing the rest of the antibody sequence will affect many developability parameters in the human body such as PK, half-life, and effector functions, which are also important for antibody design but engineering these is independent of the specificity. Typically, the CDRH3s can be grafted to antibody scaffolds with good profiles of other properties, so our approach is suitable for the specificity question and selective pressure it creates.
4. The development of antibody therapeutics in the US on average takes about 8 years until the drug receives an FDA approval (https://doi.org/10.4161/mabs.2.6.13603). Viral evolution is very different between viruses, for example, RNA viruses (e.g. HIV, SARS-Cov2) evolve faster than DNA viruses (e.g. Herpesvirus). In general, a higher mortality rate limits viral evolution and needs a faster immune response, while a higher R0 means more evolution opportunities for the virus given the higher number of infections.
5. See the section on Absolut! above.
6. In this work we mainly focus on viral escape potential and how easily targetable the virus is. However, in future work, with higher fidelity simulators we can model other factors such as mortality or co-morbidities.
7. In this work we assume a fixed sequence length for the antibody and the virus, we explain this in more detail in Section 4.1. An extension of this work where we allow insertions and deletions to the sequence is certainly possible, one way to adapt the action space to allow this would be to have a max sequence length for the insertions and an additional ‘None’ token for the deletions. Protein structure prediction models, such as AlphaFold3 (AF3), would be useful to validate the feasibility of structures resulting from the mutated sequences of antibodies and antigens. However, the action space would remain in the sequence space, and we could just include an extra ‘structure check’ of the sequence by passing it through AF3.
8. Yes, there can be multiple anti-targets in our model, but in our experiments, we stick to 1.
9. We don’t define the antibody’s policy function explicitly but we can easily add that if you think it’s helpful.
10. Figure 2b and Figure 2c show viral and antibody evolution plateauing. Eventually, an equilibrium must be reached, however, this may take an extremely long time. Do you think it would be interesting to run an ultra-long horizon experiment, to better understand the behaviour?

Conclusion

We appreciate the interest you have shown in our work! Is there anything we could change for you to raise your review to a “strong accept”? Also, any future work ideas would be great too!

Thank you for your review. Below we address your concerns and questions:

Extra Experiments

Following your feedback, we have now conducted extra experiments with 3 additional viruses: flu, MERS and the West Nile virus. In all of these cases, our experiments show that the antibody shapers generated by ADIOS successfully shape the viruses, limiting viral escape. We have now added these results to the paper, alongside correcting the enumeration of appendix figures.

Sampling Outside of Absolut!’s Distribution

Thank you for bringing up the important problem of a binding model's distribution, or in other words its domain of applicability. We have carefully considered this problem and, in fact, one of the main reasons we decided to use Absolut! as our binding simulator is to avoid sampling outside of the distribution.

As we explain in Section 3 of our paper, sequence-based ML models like Mason et al., 2021 (https://doi.org/10.1038/s41551-021-00699-9); Lim et al., 2022 (https://doi.org/10.1080/19420862.2022.2069075); Yan Huang et al., 2022 (https://doi.org/10.3389/fimmu.2022.1053617), as data-driven predictive models struggle to generalise beyond their training distribution and are therefore not suitable for our application which explores novel viral mutations.

However, Absolut! is a mechanistic model (not data-driven) that we use as an oracle for querying the binding strength of new sequences. As we write in Section 3, “For any antibody-antigen pair, Absolut! enumerates possible binding poses and computes their energy using the Miyazawa-Jernigan potential” This process doesn’t have a training distribution or a domain of applicability and is well suited to our application which requires mutating both the virus (antigen) and the antibody.

Real-World Applicability

We appreciate the challenges of binding prediction, and acknowledge that using Absolut! (or any other simulator) comes at a cost of real-world applicability. However, it has been shown that ML models that perform better on Absolut! perform well on real-world datasets too (https://doi.org/10.1038/s43588-022-00372-4). Besides, using more realistic models is extremely expensive and cheaper data-driven models are not suitable due to their limited domain, making Absolut! an ideal choice for the current work.

More broadly, breakthroughs in biology have come from building on previous computational works with initially limited real-world applicability. For example, the development of effective mRNA vaccines has been credited at least in part to RNA Folding algorithm. Original papers such as Nussinov and Jacobson, 1980 (https://doi.org/10.1073/pnas.77.11.6309) and Zucker and Stiegler, 1981 (https://doi.org/10.1093/nar/9.1.133) laid the foundations to more modern heuristics for RNA structure prediction and therefore therapeutics applications.

We are currently working on extensions to this work, where instead of using Absolut! as our binding simulator we use more complex models. These include AlphaFold3 (and similar), to predict the changing structure of antibodies and viruses, as well as protein-protein docking models to estimate the binding strength based on the structure. We also account for additional factors that influence viral escape. However, Absolut! is a necessary first step to prove that ADIOS is feasible and effective. Due to its speed, it is also a great testbed for developing methods which can then be scaled to more computationally expensive models.

Answers to Questions

1. In Figure 2d, we show the number of binding samples required to achieve a given antibody fitness for different shaping horizons. In Table 1, we report how long a single binding sample takes in our GPU-accelerated implementation: $2.1 ×10^{−4}s$. As the binding calculation is the main computational cost, a single run takes up to 4000s ~ 1h. For Figure 2d we chose to show binding samples on the x-axis (instead of runtime) because it provides a reference point for future work where one will want to use a more expensive (i.e. slower) binding simulator. Would it be helpful to discuss in the paper how binding calculation is the main computational cost? Or show a plot with the computational speed of our method (on our hardware) on the x-axis?
2. Our binding simulator, Absolut!, is not a pre-trained predictive model but a mechanistic one instead. Absolut! applies the same scoring computation to any possible binding pose to calculate the binding energy, a priori, there is no one pose where the binding calculation is systematically more “correct”. Thus, we hypothesise that mean correctness of the binding score should not systematically change as we mutate the sequences and/or change binding poses.

Conclusion

We value and respect your critique, but we believe it is in large part due to a misunderstanding of Absolut!. If you disagree, please let us know why! We are always looking to improve the quality of our work. Otherwise, we would appreciate a re-evaluation of your review score.