acceptodds
Under review as a conference paper at ICLR 2027

Shaped, not Stirred: Co-evolving Combination Therapies to Prevent Viral Escape

Abstract

Combination therapies are the gold standard for treating rapidly evolving viruses like HIV or Influenza, where single-drug approaches quickly lose efficacy due to viral adaptation. However, current therapy design is fundamentally *myopic*, and optimises for immediate efficacy against present viral strains without anticipating how therapeutic pressure will influence future viral evolution. We introduce **Co-ADIOS** (**Co**operative **A**ntibody **D**esign v**i**a **O**pponent **S**haping), a framework that designs antibody combination therapies, or *cocktails*, that steer viral evolution toward weaker variants. We formulate the interaction between antibody cocktails and the virus as a *Stackelberg Evolutionary Game*, where antibodies act as *leaders* that shape the fitness landscape to steer the adaptive trajectory of the virus (the *evolving follower*). In the game, viral evolutionary dynamics are governed by a *biophysically-grounded lattice simulator* that accounts for binding energies and evolutionary plausibility. We evaluate Co-ADIOS on the Influenza H5 virus in two clinically relevant scenarios: (1) designing de novo antibody cocktails, and (2) co-designing a partner antibody to complement an established therapeutic. We demonstrate that Co-ADIOS antibodies outperform myopic baselines by using cooperative strategies which achieve stronger shaping over viral evolution, at a modest cost in efficacy against the current strain. Finally, our analysis shows that the designed cocktails leave the virus far fewer single mutations worth taking, and that their two antibodies bind a shared site while covering different escape mutations, offering physically explainable insights for the design of effective real-world viral therapies.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multi-Agent Systems, AI for Science, Antibodies, Viruses, Opponent Shaping, Coevolution, Game Theory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KvwZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Co-ADIOS, an extension of the ADIOS framework to the joint design of antibody cocktails that steer viral evolution towards less favourable variants. The problem is formulated as a Stackelberg evolutionary game, combining an inner viral evolution simulator with an outer cooperative co-evolutionary algorithm for antibody optimisation.

The framework is evaluated on two tasks: (I) considering both the design of new antibody pairs and (II) the optimisation of a complementary antibody for an existing therapeutic. The authors report improvements over myopic and non-cooperative baselines, together with an analysis of the resulting viral escape landscapes.

Critical strengths
  • I find the general motivation compelling. Optimising therapies while accounting for the evolutionary response they induce is an interesting and potentially important problem.

  • The framework is intuitive and reasonably well presented, and the extension to coordinated antibody combinations is well motivated.

  • The experiments show consistent improvements within the simulator, with multiple seeds, held-out evaluations, and useful analyses of the resulting escape landscapes.

Critical weaknesses
  • Overall, the central formulation and evolutionary shaping principle are inherited from ADIOS. The main extension consists of introducing cooperative co-evolution for antibody pairs (an interesting application), using an established optimisation technique. I am not entirely convinced that this constitutes a sufficiently substantial methodological contribution on its own.

  • Although the ablations are useful, I would have liked to see comparisons against stronger alternatives, particularly independently optimised ADIOS antibodies, compute-matched joint optimisation, and approaches optimising robustness across a fixed set of viral variants.

  • It is not entirely clear to me how much of the improvement comes specifically from complementary antibody behaviour rather than simply stronger or broader neutralisation.

Questions for authors
  • Q1: Can the authors compare against independently optimised ADIOS antibody pairs and a compute-matched joint optimiser, to better establish the contribution of cooperative co-evolution?

  • Q2: Can they compare against a robust optimisation baseline that targets a predefined collection of viral variants, rather than modelling cocktail-induced evolutionary trajectories?

  • Q3: Can the authors assess generalisation to another antigen or an independent binding/viral fitness model?

Should the authors address my questions I would be incliden towards weak acceptance, given the remaining doubt on methodological novelty

3 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your review! We are glad you find the motivation compelling and the framework well presented, and that you value our explainability section on the escape landscapes. We have run all experiments you asked for in Q1 to Q3 and updated the manuscript with the new results. The details are below, and all new text in the revised manuscript is in blue.

W1: Novel methodological contribution

We agree that the shaping principle and the Stackelberg framing come from ADIOS, and that cooperative co-evolution is an established technique. However, we believe extending ADIOS to antibody cocktails is both important and far from a direct application. It is important because combination therapies are the standard way to limit escape in evolving pathogens, from HIV to COVID-19 (Section 1), yet to our knowledge no opponent-shaping method exists for designing them. It also required new components beyond a different optimiser:

  • A reward for a cocktail. ADIOS scores a single antibody by its binding energy, which does not extend to two antibodies acting on the same virus. We built the cocktail reward by adapting the fractional activity model of Einav et al. on top of the binding simulator, calibrating its activity curves to Absolut! energies and splitting the dose between the antibodies (Section 3, Appendices C and E), with shared-site, independent and pose-aware binding rules (Appendix D). This reward also shows why cocktail design cannot be reduced to designing antibodies one at a time: with each antibody at dose 0.5, a cocktail never neutralises a fixed virus better than its stronger antibody given alone at dose 1, so its benefit must come from how the two antibodies jointly shape escape (Appendix C, Figure 6c).
  • A cooperative optimiser designed for this problem. CCEA is a family of algorithms rather than one fixed method, and we had to make several design choices for it to work for cocktails. Each antibody has its own sub-population, and every candidate is scored over full viral escape rollouts against the partner's current champion. The sub-populations update in turn against a frozen partner champion. We also had to make sure the same setup works for both scenarios: it designs both antibodies de novo in Scenario 1, and a complement to an existing antibody in Scenario 2 by freezing one sub-population.

W2 / Q1 / Q2: Stronger baselines at matched budget

We have added all three baselines you suggested, and every baseline is now run with the same budget of binding evaluations as Co-ADIOS (Table E.2 lists the settings and budget of every method):

  • Independent ADIOS cocktail: two ADIOS antibodies designed separately and administered at dose 0.5 each, as in every cocktail.
  • Compute-matched joint optimiser: the concatenated cocktail, which optimises both antibodies as one sequence with a single population (already included previously but now also reported at matched budget).
  • Broad-binding cocktail and single: optimised for binding across a fixed panel of wildtype mutants with up to five substitutions, instead of the escape trajectories. The panel holds 61,400 viruses, the number each Co-ADIOS candidate is scored against across its escape rollouts.

At matched budget on influenza, Co-ADIOS outperforms the independent ADIOS cocktail by +0.025+0.025 (16/20 seeds), the concatenated cocktail by +0.057+0.057 (19/20), the broad-binding cocktail by +0.135+0.135 (20/20) and the broad-binding single by +0.167+0.167 (20/20) (Figure 4d). The broad-binding baselines work well against the wildtype, but the virus escapes them within a few steps (Figure 4a), so most of the improvement comes from accounting for the viral trajectory the cocktail induces. In Scenario 2, the Co-ADIOS complement also outperforms the broad-binding complement by +0.10+0.10 on all 20 seeds (Figure 5c).

(continued below)

AuthorsOpenReview ↗

Rebuttal (2/2)

W3: Complementary behaviour versus broader neutralisation

Two results demonstrate the complementary behaviour. First, the broad-binding baselines directly optimise for broad neutralisation, and Co-ADIOS outperforms them on every seed (W2 above). Second, the new Figure 6c compares the cocktail with each of its antibodies given alone at the full dose. If the gain came from one stronger antibody, that antibody would do as well on its own. Instead, the virus escapes each antibody faster when it is given alone, and the cocktail's held-out return exceeds that of its stronger antibody by +0.10+0.10 on all 20 seeds. The new Figure I.1 (Appendix I) shows the escape landscape of each antibody of the Figure 6 cocktail given alone at the full dose: against the wildtype, 139 and 104 single substitutions raise viral activity, so each antibody alone leaves many escape routes open.

Q3: Another antigen or an independent model

We have repeated Scenario 1 on a second antigen, the SARS-CoV-2 receptor-binding domain (PDB 6XDG, new Appendix G). This also changes the viral fitness model, as the viral target is the real receptor, ACE2. Co-ADIOS outperforms the myopic, broad-binding and concatenated baselines on all 20 seeds and the independent ADIOS cocktail on 14/20 seeds, and matches the ADIOS single (+0.005+0.005, p=0.43p = 0.43). We think the last result is a limitation of the Absolut! binding simulator rather than of Co-ADIOS: 6XDG has two distinct epitopes, but under Absolut!'s lattice energy the second is seldom selected as an antibody's lowest-energy pose, so the two antibodies converge to the same site (Appendix D).

We have also tested a different model of viral evolution: re-scoring the influenza designs with an inner loop that carries 30 viruses between escape steps barely changes the results (new Appendix F). We are aware of the limitations of Absolut! and are building a higher-fidelity binding simulator, but this is a separate line of work and out of scope for this paper.

You mentioned that you would lean towards acceptance if we addressed your questions. We believe we have now addressed each of them, and we would be grateful if you would consider raising your score. If you have any remaining concerns, please let us know, and we will do our best to address them during the remaining discussion period, whether through clarification, further changes to the manuscript or additional experiments.

AuthorsOpenReview ↗

Rebuttal Follow-up

Thank you again for your constructive review. As the discussion period is ending soon, we wanted to briefly follow up on our rebuttal.

We have conducted all three sets of experiments you requested: independently optimised and compute-matched baselines (Q1), broad-binding baselines that target a collection of viral variants (Q2), and evaluation on a second antigen (Q3), alongside robustness checks on the evolutionary dynamics. The new results provide further evidence of Co-ADIOS's advantages over the baselines.

We would be very grateful to know whether these additions have addressed your concerns and, if so, whether you would consider updating your score before the discussion closes.

Thank you again for your time and feedback!

Reviewer_XKTrOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+2)
Read review
Paper summary

This paper introduces Co-ADIOS, an extension of ADIOS from individual antibodies to antibody cocktails to shape viral evolution and prevent escape. The authors formulate antibody-virus interactions as a Stackelberg evolutionary game, using a lattice-based binding model to simulate viral mutation and selection. They then use a cooperative co-evolutionary algorithm to optimize two antibodies against the resulting evolutionary trajectories. The method is evaluated on influenza H5 in two settings: designing antibody pairs from scratch and designing a complementary antibody for an existing therapeutic. They show Co-ADIOIS archives better long-term performance in simulations compared to baselines, despite somewhat weaker neutralization of the initial strain. The authors also find that the optimized antibodies often bind overlapping sites while covering different escape mutations, reducing the number of favorable evolutionary escape routes.

Critical strengths
  1. well-motivated, practical question: Clinical cocktails are built from antibodies optimised one at a time against today's strain. Designing the combination against the escape it will provoke is a natural, useful step

  2. good and clear writing: Results sections open with claims, the argument can be followed from topic sentences alone, and the biology glossary (App. A) helps ML readers.

  3. Experimental rigour (within the simulator): Fig. 4d separates the contributions of shaping, the second antibody and the optimiser. The horizon sweep (Fig. 4c) shows the benefit grows steadily with how far ahead the design looks.

Critical weaknesses

Overall, my concerns fall into two groups: whether the results reflect biology (points 1–2), and whether the experiments isolate the claimed benefit of cooperative design (points 3–4).

  1. paper does not use physical binding affinity, instead authors use Lattice energy. The viral fitness model also relies on CH65 binding as a proxy for receptor binding (Section 5.2), with activity curves calibrated within the simulator rather than against experimental data. Why not use actual ML predicted/calibrated kds. eg in https://www.pnas.org/doi/10.1073/pnas.2611571123

  2. The biological interpretation depends strongly on the assumptions of the model. In Co-ADIOS cocktails the two antibodies bind the same epitope, or one heavily overlapping with it (Sec. 6.3, Fig. 6d). This is actually mostly not true for virus. e.g for sars-cov-2 class 3 and class 4 antibodies bind to different epitopes. Approved SARS-CoV-2 cocktails mostly pair class 1 with class 2 or class 3 antibodies, e.g. casirivimab + imdevimab (REGEN-COV)

3.The comparisons do not clearly establish the benefit of cooperative optimization. ADIOS already optimizes antibodies against the viral evolutionary trajectories they induce. The main new contribution of Co-ADIOS is therefore the joint design of two antibodies, rather than evolutionary shaping itself. However, I don't think the current comparisons fully establish the benefit of this cooperative design. In particular, the "non-cooperative cocktail" baseline in Section 5.3 also optimizes the joint cocktail objective, but treats the two antibody sequences as a single concatenated sequence. It can therefore also discover complementary antibodies. The difference between the two methods is primarily how they search the sequence space, rather than whether they account for cooperation between antibodies. The improved performance of Co-ADIOS shows that its optimization algorithm works better in the tested setting, but does not necessarily demonstrate a distinct benefit from cooperative shaping.

4.It is also unclear whether the methods are compared under the same computational budget, particularly in terms of binding evaluations. A more direct comparison would be to combine two independently optimized ADIOS antibodies, and to compare Co-ADIOS with a joint-pair optimizer under matched computational budgets. It would also be useful to compare against a cocktail optimized for broad binding across a fixed set of plausible viral mutants, to clarify how much of the improvement comes specifically from anticipating the induced evolutionary trajectory.

Questions for authors
  1. The authors define fitness as R_v(v) = FA(v; a₁, a₂) − λ_vt · Δ(v)²; this seems to be over-engineered and no evidence suggest this actually capture the relationship between binding energy and fitness (or maybe I am wrong? if so please point out); instead, there has empirical established studies that map binding -> fitness for pandemic virus, e.g https://www.pnas.org/doi/10.1073/pnas.2314518121; what is the reasoning for using R_v(v) = FA(v; a₁, a₂) − λ_vt · Δ(v)²?

  2. in terms of Stackelberg setup, the inner loop keeps one mutant per step from a single lineage. That roughly matches the strong-selection, weak-mutation (SSWM) regime of population genetics, where beneficial mutations arise rarely and each one sweeps to fixation before the next appears. That regime holds when population size × mutation rate ≪ 1. Influenza within a host is usually the opposite case: populations are huge and many mutations are present at once. Why does author think the Stackelberg game is a good approx for evolutionary dynamics of influenza?

  3. In addition to the 2 questions above, If authors could address my point 1-4 in weaknesses I would be happy to increase scores

4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your thorough review and feedback. We are glad you found the question well motivated and practical, the writing clear, and the ablations and horizon sweep rigorous within the simulator. We have run new experiments to address your critique and questions. All new text in the revised manuscript is in blue.

W1: Lattice energy instead of physical binding affinity

We agree that Absolut! returns a lattice contact energy which is only an approximation of the binding free energy, and that our activity curves are calibrated within the simulator. However, Mohanty and Shakhnovich (2026) build their affinity model for a single antibody-antigen complex, so we believe it cannot accurately score the new antibodies Co-ADIOS designs. Co-ADIOS needs a binding model that can score any antibody its outer loop proposes against any viral variant its inner loop produces, and extending their model to that setting is itself an open research problem. We have added this paper to the related work (Section 2).

More generally, any learned affinity predictor faces an out-of-distribution problem under our optimisation: a search over 201120^{11} antibody sequences moves away from the training data and exploits the predictor's errors. Absolut! is a mechanistic model with no training distribution, so it avoids this problem. This is why we treat the binding model as an interchangeable component of Co-ADIOS rather than a claim about physical accuracy. We also show that results transfer beyond the poses the optimiser scored: every reported number is recomputed on a larger pose set than the one used for optimisation (Appendix E). We are aware of the limitations of Absolut! and are building a higher-fidelity binding simulator, but this is a separate line of work and out of scope for this paper. Once it is ready, Co-ADIOS can use it without any changes, since the binding model is an interchangeable component.

W2: Overlapping epitopes

You are right that the antibodies of real cocktails often bind distinct epitopes. REGEN-COV is a good example, and it is in fact the structure we use for SARS-CoV-2: PDB 6XDG is the receptor-binding domain bound by REGN10933 and REGN10987, whose epitopes share no residue. The Co-ADIOS formulation does not require the antibodies to share an epitope. Besides the shared-site rule used in our main experiments, Co-ADIOS already supported antibodies binding independent sites, and we added a pose-aware rule that treats the two antibodies as binding independently whenever their lowest-energy poses touch no common antigen residue (new Appendix D). Our SARS-CoV-2 experiments use the pose-aware rule throughout, so the cocktail is free to target two separate epitopes.

Co-ADIOS still converges to a shared site because of a constraint of the Absolut! binding simulator, not of our method. Under Absolut!'s lattice energy, the REGN10987 epitope is seldom selected as an antibody's lowest-energy pose: only 9 of the 2,519 poses that are lowest-energy for 2 million sampled antibody-antigen pairs lie on it. As a result, the two members share the same pose in 82.5% of scored states, overlapping poses in 15.3%, and disjoint poses in only 2.2%. On influenza no two coarse poses are disjoint (0 of 434,778 pairs), and re-optimising under the pose-aware rule changes the held-out return by only −0.005-0.005 (p=0.57p = 0.57). With a higher-fidelity binding simulator, such as the one we are building (W1), the pose-aware rule allows Co-ADIOS to assign the two antibodies to separate epitopes without any change to the method.

W3: Does the comparison isolate cooperative design?

You are right: the concatenated cocktail also optimises the joint objective, so beating it shows that the CCEA is the better optimiser, not that co-design beats designing the antibodies separately. To isolate co-design, we have now added the baseline you suggested: an independent ADIOS cocktail, which pairs two ADIOS antibodies designed separately and administers them at dose 0.5 each, as in every cocktail. Both antibodies shape the virus, but they are optimised independently. At matched budget, Co-ADIOS outperforms the independent ADIOS cocktail by +0.025+0.025 on influenza (16/20 seeds) and +0.017+0.017 on SARS-CoV-2 (14/20 seeds). It also outperforms the concatenated cocktail by +0.057+0.057 (19/20 seeds) on influenza (Figure 4d). Together, these show that co-design adds a measurable benefit over independently designed antibodies, and that the CCEA finds it more efficiently than a single population searching over the concatenated sequence.

(continued below)

AuthorsOpenReview ↗

Rebuttal (2/2)

W4: Matched budgets and a broad-binding baseline

Every baseline is now run with the same budget of binding evaluations as Co-ADIOS (Table E.2 lists the settings and budget of every method), and Figure 4c shows the results. On influenza, Co-ADIOS outperforms every baseline at matched budget; on SARS-CoV-2 it outperforms all but the ADIOS single, which it matches. We have also added the broad-binding cocktail and single baselines you suggested. They are optimised for binding across a fixed panel of wildtype mutants with up to five substitutions, instead of the escape trajectories. The panel holds 61,400 viruses, the number each Co-ADIOS candidate is scored against across its escape rollouts. The broad-binding baselines work well against the wildtype, but the virus escapes them within a few steps (Figure 4a). Co-ADIOS outperforms the broad-binding cocktail by +0.135+0.135 on influenza and +0.195+0.195 on SARS-CoV-2, on all 20 seeds in both cases. This shows that a large part of the improvement comes from anticipating the trajectory the cocktail induces, which a fixed panel of plausible mutants does not capture.

Q1: The form of the viral reward

Our viral reward combines escape from the cocktail, FA\mathrm{FA}, with a penalty for the receptor binding the virus loses relative to the wildtype, Δ(v)\Delta(v) (Section 4.1). It rests on the assumption that the wildtype is a functional virus that already binds its receptor well. A small loss of receptor binding should therefore cost the virus little, while a large loss, which would leave it unable to infect cells, should be penalised much more heavily, hence the quadratic penalty λvt Δ(v)2\lambda_\text{vt}\,\Delta(v)^2.

We agree that the form of Wang et al. (2024) (which we now cite in Section 2) is better grounded in data: fitness is proportional to the fraction of receptor-binding domains that are folded and either free of antibodies or bound to the receptor, so receptor and antibody binding compete within one thermodynamic model. However, its scaling factor and effective concentrations were fitted to the growth rates of SARS-CoV-2 lineages using measured dissociation constants, so they do not transfer directly to Absolut! energies, and refitting them for H5 would need matched affinity and prevalence data that we do not have. We see adopting this form, together with a higher-fidelity binding simulator, as an important next step.

Q2: Is the Stackelberg game a good approximation for influenza evolution?

The Stackelberg game itself does not assume a single viral lineage: the objective FHF_H is defined for any model of viral evolution T(⋅∣vt,a1,a2)T(\cdot \mid v_t, a_1, a_2), including a large population with many mutations present at once. Keeping a single virus at each step is a modelling choice we made for simplicity, which we now state explicitly in Section 4.2. To check that our results do not depend on this choice, we re-scored all influenza designs with an inner loop that carries 30 viruses between escape steps: at each step, 30 viruses are sampled from the current population, with replacement, by the same softmax on RvR_v as in Section 4.2, and their offspring form the next population (new Appendix F). The results barely change: Co-ADIOS beats the myopic cocktail by +0.141+0.141 instead of +0.143+0.143 and the independent ADIOS cocktail by +0.025+0.025 instead of +0.026+0.026 (Table F.1). Since all designs were optimised with a single carried virus, this also shows that their advantage transfers to a polymorphic population without re-optimisation.

You mentioned that you would be happy to increase your score if we addressed points 1 to 4 and your two questions. We believe we have now addressed each of them, and we would be grateful if you would consider raising your score. If you have any remaining concerns, please let us know, and we will do our best to address them during the discussion period, whether through clarification, further changes to the manuscript or additional experiments.

AuthorsOpenReview ↗

Rebuttal Follow-up

Thank you again for your detailed review. As the discussion period is ending soon, we wanted to briefly follow up on our rebuttal.

We have added the requested independent ADIOS (W3), compute-matched (W4), and broad-binding baselines (W4), as well as experiments on SARS-CoV-2 using the ACE2 receptor, pose-aware antibody binding that allows independent binding at disjoint epitopes (W2), and evolutionary dynamics with multiple viral lineages (Q2). We also clarified the binding simulator's limitations and the rationale behind our viral fitness formulation in our rebuttal and the revised manuscript (W1, Q1).

We hope these additional experiments and clarifications address your main concerns and would be grateful to know whether any outstanding issues remain or whether you would consider updating your score.

Thank you for your time and consideration!

I appreciate the reviewers for the detailed response and extra experiments. I still have concerns on the validity of Absolut! as a proxy for free energy but would agree that there might not be a significant better solution. I have revised the scores since the response has addressed several of my technical concerns.

Reviewer_AmwKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends ADIOS from designing one antibody to designing a two-antibody cocktail against an adapting virus. The cocktail is a Stackelberg leader, evaluated by its average reward over simulated viral escape trajectories. An outer cooperative co-evolutionary algorithm maintains one population per antibody, while an inner mutation-selection process models viral adaptation. Binding is evaluated with an Absolut! lattice model, a fractional-activity model combines antibody effects, and rewards also include receptor-binding and antitarget-binding terms. Experiments on an influenza H5 antigen consider de novo cocktails and a designed complement to a fixed partner. Within this simulator, Co-ADIOS improves held-out trajectory performance over the reported myopic, single-antibody, and concatenated-sequence baselines. Mutation scans suggest complementary escape coverage by antibodies with overlapping poses.

Critical strengths
  1. The empirical evaluation directly examines the consequence of optimizing beyond the current strain. At equal total dose, Co-ADIOS improves over the single shaper on 18/20 seeds, and Figure 4c shows increasing held-out performance as the training horizon increases from 0 to 100 (Sections 5.3 and 6.1, pp. 7–8). Figure 4a shows the corresponding reversal between wildtype efficacy and later escape. Appendix E checks neutralization separately from the combined reward and openly reports that longer horizons also incur more antitarget binding. This makes the central tradeoff more interpretable. Re-scoring on 2,763 antigen poses after optimization on 933 poses also provides a modest test beyond the precise training pose set (Appendix D).

  2. The paper goes beyond a single design scenario and offers concrete explanations for its solutions. A designed complement improves each of seven fixed partners on every reported seed (Section 6.2 and Figure 5). Section 6.3 then reports 56/880 reward-increasing single substitutions for Co-ADIOS versus 188/880 for the myopic cocktail, together with overlapping binding poses and examples of mutations covered by the other member. Appendix F shows the individual activity landscapes. These are useful, inspectable observations supporting complementary coverage within the model.

Critical weaknesses
  1. The principal comparison does not isolate cooperative design from a change in search representation and budget. The “non-cooperative cocktail” concatenates both antibody sequences and optimizes them jointly in one population (Section 5.3, p. 7). It therefore already searches pairs under the joint objective. Beating it on 19/20 seeds (Section 6.1) establishes an advantage for the proposed population decomposition under the chosen settings, but does not itself establish an advantage over pairing independently optimized ADIOS shapers. That latter comparison is suggested by the contribution statement on p. 2 but is not explicitly specified among the baselines. Moreover, Sections 4.2–4.3 introduce population sizes, rollout count, mutation rates, and selection temperature without numerical settings, and the manuscript does not establish whether the two-population method and concatenated baseline receive equal fitness-evaluation or binding-evaluation budgets. Appendix D reports pose sets, calibration, and dosing, but does not resolve this comparison. A matched-budget comparison, including independently designed shapers paired at equal total dose, could materially change the evidence for the claimed coordination benefit.

  2. The mechanistic conclusion rests on consequential modeling assumptions whose robustness is untested. Appendix C assumes all cocktail members compete for one binding site, irrespective of their actual pose compatibility. Section 5.2 uses CH65, an antibody, as a proxy for the sugar receptor, and Appendix D anchors its activity at 0.5 for the wildtype and specifies a dose lifetime of 86 simulated steps. The inner process retains one selected viral lineage per step (Section 4.2). Held-out random seeds and an expanded nested pose set preserve these assumptions. No reported sensitivity analysis establishes whether the coordination advantage or overlapping-site strategy survives plausible changes in receptor constraints, competition rules, or evolutionary dynamics. Also, Appendix F evaluates individual members at their cocktail doses; under Equation C.2, the pair necessarily has lower activity than either member at that dose. Thus, Figure 6c alone cannot independently demonstrate emergent cooperation. The paper acknowledges its coarse simulator and single-antigen scope, appropriately, but the current evidence supports a narrower simulation-specific mechanism than the therapeutic framing suggests.

Questions for authors
  1. What are the exact search and rollout settings and the total binding/fitness-evaluation budgets for each arm? Under matched budgets, how does Co-ADIOS compare with cocktails assembled from independently optimized ADIOS antibodies?

  2. Which available experiments show that the coordination advantage and overlapping-site result persist when the receptor penalty/calibration or antibody-competition assumption changes? Can you distinguish useful trajectory-level complementarity from the guaranteed activity improvement obtained by adding a second binder at fixed per-member doses?

3 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your detailed review! We are glad that you found the paper clear and easy to follow, that the evaluation directly examines the consequences of optimising beyond the current strain, and that the mutation scans and per-member landscapes give inspectable evidence of complementary coverage. To address your questions and the two weaknesses you’ve pointed out, we’ve run new experiments and updated the manuscript (changes in blue).

W1 / Q1: Isolating co-design, matched budgets and experimental settings

You are right that the concatenated cocktail already searches pairs under the joint objective, so beating it shows that CCEA is the better optimiser, not that co-design beats independent design. We have now added the comparison you suggested, and matched all budgets.

  • Independent ADIOS cocktail. We pair two ADIOS antibodies, each designed separately at full dose as ADIOS would design it, and administer them at dose 0.5 each, as in every cocktail (a single antibody receives dose 1). This isolates co-design: both members shape, but they are optimised independently.
  • Matched budgets. Every baseline now receives the same number of binding evaluations as Co-ADIOS (one evaluation is one call to the binding simulator). The new Table E.2 lists the population, generations, horizon, fitness evaluations and binding evaluations of every method in both scenarios, and Appendix E lists every search and rollout setting (Pv=30P_v = 30 offspring per escape step with μv=1\mu_v = 1 substitution, selection temperature β=5×10−4\beta = 5 \times 10^{-4}, K=20K = 20 training rollouts at H=100H = 100, populations of 100 per antibody for the CCEA and 200 for single-population searches).

At matched budget, Co-ADIOS improves on the independent ADIOS cocktail by +0.025+0.025 (16/20 seeds) and on the concatenated cocktail by +0.057+0.057 (19/20) on influenza (Figure 4d). It leads every baseline at every budget above 0.45×1090.45 \times 10^9 binding evaluations (Figure 4c). On SARS-CoV-2, it improves on the independent ADIOS cocktail on 14/20 seeds and on the concatenated cocktail on 20/20 (Appendix G). Co-design therefore adds a measurable benefit over pairing independently designed shapers, and the CCEA finds it more efficiently than a joint search over the concatenated sequence.

(continued below)

AuthorsOpenReview ↗

Rebuttal (2/2)

W2 / Q2: Robustness to modelling assumptions

Competition rule. Besides the shared-site rule used in our main experiments, Co-ADIOS already supported antibodies binding independent sites. We now describe both in the new Appendix D and add a pose-aware rule that treats the two antibodies as binding independently whenever their lowest-energy poses touch no common antigen residue. Re-optimising Co-ADIOS on influenza under the pose-aware rule changes its held-out return by −0.005-0.005 (p=0.57p = 0.57). This is because no two coarse poses on 5JW4 are disjoint (0 of 434,778 pairs). On SARS-CoV-2, which uses the pose-aware rule throughout, the two members occupy the same pose in 82.5% of scored states, different but overlapping poses in 15.3%, and disjoint poses in only 2.2%. The overlapping-site strategy is therefore a property of the simulated antigens, not an artefact of the competition rule.

Receptor constraint. On SARS-CoV-2, the viral target is the real receptor, residues 24 to 43 of the ACE2 helix that binds the receptor-binding domain, rather than the CH65 antibody proxy we use for influenza. The ordering of methods is preserved, except that Co-ADIOS matches the ADIOS single (+0.005+0.005, p=0.43p = 0.43) rather than significantly outperforming it. We think this is a limitation of the lattice model rather than of Co-ADIOS. The structure we use, PDB 6XDG, is the receptor-binding domain bound by the two antibodies of the REGEN-COV cocktail, REGN10933 and REGN10987, whose epitopes share no residues. Under Absolut!'s lattice energy, however, the REGN10987 epitope is seldom selected as an antibody's lowest-energy pose: it does so in only 9 of the 2,519 poses that win for 2 million sampled antibody-antigen pairs. Co-ADIOS therefore has no second epitope to exploit under the current binding simulator, and on this antigen the two members converge to antibodies that share the same pose in 82.5% of cases. A higher-fidelity binding simulator would remove this limitation. We are currently working on building one, but it is a separate line of work and out of scope for this paper. Since Co-ADIOS treats the binding model as a replaceable component, it can be applied to such a simulator unchanged.

Evolutionary dynamics. We re-scored all influenza designs with an inner loop that carries 30 viruses between escape steps instead of one (new Appendix F). The results barely change: Co-ADIOS beats the myopic cocktail by +0.141+0.141 instead of +0.143+0.143 and the independent ADIOS cocktail by +0.025+0.025 instead of +0.026+0.026 (Table F.1).

Trajectory-level complementarity versus a guaranteed activity improvement (new Figure 6c). At half dose, Equation C.2 indeed guarantees that the pair leaves less virus active than either member. Under the same equation, a pair at dose 0.5 each can never neutralise a fixed virus better than its stronger member given dose 1. Any advantage over a full-dose member must therefore come from the escape trajectory. We now give each member the full dose and let the virus escape it on its own held-out rollouts. Each member alone is escaped faster than the pair (Figure 6c), and the pair's held-out return exceeds that of its stronger member by +0.10+0.10 [+0.085,+0.117][+0.085, +0.117] on all 20 seeds (Section 6.3). The new Figure I.1 (Appendix I) shows the escape landscape of each member of the Figure 6 cocktail given alone at the full dose: against the wildtype, 139 and 104 single substitutions raise viral activity against a1a_1 and a2a_2, respectively, so each member alone leaves many escape routes open.

We hope these additions and clarifications address your concerns. If they do, we would be grateful if you could consider raising your score. If you have any remaining concerns, please let us know, and we will do our best to address them during the discussion period, whether through clarification, further changes to the manuscript or additional experiments.

Thank you for the response. The clarifications have adequately addressed my concerns. I appreciate the authors' efforts and will raise my score accordingly.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Rebuttal Response

We thank all three reviewers for their careful and constructive reviews. We are glad that the reviewers found the motivation compelling (KvwZ) and the question well motivated and practical (XKTr), the paper clear and easy to follow (all three reviewers), and the experiments rigorous within the simulator, with held-out evaluation and useful analyses of the escape landscapes (XKTr, KvwZ, AmwK).

The reviewers raised four shared concerns: whether the comparisons isolate the benefit of co-designing the cocktail (all three reviewers), whether methods are comparable at equal budget (all three), how Co-ADIOS compares with a cocktail optimised for broad binding across a fixed panel of variants (XKTr, KvwZ), and whether the results depend on our modelling choices and on a single antigen (AmwK, XKTr, KvwZ). We have run new experiments to address each of these and uploaded a revised manuscript, with all new and changed text in blue. In summary:

1. Stronger baselines, all at matched budget (new Figures 4 and 5, Table E.2). We added the two baselines the reviewers asked for. The independent ADIOS cocktail pairs two ADIOS antibodies designed separately, each administered at dose 0.5 as in every cocktail. The broad-binding cocktail and single baselines are optimised against a fixed panel of wildtype mutants with up to five substitutions instead of escape trajectories. The panel holds 61,400 viruses, the number each Co-ADIOS candidate is scored against across its escape rollouts. The concatenated cocktail, which optimises the pair as one sequence, is now also reported at matched budget. Every baseline now receives the same budget of binding evaluations as Co-ADIOS, and Appendix E lists every search setting and budget. On influenza, Co-ADIOS outperforms all seven Scenario 1 baselines: the independent ADIOS cocktail by +0.025+0.025 (16/20 seeds), the ADIOS single by +0.039+0.039 (16/20), the concatenated cocktail by +0.057+0.057 (19/20), the broad-binding cocktail by +0.135+0.135 (20/20) and the myopic cocktail by +0.223+0.223 (20/20). In Scenario 2, the Co-ADIOS complement beats the partner alone and the myopic and broad-binding complements on all 20 seeds.

2. A second antigen (new Appendix G). We repeated Scenario 1 on the SARS-CoV-2 receptor-binding domain (PDB 6XDG), where the viral target is the real receptor, the ACE2 helix. Co-ADIOS outperforms the myopic, broad-binding and concatenated baselines on all 20 seeds and the independent ADIOS cocktail on 14/20, and matches the ADIOS single (+0.005+0.005, p=0.43p = 0.43). Across both antigens, Co-ADIOS significantly outperforms 13 of the 14 baselines at matched budget and matches the last.

3. Robustness to modelling assumptions (new Appendices D and F). Besides the shared-site rule used in our main experiments, Co-ADIOS already supported antibodies binding independent sites. We now describe both rules in the new Appendix D and add a third, pose-aware rule that switches to independent binding whenever the two antibodies' poses are disjoint. Re-optimising on influenza under the pose-aware rule changes the held-out return of Co-ADIOS by −0.005-0.005 (p=0.57p = 0.57), and the SARS-CoV-2 runs use it throughout. We also re-scored all influenza designs with an inner loop that carries 30 viruses between escape steps instead of one (Appendix F). The results barely change: Co-ADIOS beats the myopic cocktail by +0.141+0.141 instead of +0.143+0.143, and the independent ADIOS cocktail by +0.025+0.025 instead of +0.026+0.026.

We would be grateful if the reviewers could let us know whether these changes address their concerns and, if so, consider raising their scores.

Area_Chair_LDE2OpenReview ↗

Please respond

Hi reviewers,

Thanks for your work so far on this paper. Please respond to the authors before the end of the discussion period if you have not done so already, even if to acknowledge their response.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.