ADIOS: Antibody Development via Opponent Shaping

arXiv:2409.10588 · q-bio.PE, cs.AI, cs.GT, cs.MA · Submitted 2025-06-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "ADIOS: Antibody Development via Opponent Shaping".

Jane: The paper was written by Sebastian Towers, Aleksandra Kalisz, Philippe A. Robert, Alicia Higueruelo, Francesca Vianello et al. from University of Oxford and University of Basel and Isomorphic Labs and Exscientia and Epsilogen Ltd..

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everybody. Today we're looking at a paper that's got a fantastic acronym — ADIOS: Antibody Development via Opponent Shaping. Jane, I have to say, the title alone made me sit up.

Jane: It grabbed me too, Tom. And the name is actually perfect once you understand what they're doing. "Opponent shaping" comes from game theory and multi-agent AI — it's about influencing how your opponent learns and evolves, not just beating them at the current moment.

Tom: So instead of designing an antibody that just works against today's virus, they're designing one that actually steers how the virus mutates in the future. That's a wild idea.

Jane: Exactly. And that's why the title matters. ADIOS isn't just a catchy name — it's literally describing the method. The antibody is saying "adios" to the virus's ability to escape over time.

Tom: And I love that they frame it as a two-player game. The virus wants to evade the antibody, the antibody wants to bind to the virus. But the clever twist is that the antibody's moves today change what the virus learns tomorrow.

Jane: Right. And that's the "opponent shaping" part. Most therapy design is myopic — you target the current strain, and then you're surprised when the virus evolves resistance. ADIOS says, let's anticipate that evolution and push it in a direction we want.

Tom: So the title is really a promise. It's saying we're not just developing antibodies, we're developing antibodies that shape the opponent's future behavior.

Jane: And that's a fundamentally different way of thinking about medicine. It's not just "hit the target" — it's "make the target evolve into something weaker."

Tom: I'm already hooked. Let's get into what they actually did.

Summary: Jane: So Tom, let's break down what this paper actually does. They built a framework where you have two nested loops. The inner loop simulates the virus evolving to escape a given antibody. The outer loop optimizes the antibody to perform well across those simulated escape trajectories.

Tom: So it's like a training simulator for antibodies. You let the virus "practice" escaping, and then you pick the antibody that survives the longest.

Jane: Exactly. And the key insight is that the antibody's fitness isn't just about binding strength to the current virus. It's about the average payoff across a whole trajectory of viral mutations.

Tom: And they use this thing called Absolut! to simulate binding — it's a protein interaction simulator. But here's the kicker — they reimplemented it in JAX and got a ten-thousand-fold speedup.

Jane: That's huge. Because to do this opponent shaping, you need to run thousands of viral escape simulations, and each one requires millions of binding calculations. Without that speedup, the whole approach would be computationally impossible.

Tom: Right. And they tested it on the dengue virus, but also on West Nile, influenza, MERS-CoV, and even a bacterium. In every case, the "shaper" antibodies outperformed the myopic ones on the long-term objective.

Jane: And the really cool part — they showed that the shapers don't just resist escape. They actively shape the virus's evolution so that the final variants are more vulnerable to other antibodies too.

Tom: So it's not just "my antibody is better." It's "my antibody makes the virus weaker in general." That's a game-changer.

Jane: It really is. And it's worth emphasizing — the paper is a proof of concept. The binding simulator is simplified. But the framework is the contribution, and it's designed to work with better simulators as they come along.

Tom: So the summary is: they've shown that thinking about viral evolution as an opponent you can shape, rather than a static target, leads to fundamentally better therapies.

Improvements: Tom: So Jane, what does this paper actually improve upon? I mean, antibody design has been around for a while.

Jane: Right, and that's the key point. Most existing antibody design methods optimize for binding to a known antigen structure. They use energy-based methods, language models, graph neural networks — all trying to find the antibody that binds best to the current virus.

Tom: And that's the myopic approach.

Jane: Exactly. And there's also work on predicting viral escape — like EVEscape, which tries to forecast which mutations will emerge. But those methods treat the antibody as a fixed pressure. They don't ask "how should the antibody change to make escape harder?"

Tom: So ADIOS is the first to connect those two things — predicting escape and optimizing the antibody — into a single loop.

Jane: Precisely. And that's the improvement. They're not just saying "here's a better binding predictor." They're saying "here's a new objective function for antibody design that accounts for the virus's adaptive response."

Tom: And the results back it up. The shapers with horizon one hundred significantly outperformed myopic antibodies on the escape-averaged fitness. But I noticed there's a trade-off — the shapers actually did worse on immediate binding.

Jane: That's a really important finding. The shapers sacrifice some immediate potency for long-term robustness. And the paper even shows that the optimal training horizon depends on your compute budget. If you're limited, a shorter horizon like twenty can be a better proxy than one hundred.

Tom: So it's not just "bigger horizon is always better." It's about matching the horizon to your resources.

Jane: Right. And they also showed that the shaping effect persists even when there's external pressure from other antibodies — which is more realistic, because in real life you might have multiple therapies circulating.

Tom: So the improvement is really about the mindset. Instead of asking "what binds best today," you ask "what makes the virus evolve into something we can handle tomorrow."

Conclusion: Tom: Alright, let's wrap this up. We've been talking about ADIOS: Antibody Development via Opponent Shaping, and I think it's fair to say this is one of those papers that makes you see a whole field differently.

Jane: Absolutely. The core idea is simple but profound — treat the virus as an adaptive opponent, not a static target. And then optimize your therapy to shape that opponent's evolution, not just to beat its current form.

Tom: And the practical implications are enormous. Think about COVID — we saw vaccines lose effectiveness against new variants. ADIOS offers a path toward therapies that don't just fight today's strain but actively steer the virus toward weakness.

Jane: And it's not just viruses. The paper mentions cancer — monoclonal antibodies for tumors. Cancer cells evolve resistance too. The same opponent shaping principle could apply there.

Tom: And antimicrobial resistance — bacteria evolving to escape antibiotics. Same problem, same potential solution.

Jane: Right. Now, we should be clear — this is a proof of concept with a simplified simulator. You can't go out and design a real drug with this yet. But the framework is designed to plug in better simulators, like AlphaFold3, as they become available.

Tom: So it's a foundation. A new way of thinking about therapy design that could eventually lead to longer-lived vaccines and treatments.

Jane: And that's why we're excited. It's not just a better algorithm — it's a better question to ask. And sometimes, asking the right question is the biggest step forward.

Tom: Well said, Jane. That's a wrap on ADIOS. Thanks to everyone for listening, and we'll see you on the next one.

Jane: Goodbye, everyone!

Sebastian Towers, Aleksandra Kalisz, Philippe A. Robert, Alicia Higueruelo, Francesca Vianello, Ming-Han Chloe Tsai, Harrison Steel, Jakob Foerster

University of Oxford · University of Basel · Isomorphic Labs · Exscientia · Epsilogen Ltd.

q-bio.PE, cs.AI, cs.GT, cs.MA

Submitted: 2025-06-06

Updated: 2026-08-12

Comments: Accepted at ICML 2025

Journal ref: Proceedings of the 42nd International Conference on Machine Learning (ICML 2025), PMLR 267

Code: https://github.com/olakalisz/adios

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 70/100

The gist: ADIOS: Antibody Development via Opponent Shaping Summary This paper introduces ADIOS (Antibody Development vIa Opponent Shaping), a meta-learning framework for designing antibody therapies that not

Terminology

Summary

ADIOS: Antibody Development via Opponent Shaping

Summary

This paper introduces ADIOS (Antibody Development vIa Opponent Shaping), a meta-learning framework for designing antibody therapies that not only defend against current viral strains but also actively influence, or shape, which future viral variants emerge. The authors frame the interaction between antibodies and viruses as a two-player zero-sum game, where the antibody's payoff is determined by its binding strength to the virus while avoiding an anti-target, and the virus has the opposite payoff.

The method consists of two nested optimisation loops. In the inner loop, Simulated Viral Escape via Evolution, the virus adapts to a fixed antibody by repeatedly finding approximate best responses that decrease binding strength over a given horizon length. In the outer loop, Antibody Optimisation, a genetic algorithm optimises antibodies to be effective across viral evolutionary trajectories, resulting in antibodies called shapers. This contrasts with myopic antibodies that only optimise for binding to the initial virus.

To meet computational demands, the authors reimplemented the core binding calculation of the Absolut! framework in JAX, achieving a 10,000-fold speedup compared to the original implementation. The binding function uses the Miyazawa-Jernigan energy potential matrix and binding poses generated by Absolut!.

Key results from the paper include:

  1. Shapers outperform myopic antibodies: Shapers optimised with horizon H=100 significantly outperform myopic antibodies in the escape-averaged objective Fv 100(a). None of the myopic antibodies outperform any of the top 10% of shapers in this long-term objective. However, there is a trade-off: shapers underperform on the myopic objective Ra(v,a).

  2. Viral escape curves: Shapers learn to influence viral trajectories to minimise long-term viral escape, albeit at the cost of initial performance. While myopic antibodies offer better immediate control, shapers provide more sustained effectiveness against evolving viral populations.

  3. Varying horizons: Shapers optimised with longer horizons H consistently yield better performance when evaluated against a consistent true objective of H=100. However, when accounting for computational cost (total number of binding samples), H=20 shapers perform strongly, nearly matching H=100 performance, suggesting that shorter-horizon proxies can yield substantial benefits.

  4. Generalisation to other pathogens: ADIOS was evaluated on three additional viruses (West Nile, Influenza, MERS-CoV) and the bacterium Clostridium Difficile. In all cases, shapers outperform myopic antibodies in escape-averaged fitness, confirming that ADIOS can successfully achieve shaping across diverse pathogens.

  5. Attack is the best defence: The authors disentangle robustness (defence) from shaping (attack) strategies. Viruses induced by H=100 shapers are consistently more exploitable by antibodies across all optimisation horizons, suggesting that H=100 shapers actively shape escape trajectories to make resulting variants more susceptible to antibody binding in general. However, this shaping comes at a cost: H=100 shapers show slightly lower payoffs compared to shorter-horizon antibodies against the viruses they induce.

  6. Shaping with external pressure: When simulating viral escape with additional external pressure from a separate myopic antibody, the shaping effect remains clearly visible, though somewhat reduced, demonstrating transfer to scenarios with multiple therapeutic pressures.

  7. Explainability analysis: Long-horizon shapers exhibit more uniform amino acid distributions, while myopic antibodies cluster around amino acids with extreme binding energies. The authors hypothesise that maintaining diversity in amino acid composition helps shapers preserve robustness against viral mutation. Additionally, pose matrix analysis reveals that H=100 shapers actively constrain viral evolution by both preventing unfavourable binding configurations (preventing the virus from including the antibody's lowest binding amino acids in the pose) and preserving favourable ones (inhibiting the virus from removing its own high-binding amino acids from the pose).

The authors conclude that while their current implementation uses simplified binding and evolutionary escape models that prevent direct therapeutic application, ADIOS could be integrated with more sophisticated models like AlphaFold3. As computational models of protein interactions and evolutionary processes improve, ADIOS has the potential to transform therapy development against viruses, cancers, and other evolving adversaries.

Improvements for AI systems

Based on the paper, here are specific improvements I can implement in AI systems:

Improvement: Build a meta-learning framework where the outer optimization loop explicitly models and shapes the opponent's (virus/pathogen) adaptive response, rather than optimizing against a static target.

What the improved system can do:

  • Design antibodies/therapies that remain effective across evolutionary trajectories (100+ mutation steps), not just against current strains

  • Actively steer pathogen evolution toward weaker variants (shaping, not just robustness)

  • Optimize for long-horizon fitness (average payoff over escape trajectory) instead of myopic single-step binding

Improvement: Reimplement the Absolut! binding calculation in JAX with GPU acceleration, pruning to only high-relevance poses (top 100 poses cover 95% of binding configurations).

Improvement: Implement the inner loop as a stochastic viral evolution process (Algorithm 1) with fitness-proportional selection (softmax with β) and explicit anti-target/target constraints.

Improvement: Implement the outer loop (Algorithm 2) with tunable horizon H and a compute-normalized evaluation metric (number of binding samples).

Improvement: Apply the same framework to multiple antigens (dengue, West Nile, Influenza, MERS-CoV) and a bacterium (Clostridium difficile) with consistent shaping results.

Improvement: Implement pose-matrix analysis (20×20 amino acid interaction counts) and amino acid distribution tracking to explain shaping strategies.

Improvement: Extend the payoff function to include external therapeutic pressures (Equation 3), simulating scenarios with multiple concurrent therapies.

Improvement: Implement a two-tier verification: train on low-resolution (fast) simulator, verify on high-resolution (accurate) simulator.

Improvement: Use genetic algorithm with population size Pa=40, single-point mutations, and η=5 Monte Carlo rollouts for fitness estimation.

Improvement: Frame the interaction as a zero-sum game where antibody payoff = B(virus, antibody) - B(anti-target, antibody) - B(virus, target).

Bottom line: The improved AI system can design therapeutic antibodies that not only neutralize current viral strains but actively manipulate viral evolution to produce weaker, more targetable variants—while being computationally efficient enough for practical deployment and generalizable across multiple pathogens.

Sources

Related papers