Research map / Label-efficient and robust learning / Recognition and long-tail
ADIOS: Antibody Development via Opponent Shaping
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ADIOS, a meta-learning opponent-shaping framework that optimises antibodies against simulated evolutionary escape of viruses, using a GPU-accelerated JAX reimplementation of Absolut! to find "shapers" that outperform myopic antibodies.
Problem
Antiviral therapies are designed against current viral strains only (myopic), so therapy-induced selective pressure drives escape variants that render them ineffective over time.
Approach
Frame antibody-virus interaction as a two-player zero-sum game and optimise antibodies in an outer loop against escape trajectories simulated in an inner loop (virus naively evolving via mutation and selection), so antibodies both stay effective and steer viral evolution toward weaker variants.
Key results
- Shapers (H=100) have higher escape-averaged fitness F^100 than myopic antibodies; none of the myopic antibodies outperform the top 10% of shapers on this objective.
- A JAX reimplementation of Absolut! achieves about 10,000x speedup (1.8 s on Apple M2 Max CPU vs 2.1e-4 s per antigen on Nvidia A40).
- Shapers trained with H=20 nearly match H=100 per meta-step and far exceed it under equal binding-sample budgets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the JAX GPU-accelerated Absolut! reimplementation instead of the original Absolut! implementation | Absolut! · Absolut! binding query | time/antigen: 1.8 -> 2.1 × 10^{-4}; 10,000-fold speedup | Helped |
| Optimise antibodies for escape-averaged fitness with H=100 rather than for the initial virus only | ADIOS antibody optimisation · Dengue virus | escape-averaged antibody fitness F^{100}_v(a) | Helped |
| Optimise antibodies for escape-averaged fitness with H=100 rather than for the initial virus only | ADIOS antibody optimisation · Dengue virus | myopic antibody payoff R_a(v,a) | Hurt |
| Optimise antibodies with H=100 instead of the myopic objective | ADIOS antibody optimisation · Dengue virus, late viral escape stages | virus fitness/payoff R_v(v,a) | Helped |
| Optimise antibodies for escape-averaged fitness rather than only current-virus fitness | ADIOS · West Nile Virus | escape-averaged antibody fitness | Helped |
| Optimise antibodies for escape-averaged fitness rather than only current-virus fitness | ADIOS · Influenza Virus | escape-averaged antibody fitness | Helped |
| Optimise antibodies for escape-averaged fitness rather than only current-virus fitness | ADIOS · MERS-CoV | escape-averaged antibody fitness | Helped |
| Optimise antibodies for escape-averaged fitness rather than only current-virus fitness | ADIOS · Clostridium Difficile | escape-averaged antibody fitness | Helped |
| Increase the optimisation horizon to H=100 rather than using shorter horizons (ablation) | ADIOS antibody shapers · Dengue virus | escape-averaged antibody fitness F^{100}_v(a) | Helped |
| Use H=20 instead of H=100 under a fixed binding-sample budget (ablation) | ADIOS antibody shapers · Dengue virus | escape-averaged antibody fitness F^{100}_v(a) | Helped |
| Use H=100 instead of H=20 with 30 antibody optimisation steps (ablation) | ADIOS antibody shapers · West Nile Virus and MERS-CoV | escape-averaged antibody fitness | Hurt |
| Add external evolutionary pressure from a separate myopic antibody (ablation) | ADIOS viral escape simulation · Dengue virus | shaping effect | Hurt |
Surprising findings
- Shapers underperform myopic antibodies on the immediate myopic objective, reflecting a short- vs long-term trade-off.
- For West Nile and MERS-CoV, H=100 shapers performed worse than H=20 at 30 meta-steps, attributed to non-convergence.
Limitations
- Binding simulator Absolut! is a toy model and not an accurate model of antibody binding.
- Viral escape is simulated with a naive evolutionary model; the antigen structure is assumed fixed during escape.
- Results only for 30 meta-steps; long-horizon shapers not converged in some settings.
Experimental setup
Datasets: Absolut! antibody sequences PDB 2R29 (dengue) PDB 1ZTX (West Nile) PDB 4QNP (influenza neuraminidase) PDB 5DO2 (MERS-CoV) PDB 4NP4 (C. difficile)
Benchmarks: Absolut! binding simulator (dengue, West Nile, influenza, MERS-CoV, C. difficile)
Metrics: escape-averaged antibody fitness F^H myopic antibody payoff R_a viral payoff R_v over escape steps compute-normalised fitness
Baselines: Myopic antibodies (H=0 objective)
Builds on: Learning with Opponent-Learning Awareness Model-Free Opponent Shaping Absolut! JAX
Closest papers by meaning
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.85)
- Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences (ICLR 2025 · similarity 0.84)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.83)
- A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer (ICLR 2025 · similarity 0.82)
- dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen (AAAI 2025 · similarity 0.81)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.