ADIOS: Antibody Development via Opponent Shaping

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

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the value of ADIOS, we build a viral evolution simulator using the Absolut! framework, in which shapers successfully target both current and future viral variants, outperforming myopic antibodies. Furthermore, we show that shapers modify the distribution over viral evolutionary trajectories to result in weaker variants. We believe that our ADIOS paradigm will facilitate the discovery of long-lived vaccines and antibody therapies while also generalising to other domains. Specifically, domains such as antimicrobial resistance, cancer treatment, and others with evolutionarily adaptive opponents. Our code is available at https: //github.com/olakalisz/adios.
Researcher Affiliation Collaboration 1FLAIR, Foerster Lab for AI Research 2Department of Engineering, University of Oxford, Oxford, UK 3Department of Biomedicine, University of Basel, Basel, Switzerland 4Isomorphic Labs, London, UK 5Exscientia, Oxford, UK 6Epsilogen Ltd., London, UK. Correspondence to: Sebastian Towers <EMAIL>, Aleksandra Kalisz <EMAIL>.
Pseudocode Yes Algorithm 1 Ev(ˆv, a): Simulated Viral Escape (Inner Loop) Algorithm 2 Antibody Optimisation (Outer Loop)
Open Source Code Yes Our code is available at https: //github.com/olakalisz/adios.
Open Datasets Yes We use the antigen protein from the Dengue Virus for our main experiments, specifically, the structure with Protein Data Bank (PDB) code 2R29 (Berman et al., 2000; Lok et al., 2008). ... The three viral antigens we use are: West Nile Virus, PDB code 1ZTX (Nybakken et al., 2005); Influenza Neuraminidase Virus, PDB code 4QNP (Wan et al., 2015) and MERS-Co V Virus with PDB code 5DO2 (Li et al., 2015). Furthermore, we show that ADIOS can be easily applied to other pathogens, such as bacteria, too. We perform an extra experiment with the Clostridium Difficile Bacterium, PDB code 4NP4 (Orth et al., 2014).
Dataset Splits No The paper describes simulation parameters like population size (P=15 for viruses, Pa=40 for antibodies) and number of Monte Carlo roll-outs (η=5), but it does not specify explicit training/validation/test splits for a pre-collected dataset in the conventional machine learning sense. The process involves simulated viral escape and antibody optimization over 'generations' or 'steps' within the simulation.
Hardware Specification Yes Absolut! Absolut! + JAX Hardware Apple M2 Max Nvidia A40 Time/Antigen (s) 1.8 2.1 10 4
Software Dependencies No The paper mentions "JAX (Bradbury et al., 2018)" and "Absolut! (Robert et al., 2022)" but does not provide specific version numbers for these software components.
Experiment Setup Yes In our experiments P = 15. For every virus in the population, we evaluate its fitness given by Rv(vi k, a). We then sample a new virus ˆvi+1 based on the fitness values, in particular: P(ˆvi+1 = vi k) exp(βRv(vi k, a)) With duplicates in the population being considered distinct, so that the likelihood of a particular variant increases with the number of duplicates. Furthermore, β is a constant which reflects how random the selection process is, with β reflecting deterministic max-fitness selection. After H generations, a full escape trajectory ˆv = [ˆv0, ˆv1, ˆv2, . . . , ˆv H] has been generated and the simulated viral escape process ends. ... In most of our experiments N = 30. ... For our experiments, Pa = 40. ... We found η = 5 to be sufficient.