Ola Kalisz

ADIOS: Antibody Development via Opponent Shaping Anti-viral therapies are typically designed to target only the current strains of a virus, a myopic response. However, therapy-induced selective pressures drive the emergence of new viral strains, aga...

10/🔗 To learn more, check out our paper and blog post! Our code is also fully open-sourced:

📄 Paper: arxiv.org/abs/2409.10588
🌐 Blog: olakalisz.github.io/adios-blog
🤖 Code: github.com/olakalisz/ad...

Excited to see where this research leads! 🚀

9/🙏 Huge thanks to my amazing coauthors who made this work possible!

Sebastian Towers, Philippe A. Roberts, Alicia Higueruelo, Francesca Vianello, Ming-Han Chloe Tsai, Harrison Steel, @jfoerst.bsky.social

This work was done at FLAIR @flair-ox.bsky.social , @ox.ac.uk

8/🧬 This is just the beginning...

ADIOS uses simplified models, but the core insight is huge: we can design therapies that remain effective AND guide evolution itself.

Next: cancer, antimicrobial resistance, any domain where we're fighting adaptive biological opponents! 🦠

7/🤯 But wait, there's more: shapers don't just defend better - they actively shape viral evolution 🧬

Viruses that evolved under pressure from H=100 shapers become easier for ALL antibodies to target - dark colours in the right-most column.

Attack is the best defense! ⚔️

6/🚀 And longer horizons work better...

As we increase the shaping horizon (H), antibodies get better at preventing long-term viral escape. H=100 shapers consistently outperform shorter horizons.

The meta-learning approach pays off - we're generating long-lasting therapies! 💊

5/📈 So do shapers actually work? Yes! ✅

We tested on dengue virus. Myopic antibodies start strong but lose effectiveness as the virus evolves - viral fitness goes up!

Shapers start slightly worse but maintain performance much better over time. They're playing the long game 🎯

4/⚡️ But there was a problem: this approach is computationally expensive

We needed fast binding calculations for hundreds of thousands of antibody-virus interactions.

💻 Solution: Reimplement the core simulator in JAX with GPU acceleration.
🚀 Result: 10,000x speedup

3/🧠 So how does ADIOS work? Through opponent shaping and meta-learning with 2 loops:

🔄 Inner loop: simulate how the virus evolves in response to each candidate antibody
🔄 Outer loop: optimise antibodies based on their performance across entire viral evolutionary trajectories

2/📉 Here's what happens with myopic therapies
✅ Therapy works against Virus A
❌ Selective pressure pushes evolution toward resistant Virus B
🔄 Back to square one

ADIOS flips this: instead of Virus B being resistant, we steer evolution toward easily-targetable Virus C

1/💡 Here's the key insight: viruses are stuck being myopic players

They evolve through trial-and-error mutations, reacting to selective pressures, including those our therapies create

But WE can (and SHOULD!) think ahead and anticipate how our therapies shape viral evolution

Antiviral therapy design is myopic 🦠🙈 optimised only for the current strain. That's why you need a different Flu vaccine every year!

Our #ICML2025 paper ADIOS proposes "shaper therapies" that steer viral evolution in our favour & remain effective. Work done @flair-ox.bsky.social

🧵👇

If you are currently applying for PhD positions and are interested in the overlap of RL and LLMs you should definitely check out the FAIR-Oxford program!