Ola Kalisz
@olakalisz.bsky.social
PhD student at FLAIR, University of Oxford
@flair-ox.bsky.social, @universityofoxford.bsky.social
Previously a Senior AI Research Scientist at Exscientia
@flair-ox.bsky.social, @universityofoxford.bsky.social
Previously a Senior AI Research Scientist at Exscientia
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! ⚔️
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! ⚔️
July 2, 2025 at 1:11 PM
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! ⚔️
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! 💊
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! 💊
July 2, 2025 at 1:11 PM
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! 💊
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 🎯
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 🎯
July 2, 2025 at 1:11 PM
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 🎯
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
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
July 2, 2025 at 1:11 PM
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
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
🔄 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
July 2, 2025 at 1:11 PM
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
🔄 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
✅ 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
July 2, 2025 at 1:11 PM
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
✅ 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
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
July 2, 2025 at 1:11 PM
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
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
🧵👇
Our #ICML2025 paper ADIOS proposes "shaper therapies" that steer viral evolution in our favour & remain effective. Work done @flair-ox.bsky.social
🧵👇
July 2, 2025 at 1:11 PM
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
🧵👇
Our #ICML2025 paper ADIOS proposes "shaper therapies" that steer viral evolution in our favour & remain effective. Work done @flair-ox.bsky.social
🧵👇