Odysseas Vavourakis

Predicting protein conformational flexibility remains a major challenge in structural biology. While we can now accurately model static protein structures, understanding their dynamics is still difficult, largely due to a lack of suitable training data.

Huge thanks 🙌 to my fellow members of @opig.stats.ox.ac.uk:

- our lead author Alex Greenshields-Watson
- my co-authors Fabian Spoendlin and @mcagiada.bsky.social
- and our extraordinary P.I. Charlotte Deane!

Have questions or thoughts? Let’s discuss! 🧬

Challenges and compromises: Predicting unbound antibody structures with deep learning Therapeutic antibodies are manufactured, stored and administered in the free state; this makes understanding the unbound form key to designing and imp…

The future of antibody design is bright, and we’re excited to contribute to it! 🌟

Check out the paper for full details!

They also give rise to probabilistic metrics (e.g. conformational likelihoods) that could better reflect state occupancies and outperform current metrics as ranking and filtering criteria.

Plus, generative models open the door to robust, antigen-conditional de novo design. 🚀

7/

We also suggest generative approaches (like diffusion or flow matching) can help!

Here’s why:
• They target conformational distributions directly as the learning objective.
• They sample these distributions efficiently.

6/

We call for:
🧠 More ML-grade unbound data for training predictors,
✅ Better methods to rank/QC structure predictions + estimate uncertainty,
🔄 Improved flexibility/ensemble predictions,
🔬 Carrying multiple conformations into downstream analyses.

5/

In other words, designing better-targeted, more reliable antibodies demands better handling of multiple conformations!

Our paper highlights these challenges, reviews current antibody structure predictors (e.g. AF3, ESM3, ABodyBuilder3), and proposes key directions for progress.

4/

Worse, this conformational heterogeneity directly affects antibody function!
• Entropic contributions influence binding and affinity (ΔG=ΔH–TΔS).
• Flexibility impacts many therapeutic traits.
• Flexibility could even be exploited—e.g., pH-sensitive antibodies that “switch on” inside tumours! 🧪

3/

Therapeutic antibodies are manufactured, stored, and administered in their free (unbound) state.

So predicting that conformation is crucial! It’s also hard:
1️⃣ Most antibody structures in the PDB are bound forms, leaving little unbound data.
2️⃣ CDR loops are flexible—literal moving targets!

2/

Challenges and compromises: Predicting unbound antibody structures with deep learning Therapeutic antibodies are manufactured, stored and administered in the free state; this makes understanding the unbound form key to designing and imp…

It’s an exciting time in protein design! 🧬✨ But much of the therapeutic potential—especially for antibodies—remains untapped. Why? 🤔

Antibodies seem like ideal candidates for design! 💉
Here’s a quick thread summarising our new review paper on the state of antibody structure prediction. 👇

1/