7/n
Also thanks to my wonderful coauthors @igashov.bsky.social , @rne.bsky.social , @mmbronstein.bsky.social and supervisors @pschwllr.bsky.social and Bruno Correia. ❤️

6/n
I’ll be presenting this work at:
🧬 GEMbio workshop: Sun 27th (Hall 4#4)

🔬 AI4Mat workshop: Mon 28th (Topaz Concourse).

If you're around #ICLR2025, let’s chat! 😊

5/n
We also propose a robust evaluation framework:

✅ “Hard” fragment recovery

✅ “Soft” pharmacophoric similarity

This gives a nuanced view of what the model learns – and shows improvements over docking-based screening baselines.

4/n
This means:

🔹 You can flexibly explore new fragment libraries

🔹 No retraining required

🔹 Outputs stay valid & structure-aware

🔹 More expressive than vanilla virtual screening
All in one unified latent space ✨

3/n
We then extended this to a generative flow matching framework:

🧠 It learns distributions over fragment latents & spatial arrangements

🧪 Conditioned directly on protein surfaces

✅ No decoder needed

✅ Chemically realistic by construction

2/n
💡 Fragment encoder:

We first train a protein–fragment encoder with contrastive loss to map both fragments and protein surfaces into a shared latent space.
It captures interaction-relevant features, which can be used directly for fast virtual screening 🚀.

1/n
Fragment-based design = build better drugs by combining small fragment that each have key interactions.

But:

❗Fragments bind weakly

❗Standard screening is inefficient

So we built a contrastive learning model to learn how fragments interact with protein pockets. 🧬