A direct comparison with Boltz-2 hasn’t been done yet, but it would be interesting to see one between co-folding and the classical/hybrid docking benchmarks!

is this how small molecules bind?? 😼

All results, code (MIT License), and data are open and available:

📄 Paper: arxiv.org/abs/2506.20043
📦 Data: zenodo.org/records/1573...
💻 Code: github.com/BIMSBbioinfo...

Huge thanks to co-authors @al2na.bsky.social, @borauyar.bsky.social, and Vedran Franke!

We benchmarked PocketVina across four widely used datasets (PDBbind, PoseBusters, Astex, DockGen), and introduce TargetDock-AI — a large-scale benchmark of >500K protein–ligand pairs with activity labels from PubChem.

(5/n)

• Achieves state-of-the-art success rates on physically valid pose prediction
• Works across ligand flexibility levels and diverse, unseen protein targets

(4/n)

PocketVina offers a robust alternative:
• Identifies multiple pocket centers using P2Rank
• Performs GPU-accelerated docking with QuickVina 2-GPU 2.1
• Completes docking + binding affinity prediction in under 1.5 seconds, with no model training

(3/n)

...physically realistic ligand poses — and are not always as efficient or accurate as often claimed. (2/n)

I remember when I first started learning ML—Andrew Ng offered a Coursera course that uses Octave and covers neural networks for image classification with MNIST. You might find it helpful! :)