Thread by @octavianEganea

EquiDock (spotlight at ICLR'22) has been recently covered by MIT News and we just made our code & pre-trained models publicly available! 1/

Code: github.com/octavian-ganea…

news.mit.edu/2022/ai-predic…

We aim to do direct prediction of the complex, as opposed to previous computationally expensive methods. 2/

We achieve this by hard coding important geometrical constraints into the deep network: EquiDock's prediction is always the same regardless of the initial placements, orientations, and roles of the two proteins, without the need of data augmentation. 3/

We design Independent SE(3)-equivariant Graph Matching Networks to guarantee the aforementioned invariance. We also believe that this architecture would be generally useful for different types of multi-body interactions. 4/

Finally, EquiDock uses an SE(3)-equivariant multi-head attention mechanism to produce keypoint sets that are representative for the binding pocket locations. And this is enforced using optimal transport because we have to recover the actual point alignment. 5/

We achieve 80-500x speed-ups at comparable or better quality. Speed is crucial for drug-discovery applications such as virtual screening or in silico protein engineering. EquiDock can complement other recent approaches such as surface-based fingerprints (MaSIF @mmbronstein) /e

@mmbronstein Grateful for collaborating with this cool team of researchers: @xinyuan_huang_ @_bunnech @yataobian @BarzilayRegina @arkrause and Tommi !