search

google-deepmind / alphafold3

AlphaFold 3 inference pipeline.
Other
4.7k stars 495 forks source link

Were all water molecules discarded before AF3 training? #58

Closed smg3d closed 2 hours ago

smg3d commented 4 hours ago

Contrary to the AlphaFold3 server, we now can include water molecules as ligands. Using ccdCodes works really well (and probably water smiles work too, but I have not tried) :

"ccdCodes": ["HOH"]

My first tests on calcium-binding proteins with both calcium and water as ligands, without binding constraints, are actually pretty good. Moreover, according to the Alphafold paper, section 2.1 of Supplementary information :

To simplify later code, the parser performs some basic structure cleanup... waters are removed ...

So my question : Were all water molecules discarded before AF3 training?

joshabramson commented 2 hours ago

Yes, all HOH and DOD instances were excluded from training.

However other similar molecules are included in training, and the model can handle general ligands, so HOH may give reasonable output.