Covalent reaction

[xiangwenkai]

Thanks for your excellent work!
Covalent reactions occur with specific protein ligands, does boltz have a method to address this?
For instance, given protein 1AUR (https://www.rcsb.org/structure/1AUR), and ligand PMS:

The expected complex conformation should be:

However, Boltz's prediction resulted in：

Is there a way for me to incorporate covalent reaction details into the model to get better predictions?

[rcossio]

This is related to Issue #65 (about the ring aromaticity).
The extra oxigen could be removed with something similar to issue #94 in the PMS definition on the CCD library.

[benf549]

What input files/method are you using to get this output?

If you know where the ligand attaches covalently and are using the modification syntax in the .yaml, input this could be fixed by dropping leaving atoms as mentioned in #94. If you're using the smiles or CCD to input PMS as a ligand though this might be more difficult to handle.

It might be best to drop leaving atoms regardless of whether the ligand is being used as a modification or not but I believe right now they are not being considered at all... @gcorso any thoughts? I'm happy to make a PR which drops them unconditionally.

[xiangwenkai]

The input yaml file is:

constraints:
- bond:
    atom1: [A, 114, OG]
    atom2: [H, 1, S]
sequences:
- protein:
    id: [A]
    sequence: MTEPLILQPAKPADACVIWLHGLGADRYDFMPVAEALQESLLTTRFVLPQAPTRPVTINGGYEMPSWYDIKAMSPARSISLEELEVSAKMVTDLIEAQKRTGIDASRIFLAGFSQGGAVVFHTAFINWQGPLGGVIALSTYAPTFGDELELSASQQRIPALCLHGQYDDVVQNAMGRSAFEHLKSRGVTVTWQEYPMGHEVLPQEIHDIGAWLAARLG
- ligand:
    ccd: PMS
    id: [H]

And the predict command is:
boltz predict 1AUR.yaml --use_msa_server

[benf549]

If you want a quick fix for the meantime, since you know where you want the ligand to go, you can try using this notebook (https://github.com/benf549/boltz-generalized-covalent-modification). You would just need a smiles string attaching the PMS ligand to the amino acid you're interested in and then clone the branch in #93 to get boltz to drop the extra 'OXT' atom from the backbone.

The most general solution to the leaving group problem might be to check if there is a bond constraint to an atom attached to a leaving group atom and only then remove that leaving group? Not sure what the authors did when training the model but I recall jwohlwend mentioning this headache in his neurips talk.