Modification of residue leads to incorrect bonding of backbone carbon atom

[ackbar03]

Hi,

Thanks for your tool for residue modification on boltz1!

I'm having some trouble using it for one of my modifications however.

I am trying to N-Methylate the backbone nitrogen of ASN, i.e. from:
Asparagine(https://pubchem.ncbi.nlm.nih.gov/compound/6267)
to
N2-Methyl-L-asparagine (https://pubchem.ncbi.nlm.nih.gov/substance/461453433)

There is a N-Methylated variant of ASN in the CCD
MEN (https://www.rcsb.org/ligand/MEN)

But it is on the wrong nitrogen.

I used your tool to extend the CCD using the SMILES for N2-Methyl-L-asparagine. However, after prediction, the residue is incorrectly bonded. It seems as if it is recognizing the wrong Carbon atom as the backbone carbon.

(Incorrect carbon from N-Methylation in Red. The correct backbone carbon in blue)

At first, I thought it may have been due to the incorrect mapping and numbering of the residues, but after checking it seems not to be the case.

(After renumbering)

(Output of properties)

0 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'N', 'alt_name': 'N', 'leaving_atom': False}
1 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, '_ChiralityPossible': 1, '_CIPCode': 'S', 'name': 'CA', 'alt_name': 'CA', 'leaving_atom': False}
2 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'C', 'alt_name': 'C', 'leaving_atom': False}
3 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'O', 'alt_name': 'O', 'leaving_atom': False}
4 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'CB', 'alt_name': 'CB', 'leaving_atom': False}
5 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'CG', 'alt_name': 'CG', 'leaving_atom': False}
6 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'OD1', 'alt_name': 'OD1', 'leaving_atom': False}
7 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'ND2', 'alt_name': 'ND2', 'leaving_atom': False}
8 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'OXT', 'alt_name': 'OXT', 'leaving_atom': True}
9 {'__computedProps': <rdkit.rdBase._vectSs object at 0x7f7fb6321dc0>, 'name': 'C0', 'alt_name': 'C0', 'leaving_atom': False}

Is this a Boltz model issue, or am I missing something in the modification script?

Thanks

[benf549]

Hm... I'm not sure if Boltz would be able to handle backbone modifications like this since it was trained almost entirely on canonical backbone residues. Are there any examples of backbone N methylation in the CCD that we could use as a reference to figure out if its an issue of the CCD new residue injection script vs Boltz itself?

[ackbar03]

Hey Ben,

Thanks for taking the time to reply.

As per your suggestion, I've taken a look at some of the other backbone n-methylation residues I've used, this seems to be a boltz issue, but it doesn't show up in Chai1

---

For example, N-METHYLCYSTEINE (CCD NCY: https://www.rcsb.org/ligand/NCY) seems to work fine in Chai1.

But not so much in Boltz1

The oxygen group at the end of the residue doesn't seem to get processed correctly when forming the amide group?

---

Also, the N-METHYLPHENYLALANINE (CCD MEA: https://www.rcsb.org/ligand/MEA) produces something that more or less makes sense in Chai1. (The bonding is a bit messy due to the way PyMOL uses atomic distance to determine bonds, but the atoms are in the right configuration)

However, in Boltz1, it seems to have the same problem with the amide bond not being processed correctly.

---

Does this help provide some hints as to how to address the problem? Is it more of a post-processing problem (to get the oxygens at the end of the residue to behave properly) or is it just something wrong with the model?

Also, I know your tool is specifically to work with Boltz1, but do you have any idea how I can add the N2-Methyl-L-asparagine to the CCD directly so I can maybe use it for Chai1? There doesn't seem to be a lot of info online on how to work with the CCD file.

[benf549]

Hm this is interesting, you might want to make an issue on the main Boltz repo for the authors to be aware. Just checking did you clone the version of the Boltz code that removes the OXT atoms before making those predictions? I'm wondering it thats freaking the model out.

This is the link to the pull request, the authors haven't merged it yet
jwohlwend/boltz#93

Very strange that its making the methylated nitrogen tetrahedral....