[BUG] Unphysical sidechain rotamers generated for His, Lys, and more

[Astro111000]

Describe the bug
When a theozyme contains atoms until CB, the resulting RFD3 scaffolds somehow generate extremly long bonds between Ca (in the backbone) to CB, which is part of the sidechain explicilty included in the theozyme. Using HBPLUS or classifier-free guidance does not work either. Is there any error in my inputs, or is this a known limitation?

To Reproduce
Please describe any:

The JSON input, theozyme, and CLI inputs are given below :

Theozyme PDB :
ATOM 1 HE2 HIS A 1 -0.984 1.862 -1.867 1.00 0.00 A H
ATOM 2 NE2 HIS A 1 -0.049 1.068 -2.069 1.00 0.00 A N
ATOM 3 CD2 HIS A 1 -0.080 -0.308 -1.976 1.00 0.00 A C
ATOM 4 CE1 HIS A 1 1.212 1.438 -1.879 1.00 0.00 A C
ATOM 5 CG HIS A 1 1.188 -0.771 -1.728 1.00 0.00 A C
ATOM 6 HD2 HIS A 1 -1.003 -0.862 -2.061 1.00 0.00 A H
ATOM 7 HE1 HIS A 1 1.567 2.457 -1.854 1.00 0.00 A H
ATOM 8 HD1 HIS A 1 2.983 0.398 -1.366 1.00 0.00 A H
ATOM 9 ND1 HIS A 1 1.988 0.358 -1.685 1.00 0.00 A N
ATOM 10 CB HIS A 1 1.715 -2.151 -1.498 1.00 0.00 A C
ATOM 11 HB3 HIS A 1 2.498 -2.150 -0.731 1.00 0.00 A H
ATOM 12 HB2 HIS A 1 0.907 -2.802 -1.154 1.00 0.00 A H
ATOM 13 HB1 HIS A 1 2.142 -2.584 -2.412 1.00 0.00 A H
ATOM 14 CG ASP A 2 4.874 -0.535 -0.031 1.00 0.00 A C
ATOM 15 OD1 ASP A 2 4.391 -1.281 0.864 1.00 0.00 A O
ATOM 16 OD2 ASP A 2 4.370 0.582 -0.411 1.00 0.00 A O
ATOM 17 CB ASP A 2 6.151 -0.983 -0.736 1.00 0.00 A C
ATOM 18 HB2 ASP A 2 5.932 -1.199 -1.788 1.00 0.00 A H
ATOM 19 HB1 ASP A 2 6.885 -0.171 -0.724 1.00 0.00 A H
ATOM 20 HB3 ASP A 2 6.568 -1.875 -0.265 1.00 0.00 A H
ATOM 21 HZ1 LYS A 3 1.661 0.475 1.702 1.00 0.00 A H
ATOM 22 HZ2 LYS A 3 3.190 -0.378 1.643 1.00 0.00 A H
ATOM 23 HZ3 LYS A 3 3.134 1.069 1.026 1.00 0.00 A H
ATOM 24 CE LYS A 3 3.066 1.127 3.137 1.00 0.00 A C
ATOM 25 HE1 LYS A 3 4.152 1.148 3.239 1.00 0.00 A H
ATOM 26 HE2 LYS A 3 2.666 2.140 3.222 1.00 0.00 A H
ATOM 27 HE3 LYS A 3 2.636 0.504 3.923 1.00 0.00 A H
ATOM 28 NZ LYS A 3 2.708 0.554 1.819 1.00 0.00 A N
HETATM 29 C1 UNK A 4 -4.449 -0.230 -0.077 1.00 0.00 A C
HETATM 30 C2 UNK A 4 -3.359 0.523 0.343 1.00 0.00 A C
HETATM 31 C3 UNK A 4 -2.168 -0.095 0.784 1.00 0.00 A C
HETATM 32 C4 UNK A 4 -2.107 -1.514 0.792 1.00 0.00 A C
HETATM 33 C5 UNK A 4 -3.205 -2.271 0.364 1.00 0.00 A C
HETATM 34 C6 UNK A 4 -4.365 -1.630 -0.064 1.00 0.00 A C
HETATM 35 H1 UNK A 4 -5.367 0.240 -0.414 1.00 0.00 A H
HETATM 36 H2 UNK A 4 -3.127 -3.355 0.381 1.00 0.00 A H
HETATM 37 O1 UNK A 4 -1.002 -2.172 1.186 1.00 0.00 A O
HETATM 38 H3 UNK A 4 -0.372 -1.450 1.491 1.00 0.00 A H
HETATM 39 O2 UNK A 4 -5.463 -2.321 -0.491 1.00 0.00 A O
HETATM 40 H4 UNK A 4 -5.280 -3.271 -0.425 1.00 0.00 A H
HETATM 41 C7 UNK A 4 -3.443 2.030 0.392 1.00 0.00 A C
HETATM 42 H5 UNK A 4 -4.165 2.410 -0.341 1.00 0.00 A H
HETATM 43 H6 UNK A 4 -3.840 2.326 1.377 1.00 0.00 A H
HETATM 44 C8 UNK A 4 -2.102 2.714 0.266 1.00 0.00 A C
HETATM 45 C9 UNK A 4 -0.989 0.719 1.163 1.00 0.00 A C
HETATM 46 C10 UNK A 4 -1.011 2.100 0.928 1.00 0.00 A C
HETATM 47 H7 UNK A 4 -0.116 2.674 1.148 1.00 0.00 A H
HETATM 48 O3 UNK A 4 0.026 0.105 1.676 1.00 0.00 A O
HETATM 49 H9 UNK A 4 -2.154 3.798 0.326 1.00 0.00 A H
HETATM 50 O4 UNK A 4 -1.815 2.735 -1.530 1.00 0.00 A O
HETATM 51 H11 UNK A 4 -2.657 2.551 -1.980 1.00 0.00 A H

JSON :

rfd3_config = {
"scytalone_theozyme_trp_modified": {
"input": f"{drive_path}/pure_trp_alone_frozen_optim.pdb",
"unindex": "A1,A2,A3,A4",
"length": "200",
"ligand": "UNK",
"plddt_enhanced": True,
"select_fixed_atoms": {
"A1-3": "ALL"
}
}
}

CLI :

rfd3 design
out_dir={name_drive_path}/rfd3_scytalone_trp_modified_try7
inputs=/content/rfd3_scytalone_theozyme_trp_modified_config.json
n_batches=1
diffusion_batch_size=5
ckpt_path={drive_path}/rfd3_checkpoint/rfd3_latest.ckpt
prevalidate_inputs=True

Expected behavior
The Ca-Cb bond should look "normal" i.e not long and with weird dihedral angles, unlike what is seen.

Screenshots
Weird His:

Weird Lys :

Lys also sometimes completely lacks all sidechain atoms, except for the ones included in the theozyme :

As can be seen, the His Ca-Cb bond length and angles are very unphysical. How to fix this?

Edit : Was related to updating my JSON file, since I had accidentally pasted an older version initally. No changes in results.

[rclune]

I wasn't able to easily recreate this issue on a dummy example. Are the long bonds and odd angles in the residues you are holding fixed? Are you only seeing this in a few of your designs or all of them?

If it's only in a few of your designs and the bonds that are distorted are between an atom that is held fixed and one that is not, you might be over-constraining your system or you might just need to filter your designs.

A few things to try/think about:

1. Instead of holding everything in your theozyme residues fixed ("ALL") try just doing the "TIP" atoms or selecting the atoms that are relevant to the catalytic activity. These odd bonds can be coming from over-constraining your system.
2. You are unindexing 4 residues in your example, but only fixing atoms for 3 of them. This might only be the case for the example your provided above. Based on what I've seen using RFD3, if a residue is read in using "unfixed", it is by default completely held fixed.
3. Have you looked at the metrics in your JSON file? If only some of your designs are showing this, you can see if there's a certain metric that's much worse in your poor designs compared to the good ones and filter based on this.