The atom number of RF3 model doesnt match with the atom number of ProteinMPNN

[Forever8341]

Hello there, thank you for the amazing work.

I have a quesiton that the atom number of RF3 model doesnt match with the atom number of ProteinMPNN. Here is my workflow.

First I ran a rfd3 to generate backbones.
rfd3 design out_dir=/media/user/ALL_USERS/hjd/rfd3/results ckpt_path=~/.foundry/checkpoints/rfd3_latest.ckpt inputs=/media/user/ALL_USERS/hjd/rfd3/JOB.json diffusion_batch_size=500 n_batches=4
The json file looks like this:

{
    "E1": {
        "dialect": 2,
        "infer_ori_strategy": "hotspots",
        "input": "/media/user/ALL_USERS/hjd/rfd3/cleaned_A.pdb",
        "contig": "20-100,/0,A150-420",
        "select_hotspots": {
            "A181": "CG2,CG1",
            "A407": "NE1,CZ2",
            "A214": "NH2,NH1",
            "A404": "CD1,CZ",
            "A222": "CB,CG"

        }
       }
}

Then I ran ProteinMPNN to refine the sequence.

mpnn  \
    --structure_path "${cif_file}" \
    --out_directory "${OUT_DIR}" \
    --checkpoint_path "${CHECKPOINT_PATH}" \
    --model_type protein_mpnn \
    --is_legacy_weights True

After that, I saved all the fa files from last step to a file called "sequences_output.json". And ran RF3.
rf3 fold inputs="./sequences_output.json" ckpt_path="/home/jedi/.foundry/checkpoints/rf3_foundry_01_24_latest_remapped.ckpt" out_dir="./rf3_ouput"
The sequences_output.json file looks like this:
{ "name": "E1_3_model_7_b0_d0(1)", "components": [ { "seq": "GWIEGVVLEFVDDDTVLVDDGERVYRVLRSSVENPENARVGSRVRVSTLTAEEVPVVCPGGTCFSVPTL", "chain_id": "A" } ] }, { "name": "E1_3_model_421_b0_d0(2)", "components": [ { "seq": "MEELVEKIKKKLEAEGYKVLKVKVNEDGTVSVVVEKDGKYYELTFDSKGNLLSKEPVRVVVKVPVNGKATYYKCDCGDSEAGVIFEDYTIPGC", "chain_id": "A" } ] }

The CIF files generated by RF3 have more atoms than the CIF files generated by ProteinMPNN and RFD3. The CIF files have the same length between RFD3 and ProteinMPNN. Is it because RF3 used the B Chain from ProteinMPNN as input? So the CIF model of RF3 is larger?

Many thanks!!!

[AndrewKubaney]

Hi @Forever8341,

I have a few follow-up questions! The CIF file generated by RFD3 might have side chain atoms; however the CIF file generated by MPNN should have no side chain atoms (since the sequence was redesigned). Are the extra atoms in the RF3 output found in additional residues/chains or are they the RF3-predicted side chain atoms for the MPNN-designed sequence?

[Forever8341]

Hello @AndrewKubaney ,

Well. The CIF file generated by MPNN has the length and chains as the CIF generated by RFD3. Both RFD3 and MPNN CIF files contain a chain B. I assume the chain B is a part of my target protein. Is that right?

When I check my results, I found that RF3 merged the two chains from MPNN CIF files (So does the fa file generate by the MPNN).

The sequence of CIF files
**RFD3_ChainA (62 aa): DPIIANVLVVNTSPPALATVSGTNVNNGDLAVLTGNTTAAVTANAAGNATSDEAAILSVTPL
RFD3_ChainB (31aa): ITVNGQTVRYKCNCGGSNEGLTTTDKVINNC

MPNN_ChainA (62 aa): ALKTATVTVVNTTPKATATVEGSNVNDGDTAVMVGEDTAKCIADSNGNPCDETAKILSKTPL
MPNN_ChainB (31aa): IPVNGKPVYYKNDCGDTNAGWYTKDYTIPNC

RF3 (93aa): ALKTATVTVVNTTPKATATVEGSNVNDGDTAVMVGEDTAKCIADSNGNPCDETAKILSKTPLIPVNGKPVYYKNDCGDTNAGWYTKDYTIPNC**

FA file gernerated by the MPNN
**>TEST_E2_3_model_9_b0_d0, sequence_recovery=0.5161
ALKTATVTVVNTTPKATATVEGSNVNDGDTAVMVGEDTAKCIADSNGNPCDETAKILSKTPLIPVNGKPVYYKNDCGDTNAGWYTKDYTIPNC
**

Did I do anything wrong?

[Forever8341]

Is is because I didnt apply **"fixed_chains":['B']  ** during the MPNN section?

[AndrewKubaney]

There are a few things going on here!

1. Does your input PDB to RFD3 contain your full target protein? If so, is the chain B output by RFD3 a small section of your target protein? In other words, is RFD3 truncating your target? If this is happening, it is probably non-ideal for the sequence design step, and we would need to follow-up with @Ubiquinone-dot.

2. If your goal is to design a binder to a specific target, it is standard practice to fix the target chains during the MPNN sequence design step (this allows the model to "see" the sequence of the target chain). fixed_chains":['B'] would achieve this (if B is your target chain).

3. There is a bug in MPNN with FASTA writing. MPNN writes the sequence of any fixed and designed chains. Currently, there is no chain-separator character in the FASTA output. We are working to fix this issue soon! In the meantime, I think there are RF3 settings to run folding using the CIF output of MPNN as input.

4. Is your goal to predict only the designed chain? When doing binder design, predicting the binder on its own will give some insight, but it is also worthwhile to predict the binder-target complex!

[Forever8341]

There are a few things going on here!

1. Does your input PDB to RFD3 contain your full target protein? If so, is the chain B output by RFD3 a small section of your target protein? In other words, is RFD3 truncating your target? If this is happening, it is probably non-ideal for the sequence design step, and we would need to follow-up with @Ubiquinone-dot.
2. If your goal is to design a binder to a specific target, it is standard practice to fix the target chains during the MPNN sequence design step (this allows the model to "see" the sequence of the target chain). fixed_chains":['B'] would achieve this (if B is your target chain).
3. There is a bug in MPNN with FASTA writing. MPNN writes the sequence of any fixed and designed chains. Currently, there is no chain-separator character in the FASTA output. We are working to fix this issue soon! In the meantime, I think there are RF3 settings to run folding using the CIF output of MPNN as input.
4. Is your goal to predict only the designed chain? When doing binder design, predicting the binder on its own will give some insight, but it is also worthwhile to predict the binder-target complex!

1. Yes, a full target protein (439 aa) is used as input PBD for RFD3. RFD3 truncated my target. This happened whenever I tried to design proteins for a large target. Sometimes the structure of truncated target, which is outputed by RFD3, would be changed . What other info should I provide to help?

2, 3&4. I got you. Thank you so much for the explaination.

[Ubiquinone-dot]

Judging by your input, your target is A150-420 which is a selection of 270 amino acids long in chain A of the input. Does this explain the unexpected truncation?

[Forever8341]

Yes, ty so much! That explains the unexpected truncation. I found that the structure of truncated target would change after it is processed by rfd3. So is the truncation acceptable for sequence design or not?