Depth of auto-generated MSAs

[max-overath]

The MSAs that got generated for my predictions only contain a couple of sequences.
Is this due to limitations of the MMSeq2 sever or can this be adjusted?

[jwohlwend]

Yeah that d be due to the server. Maybe try this (https://zhanggroup.org/DeepMSA/) and see if you're getting a much deeper MSA? Maybe your sequence just doesn't have many homologues?

[Overathed]

Thanks for the answer! Yes when using the DeepMSA I get a much deeper MSA. However, it is also deeper when using colabfold which should also use a MMSeq2 server if I'm not mistaken?

[jwohlwend]

Would you mind sharing your input config? I can take a look

[heol1]

Is the query sequence a hetero-multimeric protein? If so, I had the same issue.
In ColabFold, it queries MMseqs2 API twice: one for each chain and the other for the "pair" mode.
https://github.com/sokrypton/ColabFold/blob/e2ca9e8f992cd65c986de5b64885d5572d8b8ad9/colabfold/batch.py#L817-L857
In contrast, the current implementation of Boltz, compute_msa, calls the API only once for the "pair" mode.

boltz/src/boltz/main.py

Line 178 in e43f910

msa = run_mmseqs2(list(data.values()), msa_dir, use_pairing=len(data) > 1)

This might the reason why you have a shallow MSA...

[jadolfbr]

So should it do the multiples then? Does that match the benchmarking?

…

On Thu, Nov 21, 2024 at 4:56 PM Lim Heo ***@***.***> wrote: Is the query sequence a hetero-multimeric protein? If so, I had the same issue. In ColabFold, it queries MMseqs2 API twice: one for each chain and the other for the "pair" mode. https://github.com/sokrypton/ColabFold/blob/e2ca9e8f992cd65c986de5b64885d5572d8b8ad9/colabfold/batch.py#L817-L857 In contrast, the current implementation of Boltz, compute_msa, calls the API only once for the "pair" mode. https://github.com/jwohlwend/boltz/blob/e43f9101886ea6c290d6b1e0ada3796f0e798d88/src/boltz/main.py#L178 This might the reason why you have a shallow MSA... — Reply to this email directly, view it on GitHub <#39 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZDHRAN6EUFIP5T432MGG32BZJJVAVCNFSM6AAAAABSHCAH3GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOJSGQYTQOJRGI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[jwohlwend]

Looking into it, will report back

[max-overath]

@heol1 yes exactly it's a hetero multimer. When I run the chains individually I get much deeper MSAs

[max-overath]

@jwohlwend input fasta for reference:

>A|protein
QVQLQESGGGLVQAGGSLRLSCAGSGDALGSYTMGWFRQAPGGGRDLVAQISVDGSSTYHLDSVRGRFTASRDNAKNTVYLEMNSLNSEDTAVYYCAAAPLLRGNYDYWGQGTQVTVSS
>B|protein
IRCFITPDITSKDCPNGHVCYTKTWCDAFCSIRGKRVDLGCAATCPTVKTGVDIQCCSTDNCNPFPTRKRP

[amelie-iska]

I'm having similar issues. I tried this:

version: 1  # Optional, defaults to 1
sequences:
  - protein:
      id: A
      sequence: MGDWSALGRLLDKVQAYSTAGGKVWLSVLFIFRILLLGTAVESAWGDEQSAFVCNTQQPGCENVCYDKSFPISHVRFWVLQIIFVSTPTLLYLAHVFYLMRKEEKLNRKEEELKMVQNEGGNVDMHLKQIEIKKFKYGLEEHGKVKMRGGLLRTYIISILFKSVFEVGFIIIQWYMYGFSLSAIYTCKRDPCPHQVDCFLSRPTEKTIFIWFMLIVSIVSLALNIIELFYVTYKSIKDGIKGKKDPFSATNDAVISGKECGSPKYAYFNGCSSPTAPMSPPGYKLVTGERNPSSCRNYNKQASEQNWANYSAEQNRMGQAGSTISNTHAQPFDFSDEHQNTKKMAPGHEMQPLTILDQRPSSRASSHASSRPRPDDLEI
  - protein:
      id: B
      sequence: MGTFEEVP

with the command:

boltz predict examples/multimer.yaml --recycling_steps 20  --diffusion_samples 10 --use_msa_server

and both MSAs for the individual, as well as the pair, are single sequences only. Using ColabFold, I get much deeper MSAs (and much better predictions).

[paul-goldsmith]

Just adding another voice to this - I'm also finding the auto-generated MSA to be very shallow (single sequence) using two proteins and the --use_msa_server flag.