Currently we are building a sequence based deep-learning model to predict binding affinity between antibody and antigen. For this we are training a sequence based model with AlphaSeq Antibody dataset (https://github.com/mit-ll/AlphaSeq_Antibody_Dataset). But in this dataset they haven't provided the sequence for SARS-CoV-2 virus.

So I read the paper and there they have mentioned they have tested against peptide in the HR2 region of the SARS-CoV-2 spike protein and they have given the sequence as "PDVDLGDISGINAS". (https://www.nature.com/articles/s41597-022-01779-4#Sec3)

But this sequence is very small. I'd like a larger protein sequence that includes non-epitope portions so that the prediction is more meaningful (e.g. predicts binding to the actual epitope when provided a sequence which includes sequence beyond the epitope). I searched online and only found a covid genome sequence which isn't very helpful.

Searching the RCSB PDB website for Spike protein HR2 gave two SARS-CoV-2 results (RCSB PDB website).

One of them had sequence like the following:

>8CZI_1|Chains A, B, C|Scaffolded Spike protein S2' HR1|Nostoc punctiforme (strain ATCC 29133 / PCC 73102) (63737)

>8CZI_2|Chains D, E, F|Spike protein S2' HR2|Severe acute respiratory syndrome coronavirus 2 (2697049)

So I am not sure which one is the correct sequence for our task. Can someone provide some insights regarding this matter?

Also why are there so many sequences popping up when I search for SARS-CoV-2? Is there a version that can be considered as the root from which other versions are derived?

  • $\begingroup$ Can you please replace the image with the actual sequence? The image is hard to read, impossible to copy from and just makes understanding your question harder. You can use the {} button to format it as code. $\endgroup$
    – terdon
    Feb 27 at 9:18
  • $\begingroup$ Is there a reason you don't just use the peptide sequence that they indicate as an antigen? if you're trying to replicate their results then why do you need a different sequence? Note that your RefSeq COVID link is to the genome (DNA) not protein sequence. $\endgroup$ Feb 27 at 16:52
  • 1
    $\begingroup$ @MaximilianPress Yeah that's a valid question. Actually I wanted to input a bigger sequence to our model rather than the specific part of the whole sequence to which it actually binds, so that our machine learning model learns to guess the correct region from the sequence as well as the correct binding affinity. $\endgroup$
    – Krishna
    Feb 27 at 19:24
  • 1
    $\begingroup$ @MaximilianPress thanks for indicating that's a DNA $\endgroup$
    – Krishna
    Feb 27 at 19:24
  • $\begingroup$ Thanks for explanation, I've edited your question to include this. $\endgroup$ Feb 27 at 19:57

1 Answer 1


PDB has a whole page talking about various slices of the covid spike protein. I'd strongly recommend reading that for a better understanding of why people choose different pieces to represent. In that page they include several different solved structures of the protein that may be helpful for your benchmarking.

In terms of the "canonical" sequence, you could look at a reference genome sequence. Here for example is the EMBL reference genome for covid in GenBank format, which includes the protein sequences. I believe in that the spike protein has ID QHD43416.

You can additionally look at NCBI, which has a slightly different version of the reference (though they should be functionally identical, with just slightly different identifiers and annotations).

Assuming that you want the ancestral initial isolate of covid, you can find a similar GenBank file here from NCBI.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.