Calculating perplexity for a list of sequences given a single structure

[rtmag]

Hello, really great project! I have a question:

Given an antibody pdb structure, would it be possible to calculate the perplexity for a list of paired heavy and light sequences?

In other words, I'd like to compare a list of antibody sequences to see which has the best fit for the structure.

Thanks a lot in advance!!

• Roberto

[mhoie]

HI rtmag,

If I understand you correctly, you could calculate the average log likelihood of sequences fitting into the given structure.

E.g. first align your sequences, then calculate the average score for each residue in the output CSV: https://github.com/oxpig/AntiFold/blob/master/output/single_pdb/6y1l_imgt.csv

The original sequence should score highly, while random antibody sequences should score poorly (for CDRs at least).

You can convert this to a "pseudo-perplexity" like this

# https://huggingface.co/blog/AmelieSchreiber/mutation-scoring
# Compute and return the pseudo-perplexity
pppl = torch.exp(-avg_log_likelihood)

[rtmag]

Awesome Magnus, thanks for your input, will test it out :)!