I want to predict the folding of a chimeric antigen receptor sequence. These sequences are completely made out of small parts of different proteins. Do you know if AlphaFold can be used for this?
2
-
$\begingroup$ Please clarify your specific problem or provide additional details to highlight exactly what you need. As it's currently written, it's hard to tell exactly what you're asking. $\endgroup$– Community BotJan 18 at 19:01
-
$\begingroup$ From @MaximillianPress's response, the answer is no because a chimera has no history the domains evolving together and that is a central part of the algorithms assumptions. Please do remember to mark an answer as "accepted" when the question has been successfully answered. $\endgroup$– M__ ♦Jan 24 at 20:47
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
I'd consult references about AlphaFold. For example, wikipedia has a "Limitations" page on the tool, though the page doesn't mention anything about synthetic proteins.
The hesitation that I would have is that the tool relies to some extent on coevolutionary models of amino acids to infer interactions at both training and prediction steps, so throwing a synthetic protein at it may confuse it. See e.g. slide 10 here.
In an interview, they acknowledge this rather directly:
Now, if you have a protein which is completely new, which does not have an evolutionary history—a synthetic protein, which has just been designed— then our system would not be as accurate as it would be for existing proteins.
I am not totally clear on whether this is due to learning a latent representation of constraints on similar proteins, or whether the tool attempts to directly find homologs. The paper methods suggest to me that it is the former, but I'm not sure. They describe training on only chunks of a big MSA. If that's the case, I'd guess that they still can use information about the natural chunks of your synthetic protein, and aren't expecting to be able to do a full-length protein alignment.
So I think that you can expect the prediction to be relatively low quality. It is still probably much better than any alternative way of predicting structure!
As always, not a bad idea to post an issue on github asking this question.
-
$\begingroup$ This is a very good point indeed, it will rely on convolution as part of its algorithm and chimeras will not follow this assumption. $\endgroup$– M__ ♦Jan 24 at 20:45