Predicting Variant Structures with Templates

[brucejwittmann]

I have a situation where I have one wild-type sequence and a few thousand variants of that wild-type sequence that I've built with a generative model. To speed up inference, I've pre-computed alignments (in a3m format) and templates (in hhr format) for my wild-type sequence. Then, I can just build one new alignment per variant by replacing the query sequence in the wild-type a3m files with the variant sequence, bypassing the need to run all of my variants through mmseqs.

My question is, do I need to modify the hhr file at all to use the precomputed templates with my variants? Obviously, these files contain alignment data for the wild-type sequence, but is that information used? In other words, is it just the list of hits at the top of the file that is important for openfold, or does the alignment data matter too? If it's just the list of hits, then I can use the same template files for each of my variants; otherwise, I'll need to update the alignments in this file as well.