Model for membrane proteins

[BohanLi0429]

Hi Kaiyi and other authors,

Congratulations on this paper! The results look very promising, and I cannot wait to apply it to my project.

As I just started to run these codes.

In the extract.py., it seems to require a pre-trained PLM model for generating mutants. I searched the GitHub and found several available PLMs. Which model you would recommend for modifying the membrane proteins? Still ESM-2?

Also, did you guys ever work with membrane proteins, and whether the prediction is still robust?

Looking forward to your reply!

[idmjky]

Hi Bohan,
Thanks for the kind words. Evolve-pro is plug and play in the sense that you can work with any pretrained PLM to generate embeddings for the input AA sequence. Now, in the paper, we tested many pretrained large PLM and found ESM2-15B and ANKH large to work very well across different families of proteins. However, I never tested this on membrane protein specifically, so my intuition is that you can try both ANKH large and ESM2. You can even fine-tune these models with a subset of membrane proteins close to what you want to see if that help. I am hopeful that you will pick up better activity variants in a few rounds if you set up a good assay.