I've been posting some free GPU-bound tool servers over the last couple of weeks, and I've got another one to share today: IgGM (paper here), an antibody design tool from Tencent's AI for Science lab. You can find it at iggm.biobench.ai

As with the other tools, I've got no affiliation with the authors, I'm just looking to practice web server skills and share some GPU power with the community that would otherwise not be used much. For this tool, however, I've added some functionality on top of it, and if anyone has feedback on said addition I'd love to hear it.

The core IgGM tool can predict an antibody structure based only on a backbone sequence, an antigen PDB file, and an epitope range where you want to target the antibody. However, from my experience in my wet lab days, knowing the target epitope is more than half the battle, and if you wanted to do some in-silico antibody design, you probably want to approach it from an epitope-agnostic way. So, I added a wrapper that treats epitope selection as a bayesian optimization problem, and runs a bunch of IgGM models to try and find regions that can have high-affinity antibodies predicted against them.

The tool I've stood up only asks you to pick an antigen and then designs up to 20 antibodies against it (it'll stop early if it hits on a really high affinity one), predicts their affinity, and then generates a PDB file of the predicted complex. I've found looking at the predicted complex structure to be really important - this is still just predictions, after all, and it can make some apparently really high-affinity antibodies that are plainly interacting with the antigen in non-physiological ways.

As with the other tools, if this is useful to anyone then great, mission already accomplished, but especially where I've actually done a bit more than just wrap this one in a flask server, I'd love to hear if anyone has thoughts on IgGM or antibody prediction in general.