Interestingly, our submission initially ranked between 300-600 on the competition leaderboard due to in-silico scoring mechanics that didn't normalize for sequence length.
This highlights a crucial point: in-silico scores don't always tell the whole story when evaluating generative methods!

Agreed that the full solution is a challenge! Adding variable length CDR edits expands the search space quite dramatically! Current experiments look quite promising, though!

I‘d be curious to see the results! We’ve been playing around with CDR manipulation (including indels) and some of the ML models actually produce some pretty reasonable results! For diversification this is absolutely the way to go in my opinion. Most of the time binding got worse for us, though.

Also, we will be talking about how our zero shot approach landed first in the #AdaptyvBio protein design competition.
See you there @patrickkidger.bsky.social :)

Agreed! That being said, the models are trained to capture higher-order epistasis, and we generally try to avoid single-mutation datasets during training if possible.

Thank you @klausenhauser.bsky.social ! We are currently working on some experiments for jointly optimizing Kon and Koff separately vs Kd as a singular metric. The former should give more fine-grained control of the trade offs of binding behavior without losing any signal for optimization.