Anshul Kundaje#1650
@LeoWanPhD @ranomics Theoretically, it shouldn't. All it predicts is: P(binding mode | it binds). Aka. assuming that it DOES bind, how confident are you about the predicted model of how/where it binds.
Users agreed that ipTM and ipSAE metrics fail to predict protein binding affinity because the claim correctly highlights their limitations for de novo filtering.
@LeoWanPhD @ranomics That being said, people have been finding that when the confidence of where and how it binds is low. It's less likely to work experimentally.
@sokrypton @LeoWanPhD @ranomics Right on the money! 🙌 @sokrypton what's the best substitute for ipTM/ipSAE for de novo filtering then? Do we train another ranking predictor on negative decoys? Should we expect this to ever generalize well?
Anshul Kundaje#1650@LeoWanPhD @ranomics Theoretically, it shouldn't. All it predicts is: P(binding mode | it binds). Aka. assuming that it DOES bind, how confident are you about the predicted model of how/where it binds.