Antibody Steering
I’m testing what post-training does to an antibody model that edits a binding loop, asking which objectives improve ranking, which collapse the model’s samples while the headline number still rises, and what it takes to get the gain without the collapse.
Post-trainingwhat changes the model
HER2 binding library
high>low
IPO learns from the pairsStarting p-IgGenReplay keeps its samples probable
Generationantibody sequence only
Fixed heavy-chain prefix…YYCSR
Post-trained
p-IgGenIPO + replay
p-IgGenIPO + replay
HCDR310 chosen residues
Likelihood gate stops a run if parent-favoured validation sequences become too unlikely.
Rank measured variants · sample new loops · study residue combinations