Summary
Antibody-specific protein language models outperform generic PLMs on intrafamily but not general thermostability prediction (1). A version of ProGen specifically trained on antibody sequences outperform generic ProGen models on intra-family thermostability prediction. On inter-family prediction, they are bested by ESM-IF (see Structure-based methods outperform sequence-based methods on protein stability prediction of point mutants, but not full sequences).
Details
One related observation (unpublished as of 19 April 2026) is that the mean-pooled CDRH3 embeddings learned by generic LMs, but not antibody LMs, are basically meaningless insofar as they match those of scrambled CDRH3 sequences with the same framework. A separate theory is that this could be because antibodies are separated by V-gene by antibody LMs.
Figures
Figure 5 from (1)
See also
- Generic PLMs outperform antibody-specific PLMs on zero-shot predictions of affinity changes
- Protein language models are better zero-shot predictors for ranking closely related sequences than distantly related sequences