Mechanistic Interpretability of Antibody Language Models Using SAEs

arXiv (Cornell University)

Abstract

Sparse autoencoders (SAEs) are a mechanistic interpretability technique that have been used to provide insight into learned concepts within large protein language models. Here, we employ TopK and Ordered SAEs to investigate autoregressive antibody language models, and steer their generation. We show that TopK SAEs can reveal biologically meaningful latent features, but high feature-concept correlation does not guarantee causal control over generation. In contrast, Ordered SAEs impose a hierarchical structure that reliably identifies steerable features, but at the expense of more complex and less interpretable activation patterns. These findings advance the mechanistic interpretability of domain-specific protein language models and suggest that, while TopK SAEs suffice for mapping latent features to concepts, Ordered SAEs are preferable when precise generative steering is required.

Authors 5

1. Oliver M. Turnbull

   University of Oxford

   Affiliation as printed

   Department of Statistics , University of Oxford , UK

2. Anisha Parsan

   IIT@MIT

   Affiliation as printed

   EECS , MIT , Cambridge MA , USA

   Reticular , San Francisco , USA

3. Nithin Parsan
   Affiliation as printed

   Reticular , San Francisco , USA

4. John Jeongseok Yang
   Affiliation as printed

   Reticular , San Francisco , USA

5. Charlotte Mary Deane