Paper List

期刊: ArXiv Preprint
发布日期: 2025-12-05
BioinformaticsComputational Biology

Mechanistic Interpretability of Antibody Language Models Using SAEs

Department of Statistics, University of Oxford, UK | Reticular, San Francisco, USA | Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology (MIT), Cambridge, MA, USA

Rebonto Haque, Oliver Turnbull, Anisha Parsan, Nithin Parsan, John J. Yang, Charlotte M. Deane

30秒速读

IN SHORT: This work addresses the core challenge of achieving both interpretability and controllable generation in domain-specific protein language models, specifically for antibody design.

核心创新

  • Methodology First application of Sparse Autoencoders (SAEs) to interrogate autoregressive antibody-specific language models (p-IgGen), moving beyond general protein language models.
  • Methodology Systematic comparison reveals a key trade-off: TopK SAEs yield highly interpretable, monosemantic features (e.g., for CDR identity with validation accuracy 0.99) but lack causal steerability, while Ordered SAEs provide reliable generative control at the cost of interpretability.
  • Biology Identifies and validates antibody-specific, biologically meaningful latent features, such as CDR identity and germline gene identity (e.g., IGHJ4 prediction with F1 macro score of 0.93), demonstrating the model's learning of immunologically relevant concepts.

主要结论

  • TopK SAEs effectively compress and preserve biological information (CDR identity prediction accuracy 0.99 vs. 0.98 for raw neurons) and yield sparse, interpretable activation patterns localized to specific regions (e.g., CDRH3), overcoming neuron polysemanticity.
  • High feature-concept correlation (e.g., F1 > 0.5 for IGHJ4 latents) does not guarantee causal steerability; steering on TopK-identified IGHJ4 features failed to consistently increase IGHJ4 proportions in generated sequences.
  • Ordered SAEs, with their enforced hierarchical latent structure (via per-index nested grouping and decreasing truncation weights), successfully identify features that enable predictable generative steering, albeit with more complex activation patterns.
研究空白： Prior mechanistic interpretability work on protein language models had not been applied to autoregressive, domain-specific models like antibody LMs, creating a gap in understanding and controlling their internal representations for rational drug design.

摘要: 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 an autoregressive antibody language model, p-IgGen, and steer its 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 an hierarchical structure that reliably identifies steerable features, but at the expense of more complex and less interpretable activation patterns. These findings advance the mecahnistic 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.

代码