Interpreting and Steering Protein Language Models through Sparse Autoencoders
On the map: Mechanistic interpretability → Interpreting biomedical and science models
Abstract (arXiv)
The rapid advancements in transformer-based language models have revolutionized natural language processing, yet understanding the internal mechanisms of these models remains a significant challenge. This paper explores the application of sparse autoencoders (SAE) to interpret the internal representations of protein language models, specifically focusing on the ESM-2 8M parameter model. By performing a statistical analysis on each latent component's relevance to distinct protein annotations, we identify potential interpretations linked to various protein characteristics, including transmembrane regions, binding sites, and specialized motifs. We then leverage these insights to guide sequence generation, shortlisting the relevant latent components that can steer the model towards desired targets such as zinc finger domains. This work contributes to the emerging field of mechanistic interpretability in biological sequence models, offering new perspectives on model steering for sequence design.
Key facts. Cited by 17 papers in the map (data through 14 Oct 2026). Placed in Interpreting biomedical and science models, one of 13 directions in Mechanistic interpretability. First posted on arXiv 13 Feb 2025.
Builds on
0 of 0 references.
No references here yet.
Built on by
Citing papers per year: 2025: 8; 2026: 9.
17 papers.
Sort:
By area:
- 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycited by 61
- 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Featurescited by 9
- 2025Sparse Autoencoders for Low-N Protein Function Prediction and Designcited by 4
- 2025Resurrecting the Salmon: Rethinking Mechanistic Interpretability with Domain-Specific Sparse Autoencoderscited by 3
- 2026Protein Circuit Tracing via Cross-layer Transcoderscited by 3
- 2025Circuits, Features, and Heuristics in Molecular Transformerscited by 2
- 2026Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunkscited by 2
- 2025Automated Neuron Labelling Enables Generative Steering and Interpretability in Protein Language Modelscited by 1
- 2026GeoSAE: Geometric Prior-Guided Layer-Wise Sparse Autoencoder Annotation of Brain MRI Foundation Modelscited by 1
- 2026Retrieval and competition: how a protein foundation model starts a proteincited by 1
- 2026Mechanistic Interpretability and Causal Feature Steering of Neural Quantum States via Sparse Autoencoderscited by 1
- 2025Quantifying the Role of OpenFold Components in Protein Structure Predictioncited by 0
5 more
- 2025Features Emerge as Discrete States: The First Application of SAEs to 3D Representationscited by 0
- 2026Circuit Tracing in Autoregressive Protein Language Modelscited by 0
- 2026PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Foldingcited by 0
- 2026Interpreting Latent Protein Language Model Features with Geometric Annotationscited by 0
- 2026Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder featurescited by 0
- 2026Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features1 Oct 2026
- 2026Interpreting Latent Protein Language Model Features with Geometric Annotations26 Aug 2026
- 2026Mechanistic Interpretability and Causal Feature Steering of Neural Quantum States via Sparse Autoencoders1 Jul 2026
- 2026PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding25 Jun 2026
- 2026Circuit Tracing in Autoregressive Protein Language Models14 Jun 2026
- 2026Retrieval and competition: how a protein foundation model starts a protein5 May 2026
- 2026GeoSAE: Geometric Prior-Guided Layer-Wise Sparse Autoencoder Annotation of Brain MRI Foundation Models3 May 2026
- 2026Protein Circuit Tracing via Cross-layer Transcoders12 Feb 2026
- 2026Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks5 Feb 2026
- 2025Features Emerge as Discrete States: The First Application of SAEs to 3D Representations12 Dec 2025
- 2025Circuits, Features, and Heuristics in Molecular Transformers10 Dec 2025
- 2025Quantifying the Role of OpenFold Components in Protein Structure Prediction6 Nov 2025
5 more
- 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features1 Oct 2025
- 2025Sparse Autoencoders for Low-N Protein Function Prediction and Design25 Aug 2025
- 2025Resurrecting the Salmon: Rethinking Mechanistic Interpretability with Domain-Specific Sparse Autoencoders12 Aug 2025
- 2025Automated Neuron Labelling Enables Generative Steering and Interpretability in Protein Language Models8 Jul 2025
- 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry3 Mar 2025
- 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycited by 61
- 2025Sparse Autoencoders for Low-N Protein Function Prediction and Designcited by 4
- 2026Protein Circuit Tracing via Cross-layer Transcoderscited by 3
- 2025Circuits, Features, and Heuristics in Molecular Transformerscited by 2
- 2025Automated Neuron Labelling Enables Generative Steering and Interpretability in Protein Language Modelscited by 1
- 2026GeoSAE: Geometric Prior-Guided Layer-Wise Sparse Autoencoder Annotation of Brain MRI Foundation Modelscited by 1
- 2026Mechanistic Interpretability and Causal Feature Steering of Neural Quantum States via Sparse Autoencoderscited by 1
- 2025Features Emerge as Discrete States: The First Application of SAEs to 3D Representationscited by 0
- 2026Circuit Tracing in Autoregressive Protein Language Modelscited by 0
- 2026PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Foldingcited by 0
- 2026Interpreting Latent Protein Language Model Features with Geometric Annotationscited by 0
- 2026Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder featurescited by 0
- 2026Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features1 Oct 2026
- 2026Interpreting Latent Protein Language Model Features with Geometric Annotations26 Aug 2026
- 2026Mechanistic Interpretability and Causal Feature Steering of Neural Quantum States via Sparse Autoencoders1 Jul 2026
- 2026PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding25 Jun 2026
- 2026Circuit Tracing in Autoregressive Protein Language Models14 Jun 2026
- 2026GeoSAE: Geometric Prior-Guided Layer-Wise Sparse Autoencoder Annotation of Brain MRI Foundation Models3 May 2026
- 2026Protein Circuit Tracing via Cross-layer Transcoders12 Feb 2026
- 2025Features Emerge as Discrete States: The First Application of SAEs to 3D Representations12 Dec 2025
- 2025Circuits, Features, and Heuristics in Molecular Transformers10 Dec 2025
- 2025Sparse Autoencoders for Low-N Protein Function Prediction and Design25 Aug 2025
- 2025Automated Neuron Labelling Enables Generative Steering and Interpretability in Protein Language Models8 Jul 2025
- 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry3 Mar 2025
Similar, not cited
Close in wording, no citation either way
- 2024InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoders
- 2025Towards Interpretable Protein Structure Prediction with Sparse Autoencoders
- 2025ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders
- 2025Mechanistic Interpretability of Antibody Language Models Using SAEs
- 2026Protein Language Models Diverge from Natural Language: Comparative Analysis and Improved Inference
- 2025SAE-RNA: A Sparse Autoencoder Model for Interpreting RNA Language Model Representations
- 2024TourSynbio: A Multi-Modal Large Model and Agent Framework to Bridge Text and Protein Sequences for Protein Engineering
- 2026AMix-2: Establishing Protein as a Native Modality in Large Language Models