Research map / Architectures and efficiency / RNNs and normalization
ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders
Xiangyu Liu, Haodi Lei, Yi Liu, Yang Liu, Wei Hu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| detach W_pred when tying W_def so reconstruction loss does not update it (add detach) (ablation) | ProtSAE · BPO dataset, varying L0 | AUC | Improved |
| add axiom loss from ELEmbeddings domain knowledge (ablation) | ProtSAE · BPO dataset, varying L0 | AUC and Loss Recovered | Improved |
| add semantic bias z_bias (forced activation) (ablation) | ProtSAE · BPO dataset, varying L0 | proportion of predicted activations used in decoding | Improved |
| add learnable rescaling r_pred for defined weights (ablation) | ProtSAE · BPO dataset, varying L0 | AUC and Loss Recovered | Improved |
| add semantic guidance during SAE training (ProtSAE vs TopK SAE) | TopK SAE · probing-based interpretation, avg of three datasets | AUC: .565 -> .797 | Improved |
| add semantic guidance during SAE training (ProtSAE vs TopK SAE) | TopK SAE · probing-based interpretation, avg of three datasets | Fmax: .444 -> .579 | Improved |
| use ProtSAE features instead of linear probe on PLM hidden states | ESM2-15B probing · probing-based interpretation, avg of three datasets | AUPR: AUPR .522 -> .487; Fmax .537 -> .579 | Mixed |
| add semantic guidance during SAE training (ProtSAE vs baseline SAEs) | SAE reconstruction · BPO dataset, varying sparsity | Loss Recovered | No clear effect |
| activate ProtSAE concept features during generation (steering) | ESM2-15B masked reconstruction · seven concepts, 50% masked tokens | TM-score and RMSD | Improved |
Closest papers by meaning
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.89)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.87)
- Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression (NeurIPS 2026 · similarity 0.87)
- Concept Bottleneck Language Models For Protein Design (ICLR 2025 · similarity 0.87)
- ProSST: Protein Language Modeling with Quantized Structure and Disentangled Attention (NeurIPS 2024 · similarity 0.87)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
