TL;DR
ProGenMech is a new mechanistic interpretability framework designed to peer inside generative, autoregressive protein language models (pLMs). By extending cross-layer transcoders to a sparse Mixture-of-Experts architecture, it successfully extracts compact, biologically meaningful AI "circuits" responsible for generating sequences and predicting protein fitness.
What Problem Does This Paper Address?
Generative protein language models (pLMs) have become incredibly powerful tools in computational biology, capable of designing entirely novel proteins and predicting how mutations affect protein fitness. However, these models operate as massive, multi-billion-parameter "black boxes."
While recent mechanistic interpretability research has used tools like Sparse Autoencoders (SAEs) to unpack representation-learning pLMs (like ESM2), these methods fail to capture the step-by-step computation required for autoregressive generation (generating sequences token-by-token from left to right). Furthermore, modern generative pLMs use complex architectures like sparse Mixture-of-Experts (MoE) and dual-task training (causal generation combined with span infilling), which standard interpretability tools aren't equipped to handle.
Existing per-layer transcoders analyze each transformer layer in isolation, missing how computational context accumulates across the depth of the network.
Methods
The authors introduce ProGenMech, a framework that adapts Cross-Layer Transcoders (CLTs) to the sparse MoE architecture of ProGen3.
Key methodological innovations include:
- Global Context Tracking: Unlike per-layer approaches, CLTs reconstruct each layer's output using sparse latent variables from all preceding layers. This allows the framework to faithfully capture how information builds up across the network.
- Unified MoE Approximation: Instead of trying to untangle individual experts within a layer, ProGenMech treats the entire MoE block as a single functional mapping, discovering a unified, sparse latent space for the experts.
- Dual-Objective Training: The transcoders are trained on a hybrid mix of causal language modeling (left-to-right generation) and generalized language modeling (span infilling) to match ProGen3's dual capabilities.
- Zero-Shot Circuit Discovery: The authors developed an iterative greedy search algorithm based on gradient attribution. This isolates the minimal subset of features (a "circuit") that drives specific model behaviors, such as sequence generation or fitness scoring.
Datasets
- Training: The CLTs were trained on 5 million protein sequences randomly sampled from UniRef50.
- Generation Evaluation: Circuit discovery for generation tasks was benchmarked using structurally diverse sequences from the Swiss-Prot database (clustered at 30% sequence identity).
- Fitness Prediction Evaluation: The zero-shot fitness circuits were evaluated using 8 Deep Mutational Scanning (DMS) assays from ProteinGym (including targets like GRB2_HUMAN, GFP, and KRAS), which measure properties like binding affinity, fluorescence, and stability.
Key Insights and Ideas
- Hierarchical Logic: ProGen3 processes biological information hierarchically. Early layers detect basic biochemical building blocks (like individual amino acids), middle layers compose these into complex functional motifs (like catalytic loops), and later layers finalize next-token predictions or score fitness.
- Extreme Computational Compression: Sparser is smarter. ProGenMech can recover roughly 95% of the original model's zero-shot fitness prediction performance using a circuit containing just 0.6% of the total latent space.
- The Generation-Scoring Discrepancy: In smaller generative models (like the 112M parameter ProGen3 variant tested), there is a fundamental gap between scoring and generation. The model is highly capable of scoring realistic fitness landscapes, but its generative component easily falls into low-complexity, repetitive sequence traps.
Key Results
- Superior Distribution Recovery: In causal generation tasks, ProGenMech achieved an average Negative Log-Likelihood (NLL) of 2.50, significantly outperforming the Per-Layer Transcoder (PLT) baseline (2.57) and closely tracking the base ProGen3 model (2.00).
- Effective Fitness Circuits: For zero-shot fitness estimation, ProGenMech circuits achieved a Spearman correlation of 0.23 (recovering 80% of the full model's performance), beating the PLT baseline (0.22).
- Biological Validation: The framework successfully traced known biological mechanisms. For example, it mapped out the circuits tracking the highly conserved HRD and DFG catalytic motifs in kinases. In the GRB2 fitness landscape, it correctly identified how specific high-fitness mutations increased structural stability, while low-fitness mutations disrupted critical binding sites.
Limitations and Future Work
- MoE Routing Abstracted: By treating the MoE layer as a single block, ProGenMech abstracts away the internal routing logic. Future work could benefit from expert-specific latent spaces.
- Quadratic Scaling Overhead: Because CLTs rely on cross-layer connections, the number of decoder matrices scales quadratically ($O(L^2)$) with the number of layers, creating a parameter overhead challenge for massive models.
- Steerability Constraints: The authors attempted to "steer" the model to generate higher-fitness proteins by intervening on the discovered circuits, but the output quality was bottlenecked by the base 112M model's tendency to generate low-complexity regions. Scaling ProGenMech to larger ProGen3 variants (219M or 339M parameters) is a crucial next step to unlock true steerable design.
Summary updated