Circuit Tracing in Autoregressive Protein Language Models

2026-06-14 · Source: Machine Learning · Field: Science & Research — Life Sciences & Biology, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

ProGenMech is a novel mechanistic interpretability framework designed for generative protein language models (pLMs), specifically extending cross-layer transcoders (CLTs) to ProGen3, a sparse Mixture-of-Experts model. ProGen3 is trained for both causal generation and span infilling. Unlike per-layer methods, CLTs in ProGenMech reconstruct each layer using sparse latent variables from all preceding layers, effectively capturing inter-layer generative computation. The framework also includes a zero-shot circuit discovery component to identify sparse latent circuits responsible for protein generation and fitness prediction. In causal generation and zero-shot fitness estimation, ProGenMech surpasses local transcoder baselines in recovering ProGen3's probability distribution and functional scoring behavior. It also matches the original model's generative distribution in span infilling tasks. The identified circuits reveal biologically meaningful motifs and functional regions associated with conserved sequence patterns and protein fitness landscapes.

Key takeaway

For research scientists developing or applying protein language models, ProGenMech offers a critical framework for understanding the complex mechanisms behind protein generation. You should consider integrating such mechanistic interpretability methods to move beyond black-box predictions, enabling the identification of biologically meaningful motifs and functional regions. This approach can significantly enhance your ability to interpret and steer novel protein design, leading to more predictable and targeted outcomes.

Key insights

ProGenMech offers a mechanistic interpretability framework for generative protein language models, revealing underlying biological circuits.

Principles

• Cross-layer transcoders (CLTs) faithfully recover inter-layer generative computation.
• Sparse latent circuits can be identified for protein generation and fitness prediction.
• Mechanistic interpretability can reveal biologically meaningful motifs in pLMs.

Method

ProGenMech extends cross-layer transcoders (CLTs) to ProGen3, a sparse Mixture-of-Experts model, then uses a zero-shot circuit discovery framework to identify sparse latent circuits.

In practice

• Interpret and steer protein generation in pLMs.
• Identify functional regions associated with protein fitness landscapes.

Topics

• Protein Language Models
• Mechanistic Interpretability
• Cross-Layer Transcoders
• ProGen3
• Protein Generation
• Circuit Discovery

Best for: AI Scientist, Research Scientist

Related on AIssential

• I built a tool that shows you what GPT-2 is "thinking" in real-time as it generates 3D graph of concept activations per token [R] — AXON visualizes LLM internal states via real-time 3D graphs of SAE feature activations, revealing pre-token generation dynamics.
• Towards the explainability of protein language models — XAI is crucial for understanding and advancing protein language models, which currently operate as black boxes.
• Scalable Circuit Learning for Interpreting Large Language Models — CircuitLasso offers a scalable, sparse linear regression approach to interpret LLMs by efficiently mapping SAE features and their influence on predictions.
• Hierarchical Latent Structures in Data Generation Process Unify Mechanistic Phenomena across Scale — Hierarchical data generation is key to understanding emergent mechanistic phenomena in Transformer language models.
• Nobody Knows What’s Inside a Neural Network. A Few People Are Trying to Find Out. — A small group of researchers is undertaking "biology on language models," dissecting them like organisms to uncover their internal mechanisms, a process revealing unexpected findings.
• Temporal Functional Circuits: From Spline Plots to Faithful Explanations in KAN Forecasting — Temporal Functional Circuits enable mechanistic explanations in KAN-based time-series forecasting by transforming latent edge functions into explicit, temporally grounded insights.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.