AMix-1: A Pathway to Test-Time Scalable Protein Foundation Model

Researchers introduce AMix-1, a 1.7-billion parameter protein foundation model that uses Bayesian Flow Networks to advance computational protein design and engineering. The model demonstrates predictable scaling laws, in-context learning capabilities, and test-time scaling algorithms that enable the design of protein variants with up to 50x improved activity, establishing a framework for lab-in-the-loop protein engineering.

AMix-1 represents a significant advancement in applying large-scale foundation models to biological challenges. The research demonstrates that protein design can benefit from systematic scaling approaches similar to those in large language models, with predictable performance improvements as model capacity increases. The integration of multiple sequence alignment-based in-context learning enables the model to recognize evolutionary patterns and generate functionally coherent proteins without task-specific fine-tuning.

This work builds on growing recognition that foundation models can capture domain-specific patterns across diverse applications beyond natural language. The use of test-time scaling through directed evolution algorithms is particularly noteworthy, showing that inference-time computation can substantially improve predictions when paired with verification budgets. This approach aligns with broader trends in AI where scaling computation at inference time delivers performance gains comparable to training larger models.

For the biotechnology and synthetic biology industries, AMix-1 offers practical value by reducing design cycles and computational costs for protein engineering. The demonstrated 50x activity improvement in the AmeR variant suggests the model can identify non-obvious optimizations that exceed traditional rational design approaches. This has implications for drug discovery, enzyme engineering, and industrial biotech applications where computational screening can accelerate development timelines.

The framework's emphasis on test-time scaling introduces interesting considerations around deployment costs and optimization strategies. Future work will likely explore how these approaches scale to larger models and whether similar principles apply to other biomolecular design challenges including RNA and small molecule optimization.

• →AMix-1 achieves predictable scaling laws for protein design, with a 1.7B parameter model demonstrating emergent structural understanding capabilities.
• →Multiple sequence alignment-based in-context learning unifies protein design without task-specific fine-tuning, recognizing deep evolutionary signals.
• →Demonstrated 50x activity improvement in AmeR protein variant validates practical applicability to real protein engineering challenges.
• →Test-time scaling algorithm enables iterative performance gains proportional to verification budgets, enabling lab-in-the-loop optimization.
• →Foundation model approach suggests broader applicability to biomolecular design beyond proteins, potentially transforming drug discovery and synthetic biology workflows.