AI Insight
Researchers conducted causal intervention experiments on three different AI protein folding models (ESMFold, OpenFold, and Boltz-1) to understand how they predict protein structures. They discovered that despite having different architectures and training methods, all three models use a shared two-stage computational process: early layers first encode biochemical properties like charge from the amino acid sequence into pairwise representations, while later layers then develop spatial information such as distances and contacts between residues. The study demonstrated these mechanisms causally and showed that pairwise representations can be aligned and swapped between models, suggesting convergent evolution toward a common computational strategy for transforming sequence chemistry into three-dimensional structure.
Why it matters
Understanding the internal mechanisms of protein folding AI could improve model design, interpretability, and reliability for drug discovery and protein engineering applications. The finding that different models converge on similar computational strategies may reveal fundamental principles of how sequence information maps to protein structure.
Understand the Science
arXiv:2602.06020v3 Announce Type: replace
Abstract: How do protein structure prediction models fold proteins? We investigate this question through causal interventions on the folding trunks of ESMFold, OpenFold, and Boltz-1. Across all three models, we find a shared two-stage computational structure. In the first stage, early blocks initialize pairwise biochemical signals: features like charge propagate from sequence into pairwise representations through architecture-specific pathways. In the second stage, late blocks develop pairwise spatial features: distance and contact information accumulate in the pairwise representation. We verify these mechanisms causally by showing that steering charge and distance features induces predictable structural changes. Furthermore, these representations are functionally interchangeable: pairwise states can be linearly aligned and substituted across models. Together, these results suggest that folding trunks with different architectures, inputs, and training procedures converge on a shared representational organization for mapping sequence chemistry into spatial geometry.
Source: Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks