Looking Inside AI Protein Folding Models (accepted at NeurIPS 2026)

Proteins make life possible. They drive chemical reactions, defend us against infection, and carry signals between cells. Their three-dimensional shapes are central to how they work. Understanding and designing proteins opens routes to new medicines, vaccines, and enzymes that break down plastic. The importance of this field was recognised by the 2024 Nobel Prize in Chemistry, awarded for protein structure prediction and computational protein design.

AI protein folding models have transformed the scale of this research. The AlphaFold database provides more than 200 million predicted protein structures. Yet much of what these models have learned about the relationship between a protein’s sequence and its shape remains hidden inside their computations. Understanding that knowledge could give scientists new ways to investigate the molecular machinery of life.

The study “Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks” helps open this black box. It is among the first to experimentally uncover mechanisms shared by the core components of modern protein folding models. Studying ESMFold, OpenFold, and Boltz-1, the team changed internal signals and produced predictable changes in the resulting structures. This demonstrates how specific information inside the models helps determine their predictions.

The researchers found two broad stages: the models first build information about chemical properties, then about spatial relationships. These findings reveal a shared way of processing protein information across systems with different designs.

The broader opportunity is to turn knowledge learned by AI into explanations that scientists can test. Understanding these models could help reveal which biological principles they have learned, explain where their predictions fail, and guide future protein design. In the longer term, such insights could support the development of proteins with desired functions, from medicines to enzymes for cleaner industrial processes.

The authors are Kevin Lu, Jannik Brinkmann, Stefan Huber, Aaron Mueller, Yonatan Belinkov, David Bau, and Chris Wendler. The paper has been accepted to the NeurIPS 2026 main conference. NeurIPS holds an A* rating, the highest category in the ICORE 2026 conference ranking.

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.