🤖 AI Summary
This study elucidates the intrinsic mechanism by which the ESMFold model achieves protein folding, with a focus on canonical structural motifs such as β-hairpins. By applying counterfactual interventions in the latent space and integrating latent variable analysis with interpretable representation tracing, the work identifies and dissects, for the first time, the causal decision pathway through which ESMFold translates biochemical features into three-dimensional conformations. Two critical computational stages are revealed: an early module that injects residue identity and biochemical properties into pairwise representations, and a late module that constructs spatial information—such as inter-residue distances and contacts—based on these enriched representations. The study not only isolates and validates the decoupled processes of biochemical signal initialization and spatial feature development but also enables strong causal intervention and dynamic tracking of the protein folding mechanism.
📝 Abstract
How do protein structure prediction models fold proteins? We investigate this question by tracing how ESMFold folds a beta hairpin, a prevalent structural motif. Through counterfactual interventions on model latents, we identify two computational stages in the folding trunk. In the first stage, early blocks initialize pairwise biochemical signals: residue identities and associated biochemical features such as charge flow from sequence representations into pairwise representations. In the second stage, late blocks develop pairwise spatial features: distance and contact information accumulate in the pairwise representation. We demonstrate that the mechanisms underlying structural decisions of ESMFold can be localized, traced through interpretable representations, and manipulated with strong causal effects.