Unraveling protein conformational plasticity with PROTEUS

Protein conformational plasticity underpins allosteric regulation, fold switching, and post-translational modification accessibility, yet no existing method can probe this property at the proteome scale without simulation. Here we show that SimpleFold, a flow-matching protein structure predictor, implicitly encodes conformational plasticity in its internal representations. By comparing per-residue embeddings between the sequence-only regime and the structure-converged regime of the denoising trajectory, we define a zero-shot conformational plasticity score, PROTEUS (PROtein TrajEctory Uncertainty Score), that requires no experimental dynamics data. PROTEUS correctly orders five independent protein classes spanning the full flexibility spectrum, from rigid de novo designed scaffolds to intrinsically disordered proteins that fold upon binding. Per-residue PROTEUS profiles correlate with atomic fluctuations from 1,290 independent molecular dynamics trajectories, and this signal persists after controlling for structure prediction confidence (pLDDT) and sequence-based disorder predictions. At the protein level, PROTEUS achieves AUROC = 0.77 for fold-switch detection, 0.81 for open/closed state discrimination, and 0.93 for identifying proteins with buried phosphorylation sites. Proteome-wide analysis of 4,188 Escherichia coli K-12 proteins reveals that fimbrial adhesins and the type II secretion machinery rank among the most conformationally plastic functional classes, consistent with the structural demands of chaperone-mediated secretion and receptor engagement, while ribosomal proteins score systematically lower. These results establish that PROTEUS provides unsupervised, proteome-scale probing of structural dynamics directly from a generative model.