BPS2026 – Probabilistic deep learning enables the data-driven design of disordered proteins

Intrinsically disordered proteins and regions (collectively IDRs) lack a stable three-dimensional structure under physiological conditions, instead forming dynamic ensembles of conformations. Despite their heterogeneity, they are ubiquitous across all life and play essential roles in signaling, regulation, and biomolecular interactions. However, the absence of a well-defined fold presents significant challenges in protein design campaigns. Traditional structure-based approaches, which rely on tertiary structures and evolutionary conservation, are ill-suited for IDR design. We present RAVEN, a disorder-focused framework built around a diffusion language model that operates directly on amino-acid sequences, learning disorder-specific sequence representations and supporting conditioning on folded-domain context when present. This context conditioning captures the influence of adjacent structured domains on disordered regions, enabling context-aware sequence design and variant-effect prediction within IDRs. We showcase how RAVEN can be used to support hypothesis generation and guide experimental design. Our tools are easy-to-use, broadly accessible, and designed to enable anyone to elucidate likely functional roles of their favorite disordered protein and protein regions. This work aims to educate the community on the capabilities of these tools and offer practical and conceptual advice on how they may be best administered, with explicit consideration of caveats and limitations.