How AI Is Unlocking the Design of Proteins That Target Disordered Regions

Background

Intrinsically disordered proteins (IDPs) and intrinsically disordered regions (IDRs) make up more than half of the human proteome. Their lack of stable tertiary structure has historically classified them as "undruggable" targets.

Logos: AI‑augmented protein‑design pipeline

David Baker’s laboratory introduced Logos , a workflow that combines physics‑based Rosetta design with deep‑learning diffusion models to generate repeat‑protein scaffolds capable of binding extended peptide motifs found in natural IDRs.

Scaffold generation

Using Rosetta, repeat‑protein backbones were engineered to present extended conformations compatible with two‑residue repeat motifs (LK, RT, YD, PV, GA). Fluorescence‑polarization assays showed nanomolar affinity for LK and PV repeats, weaker binding for RT/YD, and no detectable binding for GA.

Pocket specialization

The binding pockets were refined with the diffusion model RFdiffusion. The redesign expanded the pocket from four to five repeat units, optimized hydrogen‑bond geometry, and diversified hydrophobic contacts to improve complementarity to target peptides.

Pocket assembly

RFdiffusion

was then used to create rigid interfaces between specialized pockets, producing composite scaffolds that can accommodate diverse peptide conformations. From 70 designs covering seven chimeric targets, testing only ten designs per target yielded two‑digit nanomolar affinities for six of the seven targets.

Threading and sequence optimization

Target IDR sequences were threaded onto the scaffold library. Low‑complexity or highly redundant fragments were removed, and the remaining unique segments were optimized with ProteinMPNN. Designs were evaluated by AlphaFold‑2 confidence scores and structural consistency with the predicted models.

Experimental validation

Binding affinities were measured by bio‑layer interferometry (BLI) and fluorescence polarization. For the DYNA_1b1 binder to a dynorphin‑derived peptide, 45 of 48 designs displayed strong binding, with six achieving sub‑100 pM dissociation constants. NMR spectroscopy confirmed that the previously disordered dynorphin peptide adopts an ordered conformation upon binding, matching the computational model.

Cellular assays

GFP‑fused binders targeting the mesothelin juxtamembrane region (MSLN_1b1) selectively bound mesothelin‑expressing HPAC cells but not MCF7 control cells, demonstrating target specificity in a cellular context.

Key quantitative results

Nanomolar binding for LK and PV repeat motifs in fluorescence‑polarization assays.

Two‑digit nanomolar affinities for six of seven chimeric targets after testing only ten designs per target.

Sub‑100 pM Kd for six DYNA_1b1 designs; 45/48 designs showed strong binding.

Specific cell‑surface binding of GFP‑MSLN_1b1 to HPAC cells, no binding to MCF7 cells.

References

Design of intrinsically disordered region binding proteins – Science: https://www.science.org/doi/10.1126/science.adr8063

Intracellular protein editing – Science: https://www.science.org/doi/10.1126/science.adr5499

Advancing protein evolution with inverse folding models – Cell: https://www.cell.com/cell/abstract/S0092-8674(25)00680-4

Illustrative figure