AI‑Driven De Novo Design of Small‑Molecule Binding Proteins Selective for Cortisol

Background and Challenge

Designing small‑molecule binding proteins that combine high affinity with high specificity is a long‑standing obstacle for biosensing and molecular switches. Traditional approaches rely on screening or modifying natural proteins, limiting generality and scalability.

AI‑Driven Backbone Generation

The researchers first generated 1,615 NTF2‑like backbones using a family‑level hallucination method. These backbones were re‑designed with ProteinMPNN and screened by AlphaFold , yielding 3,230 folded backbones (Collection 2). A parallel Rosetta‑parameterized pipeline produced 6,838 backbones (Collection 3).

After filtering, the team obtained encoded oligonucleotides for thousands of designs, e.g., 630 HCY binders, 1,661 ROC binders, 16,276 WRF binders, 9,024 APX binders, 19,390 IRI binders, and 7,573 OHP binders.

Design of a Diverse NTF2 Protein Family

NTF2 folds consist of three α‑helices and a curved six‑strand β‑sheet that create a large internal pocket. The goal was to engineer a family with varied pocket geometries while keeping the loop region minimal for modularity.

Over 10,000 designs were docked with six chemically distinct ligands using RIFdock : cortisol (HCY), warfarin (WRF), rocuronium bromide (ROC), apixaban (APX), SN‑38 (IRI), and 17‑α‑hydroxyprogesterone (OHP).

Method 1: RIFdock to HBNets

HCY, WRF, ROC, APX, and IRI were docked into Collection 1 backbones, requiring at least one hydrogen‑bond network (HBNet) mediated interaction. Subsequent Rosetta sequence design was biased by a position‑specific scoring matrix derived from the NTF2 family.

Method 2: Unrestricted RIFdock

OHP, APX, and IRI were placed into Collections 2 and 3 using unconstrained RIFdock, followed by sequence design with LigandMPNN , a ProteinMPNN variant trained on protein‑ligand complexes.

Designs were evaluated by Rosetta for hydrogen‑bond count, binding energy (ddG), and contact molecular surface (CMS). For Method 2, single‑sequence AlphaFold predictions were also used to ensure correct fold and binding‑site recreation.

Experimental Validation

Structural Characterization

Crystal structures of cortisol‑binding protein hcy129 and apixaban‑binding protein apx1049 were solved at 1.5 Å and 2.1 Å, respectively. hcy129 showed a Cα RMSD of 1.1 Å to the design model, with key hydrogen bonds matching the prediction. apx1049 displayed a Cα RMSD of 0.6 Å and reproduced the designed hydrogen‑bond and π‑π stacking interactions.

Specificity Assessment

Six designed binders were tested against their cognate ligands and against bovine serum albumin (non‑specific control). High‑affinity binders such as hcy129.1, iri807.1, and apx1049 exhibited strong specificity, while BSA showed negligible binding. For the hydrophobic ligand warfarin, the designed binder wrf1071 (KD ≈ 1.1 µM) performed similarly to BSA (KD ≈ 5.0 µM), highlighting challenges with non‑polar targets.

Biosensor Construction

To create a cortisol sensor, the team performed single‑site saturation mutagenesis (SSM) on hcy129, identified beneficial mutations, and built a combinatorial library screened by yeast display. The best variant, hcy129.1, achieved a KD of 68 nM (31‑fold improvement). Structural analysis linked the affinity gain to enhanced hydrophobic contacts.

Using the optimized binder, a chemically‑induced dimerization (CID) system was engineered: hcy129.1 was fused to one half of a NanoBiT luciferase pair, while a mini‑protein (miniH11) was designed to bind both hcy129.1 and cortisol, forming a ternary complex only in the presence of cortisol. The sensor displayed an EC50 of ~72 nM, matching the binding affinity, and showed minimal background signal without cortisol.

Conclusion

The study demonstrates that AI‑driven de novo design using an NTF2 scaffold can produce diverse small‑molecule binding proteins and transform them into functional biosensors. Structural and biochemical data confirm atomic‑level accuracy, while specificity assays reveal both successes and remaining challenges for hydrophobic ligands. This workflow expands the frontier of protein engineering from natural‑protein modification to on‑demand creation of programmable sensing modules.

Reference: "Small‑molecule binding and sensing with a designed protein family", Nature Communications (2026). DOI: https://www.nature.com/articles/s41467-026-70953-8