Deep-learning-designed miniprotein Denasin1 cuts stroke infarct volume in mice from 19.5% to 8.6%

A deep-learning-designed miniprotein reduces infarct volume in an ischemic stroke model, exceeding the established venom-derived hASIC1a inhibitor. Prior hASIC1a candidates included small molecules, the venom peptides PcTx1 and Hi1a, and the antibody ASC06-IgG1, limited respectively by insufficient selectivity or potency, pH-dependent pharmacology with multiple reducible disulfide bonds that hinder large-scale production, and high production cost with restricted tissue penetration. In a mouse transient middle cerebral artery occlusion model with intrathecal dosing 15 minutes after reperfusion and TTC quantification at 24 hours, infarct volume was 8.6 ± 5.4% (n = 16) for Denasin1, 19.5 ± 9% (n = 16) for aCSF vehicle and 13.1 ± 6.6% (n = 16) for PcTx1, with Denasin1 significantly lower than PcTx1.

The candidates were not screened from a compound library; a generative model first designed miniproteins around the hydrophobic hotspot F352 in the hASIC1a extracellular acidic pocket, and electrophysiology then judged function. Conventional high-throughput screening must search libraries of millions of molecules for a lead, whereas here only 76 designs were expressed and purified in the first round. The pipeline generated backbones with RFdiffusion under a fixed structure and hotspot constraints, assigned six sequences per backbone with ProteinMPNN, filtered with AlphaFold2 at pLDDT > 85 and interface pAE < 14 Å, produced about 12,000 backbones and 70,000 sequences, and retained 76 candidates; in the single-dose automated patch clamp screen, Denasin1 and Denasin2 inhibited more than 50%.

Denasin1 is largely unstructured in solution and folds into the designed helix-turn-helix conformation only on contacting the channel. Circular dichroism showed helical content below 10%, whereas the companion design Denasin2 retained more than 95% helicity even after heating to 95 °C. Single-particle cryo-EM of the hASIC1a-Denasin1 complex reconstituted in MSP1D1 nanodiscs yielded a 2.7 Å reconstruction of a single protomer (C1) and a 2.73 Å reconstruction of the trimer (C3), with density showing the predicted helix-turn-helix fold and an interface centered on an edge-to-face aromatic interaction between F352 and Denasin1 Y36.

The demonstrated scope is hASIC1a inhibition and subtype selectivity in automated patch clamp on HEK cells and two-electrode voltage clamp in Xenopus oocytes, the bound structure solved in MSP1D1 nanodiscs, and a transient middle cerebral artery occlusion model in adult male C57BL/6J mice with 30-minute occlusion and intrathecal dosing; the authors state that intrathecal administration was a direct route for proof of principle and not a clinical delivery strategy.

The authors note that computational scores cannot be assumed to translate into functional modulation or in vivo efficacy, and Denasin1 had one of the least favorable pAE scores among tested designs; the first-round hit rate was about 3% and rose to 32% after scaffold enrichment in the second round; Denasin1 and Denasin2 bind the same region yet differ functionally, which the authors attribute to the pH dependence and allosteric coupling of ASIC gating; pharmacokinetics, plasma and central nervous system half-life, stability in biofluids, toxicity and blood-brain barrier penetration remain to be studied.