ML-designed synuclein aggregation inhibitors

The research record

This is the project's file for this landmark, as written and checked. Sources are listed at the end; anything not yet confirmed is marked "Not yet verified".

Who

Robert Horne, Tuomas Knowles, Michele Vendruscolo and colleagues, University of Cambridge, with Byron Caughey's group at the US National Institute of Allergy and Infectious Diseases and Bernardino Ghetti. Funding: UKRI, the NIAID intramural programme and a US National Institute on Aging grant. Several authors are consultants to, employees of, or founders of WaveBreak Therapeutics (formerly Wren Therapeutics), a company working on synuclein drugs.

What happened

Horne et al. (Nature Chemical Biology 2024;20:634–645, DOI 10.1038/s41589-024-01580-x) targeted secondary nucleation, where existing synuclein fibrils catalyse new clumps. Using the fibril structures in landmark 27, they docked molecules to the fibril surface, then trained a model on 161 data points and tested 184 molecules over three machine-learning rounds, with hit rates rising from 12.5% to 21.4%. The best, I4.05, had a 50% inhibition concentration of 0.52 µM, against extrapolated 36.4 µM for anle-138b, about 70-fold stronger (the abstract says "two orders of magnitude"). It bound fibrils tightly (K_D 13–68 nM), spared amyloid-beta, and blocked seeded aggregation in brain samples from dementia with Lewy bodies and multiple system atrophy.

Why it moved us toward a cure

It shows that drug design aimed at the aggregate surface can be sped up, a route to a small-molecule anti-aggregant after earlier ones failed.

Outcome

Open. Only test-tube and patient-tissue assays so far. No cell, animal or human results were seen in the text I could read.

Sources

• Horne et al., Nat Chem Biol 2024, DOI 10.1038/s41589-024-01580-x, PMID 38632492, PMC11062903 (no correction or retraction listed)
• Earlier generative-ML paper, J Chem Inf Model 2023, PMC10373478
• Not yet verified: any cell or animal data for I4.05 (none found in the paper's text or in a search to 20 Oct 2026; a 2025 bioRxiv preprint by Zhang, Horne and others generates new inhibitors by machine learning, not read); whether WaveBreak's clinical candidate derives from this work. AlphaFold-designed binders were not included because no concrete landmark was verified.