RFdiffusion3 scores low against all five tau fibril folds
A Columbia University master's student found that RFdiffusion3 produced predicted docking scores (i_pTM) of only 0.116 to 0.401 against all five tau fibril folds, while the same pipeline reached above 0.7 for some fibril forms of other amyloid proteins, and the beta-sheet propensity correlation that predicted success elsewhere (0.80) vanished for tau (−0.19).
15 candidate binders, five tau fibril folds, all predicted binding scores from 0.116 to 0.401 out of 1
All results are computational.
Tau stabilizes microtubules, the internal rails a neuron uses to transport cargo. In several brain diseases the protein aggregates into insoluble fibers called fibrils, each disease producing its own atomic fold: Alzheimer's, chronic traumatic encephalopathy (from repeated blows to the head in boxers and American football players), Pick's disease, corticobasal degeneration, and progressive supranuclear palsy. The laboratory of Michel Goedert and Sjors Scheres at the MRC Laboratory of Molecular Biology in Cambridge solved the atomic structures of all five between 2017 and 2021 using cryo-electron microscopy (freezing the sample and imaging its atomic architecture).
Siddhartha Nanda, a biomedical engineering master's student at Columbia University, posted a preprint on September 22. He ran RFdiffusion3, a generative model that builds a protein backbone from scratch, against all five tau folds. ProteinMPNN then assigned an amino acid sequence to each backbone, and AlphaFold2-Multimer (a model for predicting protein complexes) scored each candidate's predicted docking confidence. The metric, i_pTM, has a maximum of 1; 0.7 is the threshold for a successful design. The method has worked before: the same pipeline designed minibinders against the GDF15 signal that reduced wasting in tumor-bearing mice. Those successes, however, came on globular proteins with binding pockets. Amyloid fibrils are flat, pocketless, built from repeating layers of beta-sheet, a pleated protein conformation.
In a companion preprint posted the same day, the same pipeline handled fibrils of three other proteins unevenly. For AL amyloidosis (protein deposits in tissues) and alpha-synuclein (in Parkinson's), one fibril form scored above 0.7 while another form of the same protein stayed below 0.15. Amyloid-beta, the other protein implicated in Alzheimer's, resisted uniformly: scores from 0.09 to 0.15. One parameter predicted success across these proteins: the more residues in the binder prone to fold into beta-sheet, the higher the score. The correlation was 0.80.
For tau, that pattern vanished. All five folds received low, tightly clustered scores; the best (0.401, for progressive supranuclear palsy) sat nearer to the amyloid-beta failure zone than to the 0.7 threshold. The correlation between the binder's beta-sheet propensity and score became statistically insignificant: −0.19. The one comparatively successful design folded as a bundle of two alpha-helices (twisted protein strands) instead of the expected beta-sheet, a form unexpected for an amyloid target.
The scores reflect the program's predicted docking reliability in one specific design mode. A single follow-up could settle this: force RFdiffusion3 to build beta-sheet binders against tau, the way BindCraft was previously tuned to do. If tau still resists, the problem is in the protein's chemistry. If it yields, the first run simply never tried the right structural form.
Open the related Eternal Search page
Sources
[1] t.me
This preprint delivers a concrete computational result: generative protein design produced uniformly low predicted binding scores across all five disease-associated tau folds, in contrast to partial successes against other amyloid proteins tested with the same method. The Columbia University Eternal Search page connects through the author's institutional affiliation — Siddhartha Nanda is enrolled in Columbia's biomedical engineering master's program — giving readers an organizational entry point grounded in a specific scientific question.