Generative AI protein design failed against all five tau fibril folds in Alzheimer's and four other brain diseases, even though for similar amyloid proteins success previously tracked fibril shape rather than disease
Generative AI protein design failed against all five tau fibril folds in Alzheimer's and four other brain diseases, even though for similar amyloid proteins success previously tracked fibril shape rather than disease
In a preprint posted on September 22, Siddhartha Nanda, a biomedical engineering master's student at Columbia University, ran the diffusion model RFdiffusion3 against five tau protein structures, each resolved by cryo-electron microscopy (freezing the sample and imaging its atomic architecture), one per brain disease from Alzheimer's to progressive supranuclear palsy. AlphaFold2-Multimer scored all fifteen candidate binders. All five folds received low scores close to one another, ranging from 0.116 to 0.401 out of a maximum of 1.
Under normal conditions tau stabilizes microtubules, the internal rails along which a neuron transports cargo. In several brain diseases the protein aggregates into insoluble fibers called fibrils. The atomic structures of these fibrils in five such diseases were solved between 2017 and 2021 by a single laboratory, that of Michel Goedert and Sjors Scheres at the MRC Laboratory of Molecular Biology in Cambridge. Each disease produces its own fold: chronic traumatic encephalopathy, the disease of boxers and American football players caused by repeated blows to the head, folds tau differently from Alzheimer's, which in turn differs from Pick's disease, corticobasal degeneration, or progressive supranuclear palsy.
Generative models such as RFdiffusion build a protein backbone from scratch, and ProteinMPNN then assigns an amino acid sequence to fit that backbone. The method works: the same pipeline designed minibinders against the GDF15 signal and reduced wasting in tumor-bearing mice. Those successes, however, were obtained on globular proteins that present binding pockets, whereas amyloid fibrils are flat, pocketless, and built from repeating layers of beta-sheet, a flat pleated form.
On the same day Nanda posted a second preprint: the same pipeline handled fibrils of three other proteins unevenly. For AL amyloidosis (protein deposits in tissues) and for alpha-synuclein in Parkinson's, both cases examined within a single disease, one fibril polymorph yielded to design (i_pTM, the program's confidence that the binder docks correctly, rose above 0.7 out of 1) while another remained intractable, below 0.15. The third protein, amyloid-beta in Alzheimer's, resisted equally and weakly regardless of polymorph, with the full range spanning 0.09 to 0.15. One parameter predicted success: the more residues in the binder that were prone to fold into the same beta-sheet conformation as the fibril itself, the higher the score, with a correlation of 0.80.
For tau this pattern vanished. All five folds received low scores close to one another, and even the best, for progressive supranuclear palsy at 0.401, sat closer to the hopeless amyloid-beta range than to the threshold for a successful design. The link between the binder's beta-sheet propensity and success disappeared: in place of a 0.80 correlation, a statistically insignificant −0.19. The one comparatively successful design folded not as a beta-sheet but as a bundle of two alpha-helices (twisted, not flat, protein strands), a form unexpected for an amyloid target.
We refer to this finding as uniform resistance with minor variation, not as evidence that this tau fold is amenable to design.
The explanation may lie in the fibril surface itself: a tool that works against other targets finds no grip here, because the edge of the layered structure and the side chains do not accommodate the docking geometry that RFdiffusion3 learned from other amyloids. One test would 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 problem is in the fold that the first run simply never attempted.