Non-standard proteins in the lens of AlphaFold 3: a case study of amyloids
TL;DR
AlphaFold 3 struggles to accurately predict the structures of amyloid proteins, correctly identifying only 34% of known amyloids while falsely predicting amyloid structures for 54% of non-amyloid peptides, largely due to a bias toward globular proteins in its training data.
Problem / question
Amyloid proteins, which form fibril structures linked to diseases like Alzheimer's and Parkinson's, are difficult to study experimentally and are severely underrepresented in the Protein Data Bank (PDB). Because AI models like AlphaFold 3 (AF3) are trained primarily on globular proteins, it is unclear if they can accurately predict the complex, polymorphic structures of amyloids.
Methods
The researchers evaluated AF3 using three datasets: 153 known amyloid sequences without solved 3D structures, 56 non-amyloid peptides, and 7 proteins with known amyloid structures (including Aβ-42 and α-synuclein). They generated hexameric (n=6) models using the AF3 web server and compared the predicted structures to experimental data in the PDB and STAMP databases using Foldseek (TM-score) and US-align, while also comparing AF3's performance against other tools like Boltz and RibbonFold.
Key findings
AF3 correctly predicted amyloid structures for only 34% (52 of 153) of the known amyloid sequences, but incorrectly predicted amyloid folds for 54% (30 of 56) of the non-amyloid peptides. For the 7 known amyloid structures, AF3 successfully modeled 5 (such as Aβ-42 and α-synuclein) but completely failed on transthyretin and immunoglobulin, predicting them as globular instead. Notably, incorrect globular predictions received higher confidence scores (pLDDT was 18 points higher on average) than correct amyloid predictions, strongly correlating with the presence of globular homologs in the PDB training data.
Why it matters
This reveals that AF3's high confidence scores can be misleading for polymorphic proteins, demonstrating that AI structure prediction still requires significant improvement and human oversight, such as manually setting oligomeric states, to accurately model disease-relevant amyloid fibrils.
Limitations
The non-amyloid control dataset was limited to short sequences (under 36 amino acids), and the operational choice to model all amyloids as hexamers (n=6) may not be optimal for every protein. Additionally, the explanation for AF3's failures is correlational, limited by the black-box nature of the AI.
Takeaway
AlphaFold 3 often defaults to predicting globular structures for amyloid proteins due to training data bias, meaning researchers cannot blindly trust its outputs or high confidence scores (pLDDT) when modeling amyloid fibrils.