Test of AlphaFold 3 exposes limitations in amyloid structure prediction models

AI model generates amyloid-like structures for many non-amyloid controls underlining the need for experimental evidence when researchers assess protein aggregation

A study of AlphaFold 3 has identified substantial limitations in its ability to distinguish amyloid-forming sequences from non-amyloid peptides under the tested conditions. Although the model reproduced five of seven known amyloid structures, it also generated amyloid-like structures for 54 per cent of the negative controls. AlphaFold 3 is an AI model that predicts the three-dimensional structures of biological molecules and how they fit together and was developed by Google DeepMind and Isomorphic Labs.

The investigation has highlighted the difference between a plausible structural model and evidence that a sequence forms an amyloid. The results support caution when researchers interpret confidence scores or use a general protein-structure prediction system for a specialised classification task.

Amyloids are ordered protein assemblies with characteristic structural organisation and disorganisation of this process is associated with Parkinson’s and Alzheimer’s disease. Their biological significance depends on the protein and context, and an amyloid form can differ substantially from a protein’s soluble or membrane-associated structure. A sequence may therefore raise several distinct questions: whether it can assemble into an amyloid, under which conditions that occurs, and what the resulting assembly looks like.

The researchers used three datasets to examine performance. The largest contained 153 amyloid sequences without experimentally resolved three-dimensional amyloid structures. A second set contained 56 non-amyloid peptides, while a third comprised seven proteins with known amyloid structures. These groups allowed the study to assess both apparent amyloid generation and agreement with an established structural reference.

AlphaFold 3 generated amyloid-like structures for 34 per cent of the positive-control sequences. That result describes the model’s output for sequences already classified as amyloid-forming in the dataset. It does not mean that the remaining sequences had lost their biological capacity to form amyloids. Rather that the model failed to produce the relevant form under the evaluation conditions.

The negative controls created a more serious problem for classification. Amyloid-like outputs appeared for 54 per cent of peptides that belonged to the non-amyloid set, a higher proportion than in the positive-control set. A model output of this kind could therefore not serve as a dependable indicator of amyloid formation within the benchmark.

The successful reproduction of five known structures remains meaningful. It showed that the model could represent some experimentally established amyloid architectures. 

But the size of the resolved-structure set also limits the breadth of that positive result. Seven examples cannot represent the diversity of amyloid assemblies, sequence lengths and experimental conditions. A strong outcome for several familiar structural classes may not transfer to an assembly with different organisation. Larger and more varied reference collections would provide a more demanding test.

The researchers found that shorter sequence fragments were more likely to produce an appropriate amyloid model. Fragment selection can therefore influence the apparent result. This is relevant because an assembly-forming region may behave differently from the complete protein, whose other regions introduce further structural possibilities. A fragment-based prediction needs to retain that context when it is used to interpret the full molecule.

Confidence scores presented another difficulty. The model often favoured globular oligomers rather than amyloid forms, even where an amyloid structure was the relevant target. A high score reflects the model’s assessment of its own structural prediction according to its internal criteria. It does not independently establish that the chosen assembly is the state formed in the experiment of interest.

This distinction matters across computational structural biology. A visually convincing model can encourage a precise mechanistic explanation but the precision of its appearance is not a measure of evidential strength. Before a model supports a biological conclusion, researchers need to ask whether the system was validated for that class of structure and whether the proposed state agrees with independent observations.

For laboratories, the practical consequence is to treat amyloid predictions as hypotheses that require experimental support. Structural and biochemical measurements can test whether an assembly forms and whether its properties agree with the model. The study has supplied a valuable boundary test for AlphaFold 3 by selected structural success demonstrating useful capability while the negative-control results show why that capability cannot be assumed to provide a reliable amyloid classifier.

For further reading please visit: 10.1038/s41598-026-68041-4