AI predicts shapeless proteins well—until they join together

A study of AlphaFold3 finds consistent predictions for individual disordered proteins but uneven results for protein assemblies.

Editorial illustration — not from the study.

Researchers systematically evaluated AlphaFold3 on proteins that lack a single stable shape, both individually and in assemblies. For individual proteins, AlphaFold3’s confidence scores were a reliable guide to disorder and performed similarly to AlphaFold2 and dedicated disorder predictors.

For assemblies, AlphaFold3 had aggregate docking performance similar to AlphaFold2, but the two systems appeared to achieve those results differently. AlphaFold3 also often produced substantially different results from independent runs, with occasional strong predictions among otherwise weaker ones. Both systems performed equally well on cases where disordered proteins fold when they interact, predicting the final folded states.

Where predictions hold up

For individual proteins, AlphaFold3’s pLDDT confidence scores predicted disorder with a Matthews correlation coefficient of 0.693, matching AlphaFold2 and approaching the performance of specialized predictors. The authors say this consistency across different system designs suggests that the ability to recognize disorder comes from the training data rather than from a particular model architecture.

For protein assemblies, AlphaFold3 and AlphaFold2 had similar average DockQ scores: 0.563 for AlphaFold3 and 0.571 for AlphaFold2. However, conventional features of the contact area explained 42% of the variation in AlphaFold3’s results, compared with 58% for AlphaFold2. This means interface properties alone did not show when AlphaFold3 would succeed or fail.

On the MFIB benchmark, which includes cases where proteins become ordered when they interact, both systems predicted the final folded states equally well. Results from repeated AlphaFold3 runs often varied substantially. In many cases, an occasional high-scoring run appeared within a broader set of lower-scoring results, suggesting that the system can sometimes find a correct structure without finding it consistently.

Why the distinction matters

The findings distinguish two uses of AlphaFold3. Its confidence scores can help identify disorder in individual proteins, but its predictions for assemblies of disordered proteins are less predictable. Researchers using it for these assemblies may therefore need to examine multiple independent predictions rather than rely on a single result or on interface features alone.

The study also points to a data problem: changing the model architecture by itself may not resolve these difficulties. The authors argue that progress will require more training data focused on disordered proteins and methods that combine results from multiple prediction runs.

Evidence and limits

This is a systematic evaluation published in the journal Proteins: Structure, Function, and Bioinformatics. It compares AlphaFold3 with AlphaFold2 on monomer and multimer cases, including the MFIB benchmark, and analyzes variation across independent runs.

The abstract does not report the number of proteins or assemblies tested, so the breadth of the evaluation cannot be assessed from the available information. The reported aggregate docking scores also do not show that either system is consistently accurate for every assembly. The results specifically concern the tested disordered-protein systems and do not establish performance for all protein-prediction tasks.

// Source

Proteins Structure Function and Bioinformatics · 2026 · DOI: 10.1002/prot.70189

View paper (DOI)Open access versionPublished 2026-10-12

Institutions: University of South Florida

Human brain organoids classified Braille letters with up to 83% accuracy

Human forebrain organoids classified artificial tactile Braille signals after a sensor converted the touch data into electrical stimulation. Accuracy averaged 61% with one organoid and rose to 83% when responses from three organoids were combined.

1 hour ago

Biology

A peptide reduced persistent traumatic memories in mice

A study in mice found that a peptide targeting the memory-related protein CPEB3 reduced the persistence of traumatic memories and anhedonia, a loss of pleasure. The findings are an early genetic preclinical test of a possible approach for PTSD, not evidence of a treatment in people.

6 hours ago·5 views

Biology

In fruit flies, heat and pesticide stress cut fertility across generations

Acute heat and sublethal imidacloprid exposure interacted to reduce male fruit flies’ reproductive success more than either stressor alone in some conditions. Stress experienced by fathers also reduced offspring fertility, while a more diverse gut microbiome partly lessened the effects on males.

20 hours ago·5 views