💻 bioinformatics

AlphaFold3 and Intrinsically Disordered Proteins: Reliable Monomer Prediction, Unpredictable Multimer Performance

This study reveals that while AlphaFold3 reliably predicts intrinsically disordered monomers by learning from training data, its multimer predictions remain unpredictable due to rigid structural priors that override input information, highlighting the critical need for disorder-enriched datasets and ensemble sampling over mere architectural innovation.

Original authors: Dao, T. M., Ghent, S., Uversky, V. N., Rahman, T.

Published 2026-10-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dao, T. M., Ghent, S., Uversky, V. N., Rahman, T.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Proteins are the workhorses of life, tiny machines that build cells, fight infections, and carry out the chemical reactions that keep us alive. For decades, scientists believed that to function, these molecules had to fold into a single, rigid, three-dimensional shape, much like a key fits into a lock. However, nature is more flexible than that. A significant portion of the proteins in our bodies, particularly in complex organisms like humans, are intrinsically disordered. These are not misfolded mistakes; they are functional by design. They remain floppy and shapeless until they meet a partner, at which point they might snap into a specific form to do their job, or they might stay flexible to act as a tether or a signal. Understanding how these shape-shifting molecules behave is crucial for medicine, as their malfunction is linked to diseases ranging from cancer to neurodegeneration.

For a long time, predicting the shape of these floppy proteins was a major hurdle for computers. Then came a breakthrough in artificial intelligence known as AlphaFold. The first version, released in 2021, revolutionized the field by predicting the shapes of rigid proteins with near-perfect accuracy. In 2024, a new version, AlphaFold3, arrived with a completely redesigned engine, promising to handle even more complex biological puzzles, including how proteins interact with one another. But a critical question remained unanswered: did this new, more powerful machine actually get better at handling the floppy, disordered proteins that make up so much of our biology, or did it simply carry over the same limitations as its predecessor?

A team of researchers set out to answer this by putting AlphaFold3 through a rigorous test specifically designed for disordered systems. They did not just look at single proteins; they examined how these molecules behave when they try to grab onto other proteins, a process that often involves a chaotic dance of folding and unfolding. The researchers gathered a large collection of real-world protein complexes, many of which had been solved in laboratories but were not part of the data the computer had seen during its training. They then asked the computer to predict the shapes of these interactions and compared the results against the known experimental structures.

The findings revealed a story of two different worlds. When the computer looked at single, isolated proteins, it performed with remarkable reliability. The new system could accurately identify which parts of a protein were floppy and which were rigid, matching the performance of the older version and rivaling specialized tools built just for this purpose. This suggests that the ability to spot disorder is not a trick of the specific computer architecture, but a skill learned from the vast library of biological data the system was trained on. If a protein is naturally floppy, the computer knows it.

However, the picture changed dramatically when the researchers asked the computer to predict how these proteins interact in pairs or larger groups. Here, the new system did not show a clear improvement over the old one. While the average quality of the predictions was similar, the way the computer arrived at those answers was fundamentally different and, in some ways, more unpredictable. The researchers found that for the older system, they could explain most of its successes and failures by looking at simple physical features of the interaction, such as how much surface area the proteins covered or how much disorder was present at the contact point. For the new system, these simple clues explained far less of the outcome. The computer's performance seemed to depend on factors that were harder to see and harder to predict.

The study uncovered that the new system sometimes succeeds where the old one fails, and vice versa, but not in a consistent pattern that a user could easily anticipate. In some cases, the new system would produce a perfect prediction, while in others, it would fail completely, even when the physical conditions looked identical. The researchers ran the same prediction multiple times with slight variations to see if the results would change. They found that for many difficult cases, the system would almost always produce the same wrong answer, but occasionally, just by chance, it would stumble upon the correct shape. This means that unlike a tool that consistently explores different possibilities, the new system often locks onto a single outcome, and that outcome is not always the right one.

Despite these quirks, the new system did show strength in a specific, challenging scenario: when a disordered protein folds into a solid shape only after it binds to a partner. In these cases, both the old and new versions performed equally well, successfully predicting the final, stable structure. This is likely because the final shapes in these cases look like the rigid proteins the computer has seen millions of times in its training data. However, the system still struggled to predict which parts of a protein would fold and which would remain floppy before the binding happened, a capability essential for understanding how these interactions start.

The researchers concluded that while the new artificial intelligence model is a powerful tool for predicting the shapes of single proteins and their disordered regions, it has not yet solved the mystery of how these molecules interact in complex groups. The improvements in its design did not translate to a better understanding of disorder in multi-protein systems. Instead, the new system introduced a layer of unpredictability that makes it difficult for scientists to know when to trust its output. The study suggests that simply building a more complex computer model is not enough to solve this problem. To truly master the behavior of these shape-shifting molecules, the field needs training data that includes more examples of disordered proteins and methods that can capture the full range of their possible shapes, rather than just the single, most likely one. Until then, scientists must use these powerful tools with caution, knowing that for the most dynamic parts of biology, the answers are not always as clear as they appear.

Technical Summary: AlphaFold3 and Intrinsically Disordered Proteins

Problem Statement
Intrinsically disordered proteins (IDPs) and regions (IDRs) constitute 30–40% of eukaryotic proteomes and are critical for cellular regulation, signaling, and condensate formation. Unlike folded proteins, IDPs lack stable 3D structures, often mediating interactions through disorder-to-order transitions, fuzzy complexes, or fly-casting mechanisms. While AlphaFold2 (AF2) demonstrated an emergent capability to predict disorder via its confidence metric (pLDDT) despite not being explicitly trained for it, the performance of AlphaFold3 (AF3) on disordered systems remained uncharacterized. AF3 introduced a diffusion-based generation architecture and unified molecular representations, raising critical questions: Does this new architecture preserve AF2's proficiency in disorder prediction? Does diffusion sampling better explore conformational ensembles of IDPs, or does it inappropriately collapse diverse states? Without systematic benchmarking, users lack guidance on trusting AF3 predictions for the substantial fraction of biological interactions involving disordered regions.

Methodology
The authors conducted the first systematic evaluation of AF3 on disordered systems using three distinct datasets and rigorous statistical analysis:

Monomer Disorder Prediction: Utilized the DisProt-PDB dataset (478 targets with experimentally validated disorder annotations). AF3's pLDDT scores were optimized against 11 established disorder predictors (including specialized tools like spot_disorder and sequence-based methods) using Matthews Correlation Coefficient (MCC), AUC, and F-max metrics.
Multimer Prediction (B90 Benchmark): Constructed a dataset of 90 non-redundant IDR-receptor dimers from PDB entries released after September 30, 2021, to ensure zero training contamination for AF3. This dataset was stratified by buried surface area, interface coil content, and disorder fraction. Both AF3 and AF2-Multimer were run with default parameters, and predictions were evaluated using DockQ scores, interface contact accuracy, and secondary structure matching (DSSP).
Disorder-to-Order Transitions (MFIB Benchmark): Evaluated 90 complexes from the Mutual Folding-Induced Binding (MFIB) 2.0 database, where IDRs fold upon binding. Performance was assessed via Interface RMSD and Ligand RMSD.
Mechanistic and Variance Analysis:
- Regression Analysis: Linear regression models were built to determine how much variance in DockQ scores could be explained by interface disorder fraction and secondary structure matching (DSSP accuracy) for both models.
- Seed Variance Analysis: For six high-divergence cases, 30 independent AF3 predictions were generated using different random seeds to distinguish between stochastic sampling and systematic failure.
- MoRF Analysis: Assessed the ability to identify Molecular Recognition Features (short disordered segments that fold upon binding) via delta pLDDT and confidence metrics.

Key Contributions and Results

Reliable Monomer Disorder Prediction: AF3's pLDDT scores reliably predict disorder, achieving an MCC of 0.693, which is comparable to AF2 (0.701) and rivals dedicated disorder predictors. Both models accurately recapitulate amino acid-specific disorder propensities (correlation $r=0.985$ with ground truth), confirming that disorder prediction emerges from training data composition rather than specific architectural design.
Comparable Aggregate Multimer Performance: On the B90 benchmark, AF3 and AF2 achieved nearly identical aggregate performance (Mean DockQ: 0.563 vs. 0.571). Both models struggle similarly with high-disorder interfaces (>40% disorder), showing reduced contact accuracy.
Divergent Mechanisms of Failure: Despite similar aggregate scores, the models operate via fundamentally different mechanisms. Conventional structural features (interface disorder and secondary structure) explain 58% of the variance in AF2's DockQ scores but only 42% of AF3's. This suggests AF3's performance depends on factors beyond simple linear relationships with interface properties, making it harder to predict success or failure based on standard structural metrics alone.
Disorder-to-Order Transitions: On the MFIB benchmark, both models perform equally well, successfully predicting the final folded states of complexes where IDRs fold upon binding. However, neither model can prospectively identify which disordered regions will undergo folding (MoRFs), as confidence metrics (pLDDT) show no correlation with binding-induced folding propensity.
Seed Variance and Failure Modes: Analysis of 30 seeds for high-divergence cases revealed that AF3 rarely converges reliably. While some cases are deterministic, others exhibit bimodal distributions where a small minority of seeds (7–17%) locate a correct solution while the majority fail. This indicates AF3 can occasionally find correct solutions without consistently identifying them, rather than engaging in productive ensemble exploration. Four distinct failure modes were identified: structured IDPs where AF2's MSA signal outperforms AF3's reduced signal; diffusion collapse; genuine disorder where AF3 outperforms AF2; and MSA over-assignment where AF3 avoids noise that misleads AF2.

Significance and Claims
The paper establishes a dichotomy in AF3's utility: it is reliable for monomer disorder prediction but unpredictable for multimeric assemblies involving disorder. The authors argue that architectural innovation alone (e.g., switching from Evoformer to diffusion) cannot overcome biases inherent in the training data, which is heavily skewed toward ordered structures.

The study concludes that progress in modeling IDPs requires disorder-enriched datasets and ensemble sampling strategies rather than merely novel architectures trained on ordered-structure-biased data. For practitioners, the findings imply that AF3 predictions for disordered complexes should be treated with caution; users are advised to compare multiple seeds and cross-validate with AF2, as large divergences indicate structural ambiguity requiring experimental validation. The paper does not claim AF3 has solved the challenge of IDP modeling but highlights that its multimer performance operates through incompletely understood principles with unpredictable outcomes for disorder-containing complexes.

Technical Summary: AlphaFold3 and Intrinsically Disordered Proteins

More like this