Related Experiment Video
Updated: Jun 15, 2025
12:42
Microfluidic Mixers for Studying Protein Folding
Published on: April 10, 2012
15.1K
AlphaFold predictions of fold-switched conformations are driven by structure memorization
Devlina Chakravarty1, Joseph W Schafer1, Ethan A Chen1
1National Center for Biotechnology Information, National Library of Medicine, National Institutes of Health, Bethesda, MD, 20894, USA.
Nature Communications
|August 24, 2024
Summary
Deep learning models like AlphaFold struggle to predict protein fold switching, often memorizing training data instead of learning protein energetics. Physically based methods are needed for accurate prediction of multiple protein conformations.
Area of Science:
- Protein structure prediction
- Computational biology
- Biophysics
Background:
- Deep learning models, such as AlphaFold (AF), have shown promise in predicting protein structures from sequences.
- Protein energy landscapes, characterized by conformational diversity, are crucial for understanding protein function.
- Fold-switching proteins present a challenge due to their ability to adopt multiple distinct structures.
Purpose of the Study:
- To evaluate the predictive capabilities of AlphaFold for fold-switching proteins.
- To determine if AlphaFold can accurately model protein energy landscapes and conformational dynamics.
- To identify limitations of deep learning models in predicting complex protein behaviors.
Main Methods:
- Testing AlphaFold (AF2 and AF3) on known fold-switching proteins.
- Analyzing over 560,000 predicted models from multiple AlphaFold implementations.
- Comparing predictions with experimentally determined structures and assessing AlphaFold's confidence metrics.
- Investigating potential memorization of training data and misassignment of coevolutionary restraints.
Main Results:
- AlphaFold demonstrates weak predictive power for protein fold switching.
- A significant portion of AlphaFold's successes on fold-switching proteins stems from memorization of training data.
- AlphaFold's confidence metrics are unreliable for identifying correct fold-switching conformations.
- Performance was limited to 35% success for proteins within the training set and only 1 out of 7 for proteins outside the training set.
Conclusions:
- Current deep learning models like AlphaFold have significant limitations in predicting protein fold switching.
- The models appear to memorize structural information rather than learning underlying biophysical principles.
- There is a critical need for physically based computational methods to accurately predict multiple protein conformations and energy landscapes.
Related Concept Videos
01:25Protein Folding
7.8K
Proteins are chains of amino acids linked together by peptide bonds. Upon synthesis, a protein folds into a three-dimensional conformation, critical to its biological function. Interactions between its constituent amino acids guide protein folding, and hence the protein structure is primarily dependent on its amino acid sequence.
Protein Structure Is Critical to Its Biological Function
Proteins perform a wide range of biological functions such as catalyzing chemical reactions, providing...
Protein Structure Is Critical to Its Biological Function
Proteins perform a wide range of biological functions such as catalyzing chemical reactions, providing...
7.8K
03:00Molecular Chaperones and Protein Folding
17.8K
The native conformation of a protein is formed by interactions between the side chains of its constituent amino acids. When the amino acids cannot form these interactions, the protein cannot fold by itself and needs chaperones. Notably, chaperones do not relay any additional information required for the folding of polypeptides; the native conformation of a protein is determined solely by its amino acid sequence. Chaperones catalyze protein folding without being a part of the folded protein.
The...
The...
17.8K
03:03Amyloid Fibrils
9.3K
Amyloid fibrils are aggregates of misfolded proteins. Under most circumstances, misfolded proteins are either refolded by chaperone proteins or degraded by the proteasome. However, in the case of a mutation or a disease, these proteins can accumulate to form large clusters and often further assemble to form elongated fibers, called fibrils.
Amyloid deposits were observed as early as 1639 in the liver and the spleen. In 1854, Rudolph Virchow performed iodine staining,...
Amyloid deposits were observed as early as 1639 in the liver and the spleen. In 1854, Rudolph Virchow performed iodine staining,...
9.3K
02:26Conservation of Protein Domains Over Different Proteins
10.8K
Protein domains are small structurally independent units that are part of a single amino acid chain. Although these domains are often structurally independent, they may rely on synergistic effects to perform their functions as part of a larger protein. Protein domains may be conserved within the same organism, as well as across different organisms.
A limited set of protein domains often duplicate and recombine during evolution. These domains can be organized in different combinations to...
A limited set of protein domains often duplicate and recombine during evolution. These domains can be organized in different combinations to...
10.8K
01:58Cooperative Allosteric Transitions
7.9K
Cooperative allosteric transitions can occur in multimeric proteins, where each subunit of the protein has its own ligand-binding site. When a ligand binds to any of these subunits, it triggers a conformational change that affects the binding sites in the other subunits; this can change the affinity of the other sites for their respective ligands. The ability of the protein to change the shape of its binding site is attributed to the presence of a mix of flexible and stable segments in the...
7.9K
01:49Conserved Binding Sites
4.2K
Many proteins’ biological role depends on their interactions with their ligands, small molecules that bind to specific locations on the protein known as ligand-binding sites. Ligand-binding sites are often conserved among homologous proteins as these sites are critical for protein function.
Binding sites are often located in large pockets, and if their location on a protein’s surface is unknown, it can be predicted using various approaches. The energetic method computationally...
Binding sites are often located in large pockets, and if their location on a protein’s surface is unknown, it can be predicted using various approaches. The energetic method computationally...
4.2K