Predicting fold-switching proteins and both of their folds from sequence
Research-grade (50-69): the leading predictors are documented to fail outside memorised examples, and workarounds are disputed, but benchmarks are clear and progress is active. It sits above the 3D core-collapse explosion-energy problem (50) because no method works prospectively, and far below the defining open problems of other fields. Application value 12 matches niche target biology (compare dark-matter detection spin-offs at 14); leverage 30 reflects that success would mostly inform multi-state prediction rather than gate a programme.
The problem
Some proteins can fold into two different shapes; current AI mostly finds only one.
- 1Most proteins: one sequence, one fold
- 2Switchers have two stable shapes
- 3Evolution clue for second shape is weak
- 4AI recalls the common shape
- 5Need a test it cannot memorise
What is being asked. A fold-switching protein populates two or more folds that differ in secondary structure or topology, with a trigger such as a ligand, redox state or binding partner (examples: RfaH, KaiB, lymphotactin). The task is to decide from sequence alone whether a protein switches and to predict both folds.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- Predicting fold-switching proteins and both of their folds from sequenceThis problem
- Then
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceComputational Medicine · D64
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Reverse-mode automatic differentiationAlgorithm
- Fast matrix multiplicationAlgorithm
- Step 2
- The transformer architectureAlgorithm
- Step 3
- Biomolecular structure prediction (AlphaFold-style)Algorithm · host
- Then
- Predicting fold-switching proteins and both of their folds from sequenceThe problem
State of the art
- AlphaFold2 default behaviour. Chakravarty and Porter (Protein Science, 2022) tested 98 fold-switching proteins and found AlphaFold2 predicts only one of the two folds for most; confidence scores favoured the wrong state ("AlphaFold2 fails to predict protein fold switching").
- MSA manipulation. Wayment-Steele et al. (Nature 2024) showed that clustering the MSA by sequence similarity (AF-Cluster) lets AF2 sample alternate states of known metamorphic proteins (KaiB, RfaH, others). Schafer et al. (preprint, 2024) reported the method also predicts spurious switches for KaiB homologs that experimentally do not switch, and other groups found random sampling can match clustering.
- Memorisation. Chakravarty et al. (Nature Communications 2024) pooled over 280,000 AF2 and AF3 models: for fold switchers likely present in the training set the success rate was 35%, but AlphaFold captured only 1 of 7 experimentally confirmed fold switchers outside its training sets despite extensive sampling; AF3 was reported to mis-assign coevolutionary restraints.
- CASP16 (2024). In the first ensemble-style experiment with multiple conformational states, full or partial success on both states was reached for about half of the targets, and the best groups ran many AlphaFold2 models with enhanced MSAs and sampling (bioRxiv preprint; the AlphaFold3 server also did well on several targets).
- Reviews and preprints. A 2025 review by Chakravarty and Porter (arXiv:2507.10839) summarises that evolution has selected for dual-folding in some families and that the models fail even when coevolutionary signal for both folds is present in the MSA. A bioRxiv preprint (Oct 2025) proposes modifying AF2's coevolutionary input to uncover distinct conformations; it has not been refereed.
- No method meets the benchmark above as of October 2026 in sources reviewed; reported successes depend on known examples or on training-set overlap.
Realistic applications
Reliable fold-switch detection would flag drug targets whose binding pockets exist only in one state (as for the KaiB clock protein or some transcription factors), improve annotation of unannotated genomes (a switched fold changes function), and give a testbed for any method that claims to model protein energy landscapes. No near-term clinical deployment is claimed.
What counts as solved
A peer-reviewed, prospective blind evaluation (CASP/CAMEO-style, or a pre-registered test on newly solved structures) in which a single method meets the benchmark above: both folds correct (TM-score ) for at least 70% of held-out fold switchers with at most 5% false positives among single-fold controls, with code and weights released.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
A single method must pass a prospective blind evaluation (CASP/CAMEO-style or a pre-registered test on newly solved structures) with both folds correct (TM-score 0.8 or higher) for at least 70% of held-out fold switchers and at most 5% false positives among single-fold controls, with code and weights released so the evaluation can be rerun.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
The central need is a model of why a sequence can encode two folds and how to detect weak coevolutionary signal for the minor fold, plus more experimentally verified fold switchers to train and test on.
Bottlenecks
ranked by votes- #1TheoryNo energy-landscape model for dual-fold sequences
Predictors learn the dominant fold of a family; the evolutionary signal for a second fold is weak and mixed with the first. A principled model of when sequence encodes two basins is missing.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Sep 2, 2025CASP16 ensembles experimentsource
Both states reproduced fully or partly for roughly half of the multi-state targets; enhanced-sampling AF2 variants worked best.
- Jul 14, 2025Annual Review of Biophysics reviewsource
Review estimates nearly 100 characterised fold switchers and up to 4% of PDB entries.
- Aug 24, 2024Memorisation explains many AF successessource
Over 280,000 models: 35% success for likely-trained fold switchers, 1 of 7 outside the training set.
- Jan 1, 2024AF-Cluster samples alternate statessource
MSA clustering lets AF2 sample both folds of KaiB, RfaH and others; a later preprint shows false positives on non-switching KaiB homologs.
- May 1, 2022AlphaFold2 fails on fold switchers