structural-biology/ai produced the result/Proteomes 2026 · v2
Software predictions map floppy stretches of touch-sensing PIEZO channels
Researchers ran the human PIEZO1 and PIEZO2 protein sequences through a set of off-the-shelf sequence predictors, several of them neural networks, to estimate which stretches are disordered, which may bind partners, and how clinical mutations line up with them.
spectrum · one line per step, placed by what the step does · bright lines used AI
A Propeller with a Flexible Twist: A Computational Analysis of Intrinsically Disordered Regions in PIEZO Gating and PIEZO-Associated Channelopathies
Proteomes, 2026
doi:10.3390/proteomes14030041 · record aix-00074 v2 · checked 2026-10-08
- AI was for
- Classification, Property prediction
- Model family
- Multilayer perceptron
- Checked by
- None stated
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Most proteins are usually pictured as having a fixed shape, like a tool machined to fit a job. But many contain stretches that never settle into one form, wriggling instead between shapes. These are called intrinsically disordered regions. They are hard to study in the laboratory precisely because they do not hold still: the usual methods for working out a protein's shape rely on it being rigid enough to measure. PIEZO1 and PIEZO2 are channels that sit in the membranes of our cells and open when the cell is pushed or stretched, which is part of how we sense touch and how blood vessels respond to flow. Their rigid, propeller-shaped parts are better described than their floppy parts.
The researchers set out to chart those floppy parts from sequence alone, without new experiments. They asked how much of each protein is predicted to be disordered, which disordered stretches look likely to grab onto other proteins, and whether those stretches might separate out into droplets inside the cell. They then compared these predicted regions against mutations recorded in patients, to see whether disease-linked changes and harmless ones fall in different kinds of territory.
Where AI came in
Everything about this study's picture of PIEZO disorder came from software run on the two protein sequences. Six predictors of disorder, some of them trained neural networks, were run together through one platform to score every amino acid, giving the proportion predicted to be disordered and a combined profile; a second platform flagged regions where at least three quarters of predictors agreed. Further tools predicted stretches that fold up only on meeting a partner, and the chance that each protein would separate into droplets. AlphaFold models were also looked at for segments the software itself marked as uncertain.
These predictors stood in for the bench experiments that would otherwise be needed to find where a large membrane protein is flexible and what it may bind. Nothing was measured physically. Later steps, including mapping clinical variants onto the predicted regions, the statistical tests, the peptide modelling of two mutation sites and the conservation comparisons, used non-AI tools, but they took the predicted regions as their starting point. The paper states that the results still require experimental validation.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Sequence-based predictors were run on the canonical human PIEZO1 and PIEZO2 sequences to estimate per-residue intrinsic disorder, disorder-based binding regions and phase-separation propensity. The consensus mean disorder profile gave 20.31% predicted disordered residues for PIEZO1 and 23.76% for PIEZO2, and FuzDrop gave spontaneous phase-separation probabilities of 0.66 for PIEZO1 and 0.92 for PIEZO2. ClinVar variants were then mapped onto these predictions; in PIEZO1, 4.5% of pathogenic variants fell within predicted disordered regions compared with 23.1% of benign variants (Fisher's exact test, p = 0.075), and pathogenic variants had lower mean PONDR VSL2 scores at the mutation site than benign variants. Peptide models of one selected variant site in each protein were compared between wild-type and mutant sequences, and ConSurf grades were compared between disordered and non-disordered residues.
How AI was used
Canonical human PIEZO1 and PIEZO2 sequences and their UniProt feature tables were retrieved and passed to a set of off-the-shelf sequence-based predictors. Six per-residue disorder predictors (PONDR VLXT, VSL2, VL3 and FIT, IUPred-S and IUPred-L) were run through the RIDAO platform to produce per-residue scores, the percentage of predicted disordered residues, average disorder scores and a consensus mean disorder profile, with the D2P2 platform used to define high-confidence regions where at least 75% of predictors agreed. ANCHOR2 and MoRFpred were run to predict disorder-based binding regions, and FuzDrop, MolPhase, ParSe 2.0 and AutoFuz to predict phase-separation probability, droplet-promoting regions, aggregation hotspots and context-dependent interaction regions. Interaction networks and Gene Ontology enrichment were obtained from STRING. ClinVar variants were mapped onto the predicted disordered and binding-prone regions, with local disorder quantified as the PONDR VSL2 score at the mutation site and averaged over a seven-residue window, and overlaps tested with Fisher's exact and Wilcoxon rank-sum tests in R. Peptide fragments spanning two selected variant sites were modelled de novo in wild-type and mutant form with PEP-FOLD4, top models being chosen by sOPEP energy. Conservation was assessed with Clustal Omega and MAFFT alignments in Jalview and residue-level ConSurf grading over 150 selected homologs per protein. AlphaFold structural models were inspected for low-confidence, flexible segments.
The shape of the work
Structural · the record, drawn
no AI
Retrieve canonical sequences and UniProt annotations
Obtaining raw data, whether by measurement, download or retrieval.
Protein sequence data for the mechanosensitive ion channels PIEZO1 and PIEZO2 were obtained from the UniProt Consortium databasewhere the paper describes this · verbatim
AI
Predict per-residue intrinsic disorder
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
Intrinsic disorder within the full-length protein sequences of PIEZO1 and PIEZO2 was evaluated using a multi-predictor approach through the RIDAO platformwhere the paper describes this · verbatim
AI
Predict disorder-based binding regions (MoRFs)
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
disorder-based binding regions were predicted using ANCHOR2where the paper describes this · verbatim
AI
Predict phase-separation propensity
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
The probability of PIEZO1 and PIEZO2 to undergo liquid–liquid phase separation (LLPS) was evaluated using the FuzDrop algorithmwhere the paper describes this · verbatim
no AI
Build interaction networks and run functional enrichment
Obtaining raw data, whether by measurement, download or retrieval.
Protein–protein interaction (PPI) networks for PIEZO1 and PIEZO2 were generated using the STRING database (version 12.0)where the paper describes this · verbatim
no AI
Map clinical variants onto predicted regions and test enrichment
Extracting understanding from model behaviour.
Residue-level mappings were used to determine whether variants occurred within (i) long intrinsically disordered regions (IDRs)where the paper describes this · verbatim
no AI
Model wild-type and mutant peptide conformations
Numerical or physics simulation, including where a learned surrogate replaces it.
Both wild-type and mutant peptide sequences were submitted independently to PEP-FOLD4 for de novo structure predictionwhere the paper describes this · verbatim
no AI
Assess evolutionary conservation of full-length proteins and IDRs
Extracting understanding from model behaviour.
Residue-level conservation was further assessed using ConSurfwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The study is entirely computational; the reported disorder content, binding regions, phase-separation propensities and variant-context comparisons are all outputs of sequence-based predictors, so the findings rest on model output.
the results require experimental validationwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- ValidationNo validation of the AI is described.
- Version of PONDR VLXTWhich version of the model was used is not stated.
- Version of PONDR VSL2Which version of the model was used is not stated.
- Version of PONDR VL3Which version of the model was used is not stated.
- Version of PONDR FITWhich version of the model was used is not stated.
- Version of IUPred-SWhich version of the model was used is not stated.
- Version of IUPred-LWhich version of the model was used is not stated.
- Version of ANCHOR2Which version of the model was used is not stated.
- Version of MoRFpredWhich version of the model was used is not stated.
- Version of FuzDropWhich version of the model was used is not stated.
- Version of MolPhaseWhich version of the model was used is not stated.
- Version of PrDOS (via D2P2)Which version of the model was used is not stated.
- Version of PV2 (via D2P2)Which version of the model was used is not stated.
- Version of ESpritz (NMR, DisProt and X-ray variants, via D2P2)Which version of the model was used is not stated.
- Version of AlphaFold (structural models inspected)Which version of the model was used is not stated.
About this article
Record aix-00074, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error