This repository contains predicted structural ensembles of voltage-gated ion channels generated for the study “Targeted MSA Masking Reshapes Structural Sampling in Mutant Voltage-Gated Ion Channels.”

The dataset was created to investigate how targeted masking of multiple-sequence-alignment information reshapes the conformational states sampled for wild-type and mutant voltage-gated ion channels.

The current release contains AlphaFold2-derived structural ensembles for three channel families:

Kv2.1
Nav1.5
Cav1.2

The deposited ensembles include wild-type and mutant sequences generated under conventional, or vanilla, prediction conditions and under targeted MSA masking conditions.

The current release contains:

23 AlphaFold2 structural ensembles
Wild-type and mutant channel sequences
Vanilla and targeted-MSA-masked prediction conditions
Final and intermediate AlphaFold2 prediction snapshots where available
A manifest describing the contents and completeness of each deposited ensemble

AlphaFold3 and ESMFold ensembles may be added in later releases using the same channel- and variant-centered organization.

Associated manuscript

This dataset was generated for the manuscript:

Targeted MSA Masking Reshapes Structural Sampling in Mutant Voltage-Gated Ion Channels

The deposited models support analyses of:

Structural diversity within predicted ensembles
Differences between wild-type and mutant channels
Changes in conformational sampling caused by targeted MSA masking
Sampling of alternative channel states
Structural-distance distributions
Recycle-convergence quality control
Comparisons between complete and controlled model subsets

The manuscript author list, preprint link, journal information, and DOI will be added when publicly available.

Biological systems

Kv2.1

The Kv2.1 data include wild-type and mutant channel sequences such as:

WT
L403A
F412L

Vanilla and targeted-MSA-masked conditions are deposited where available.

Nav1.5

The Nav1.5 data include channel variants such as:

Multiple masking configurations are retained as separate ensembles when they represent distinct modeling experiments.

Cav1.2

The Cav1.2 data include wild-type and mutant channel sequences such as:

WT
G402S
G406R

Vanilla and targeted-MSA-masked conditions are deposited where available.

The folder structure and ensemble manifest are the authoritative records of the conditions currently included in the repository.

Repository organization

Structural models are organized first by channel, then by sequence variant, and finally by prediction method and modeling condition.

data/
├── Kv2.1/
│   ├── WT/
│   │   ├── af2_vanilla_models/
│   │   └── af2_masked_models/
│   ├── L403A/
│   │   ├── af2_vanilla_models/
│   │   └── af2_masked_models/
│   └── F412L/
│       ├── af2_vanilla_models/
│       └── af2_masked_models/
│
├── Nav1.5/
│   ├── WT/
│   │   ├── af2_vanilla_models/
│   │   ├── af2_masked_models/
│   │   └── additional masking conditions where available
│   └── QQQ/
│       ├── af2_vanilla_models/
│       └── af2_masked_models/
│
└── Cav1.2/
    ├── WT/
    │   ├── af2_vanilla_models/
    │   └── af2_masked_models/
    ├── G402S/
    │   ├── af2_vanilla_models/
    │   └── af2_masked_models/
    └── G406R/
        ├── af2_vanilla_models/
        └── af2_masked_models/

Future prediction methods may be added in parallel folders such as:

data/Kv2.1/WT/af3_models/
data/Kv2.1/WT/esmfold_models/

Ensemble manifest

The per-ensemble inventory is available at:

metadata/af2_ensemble_manifest.tsv

The manifest records information such as:

Channel
Sequence variant
Modeling condition
Number of deposited PDB structures
Approximate ensemble size
Hugging Face destination
Pre-upload completeness status

The manifest should be consulted before comparing ensemble sizes or selecting conditions for downstream analyses.

Prediction conditions

Vanilla ensembles

Vanilla ensembles were generated using the standard MSA information prepared for the corresponding sequence and AlphaFold2 prediction workflow.

These ensembles provide the baseline for evaluating the conformational distributions sampled without targeted removal of evolutionary information.

Targeted-MSA-masked ensembles

Masked ensembles were generated after selectively masking positions in the input MSA.

The masking strategy was designed to alter the evolutionary constraints presented to the structure prediction model and evaluate whether this modification changes or expands the predicted structural states sampled by the ensemble.

Different masking configurations are retained as separate conditions when they correspond to distinct modeling experiments.

Test and exploratory ensembles

Folders containing labels such as test, masked_test, v2, noIFM, or other configuration identifiers represent distinct exploratory or methodological conditions.

These conditions are preserved for transparency and reproducibility. They should not automatically be treated as interchangeable with the primary masked condition.

AlphaFold2 ensemble composition

Most complete AlphaFold2 conditions contain approximately 6,000 PDB structures.

The ensembles may include structures generated across combinations of:

Prediction seeds
AlphaFold2 model configurations
Ranking outputs
Recycle stages
Final prediction structures
Intermediate recycle snapshots

Original PDB filenames have been retained whenever possible.

The filenames may encode information such as:

Channel and variant
Prediction condition
Seed
AlphaFold2 model configuration
Rank
Recycle stage
Final or intermediate prediction status

Original filenames should be preserved when downloading or reorganizing the dataset because they are used as identifiers in associated analysis tables.

Known ensemble-count exceptions

Two recognized AlphaFold2 conditions differ from the typical 6,000-structure ensemble size:

Channel	Variant	Condition	PDB structures	Explanation
Kv2.1	L403A	vanilla	5,994	Six expected structures were not present in the source directory at deposition
Nav1.5	WT	masked_test	1,200	Exploratory targeted-masking test ensemble

These ensembles were retained as they existed in the source project rather than excluded or artificially completed.

Users should account for differences in ensemble size when comparing structural distributions across conditions.

Associated analysis tables

Selected analysis tables may be deposited under:

analysis/
├── distances/
│   ├── Kv2.1/
│   ├── Nav1.5/
│   └── Cav1.2/
│
├── rmsd_convergence/
│   ├── Kv2.1/
│   ├── Nav1.5/
│   └── Cav1.2/
│
└── subsets/
    ├── all_ok/
    ├── earliest_converged/
    └── first100/

Only curated analysis tables used for the associated study should be treated as final.

Historical, exploratory, test, or intermediate tables may be deposited separately and identified explicitly.

Structural-distance tables

Structural-distance tables contain geometric measurements calculated from individual predicted structures.

Depending on the channel and analysis, these tables may include:

Inter-residue distances
Cα-to-Cα distances
Minimum heavy-atom distances
Pore-related structural measurements
Voltage-sensor-related measurements
Activation-gate measurements
Inactivation-related measurements
Model identifiers
Source PDB filenames

The exact definitions of individual columns should be obtained from the accompanying analysis code, table-specific documentation, and associated manuscript.

Recycle-convergence quality control

AlphaFold2 structures were evaluated for structural convergence across recycle stages using Cα RMSD calculated over predefined stable-core selections.

The convergence analysis was designed to identify prediction trajectories whose channel core had stabilized across recycle stages.

Structural-distance measurements used for biological interpretation were not used to determine whether a model passed recycle-convergence quality control.

Quality-control and subset tables may contain the following flags.

`all_ok`

Identifies every acceptable model that passed the defined recycle-convergence criteria.

`earliest_converged_selected`

Identifies one earliest acceptable converged model from each model-seed prediction trajectory.

This subset reduces repeated representation of later recycle snapshots from the same prediction trajectory.

`first100`

Identifies models belonging to the first 100 generated models for the corresponding ensemble.

This subset refers to generation order and does not represent the top 100 ranked structures.

Model-level indexing

Where provided, model-level metadata tables relate each PDB filename to attributes such as:

Channel
Variant
Prediction condition
Prediction method
Seed
Model configuration
Recycle stage
Rank
Original filename
Recycle-convergence status
Controlled-subset membership
Associated structural-distance table

Intended uses

This dataset is intended for research involving:

Predicted protein structural ensembles
Voltage-gated ion channel structure
Structural effects of protein mutations
Targeted MSA manipulation
Alternative conformational-state sampling
Comparison of vanilla and masked prediction workflows
Evaluation of structural diversity across prediction conditions
Development of structural-ensemble analysis methods
Reproduction or extension of analyses from the associated manuscript
Benchmarking approaches for filtering and comparing large predicted ensembles

Out-of-scope interpretations

The deposited structures should not be interpreted as:

Experimentally determined structures
Direct measurements of physiological state populations
Direct measurements of thermodynamic probabilities
Proof that a predicted conformation is populated in vivo
Standalone evidence of channel function
Standalone evidence of variant pathogenicity
Clinical or diagnostic evidence
A replacement for electrophysiology, molecular dynamics, biochemical experiments, or experimental structural biology

Limitations

All structures in the current release are computational predictions.

The conformational diversity observed in a predicted ensemble can depend on:

Sequence preparation
MSA construction
MSA depth and composition
Masking strategy
Prediction method
Model configuration
Random seed
Recycle count
Software version
Ranking procedure
Model filtering
Ensemble size

Differences between masked and vanilla ensembles may reflect changes in the information supplied to the prediction model and should not automatically be interpreted as thermodynamic shifts.

AlphaFold confidence metrics estimate confidence in a prediction. They do not establish that a conformation is physiologically populated and do not provide a thermodynamic probability for that state.

Ensemble sizes are not identical for every condition. Analyses comparing conditions should use appropriate sampling controls, normalization, or matched subsets.

Some exploratory conditions have been retained for reproducibility and are identified through their folder and metadata names.

File formats

Structural models are stored as:

.pdb

Metadata and analysis tables may be provided as:

.tsv
.csv
.csv.gz
.parquet

Compressed tabular files should be decompressed or read using software that supports gzip-compressed CSV input.

Downloading the data

Users may download the complete repository or select only the channel, variant, condition, or analysis tables needed for their work.

Because the repository contains a large number of structural files, selective downloading is recommended when only a subset of conditions is required.

Users should record the exact repository revision or commit used for an analysis because the dataset may be expanded with additional prediction methods, metadata, and analysis tables.

Reproducibility and analysis code

Analysis code and notebooks associated with this project are maintained in the GitHub repository:

https://github.com/adrishg/vgci_mutants

The analysis repository contains or will contain workflows for:

Structural-distance calculations
Recycle-convergence RMSD analysis
Quality-control filtering
Selection of controlled model subsets
Ensemble comparisons
Statistical summaries
Figure generation

Exact model-generation settings, stable-core definitions, masking definitions, quality-control thresholds, and figure-specific procedures are documented in the associated code, notebooks, and manuscript.

Dataset maintenance

The repository may be extended with:

Additional AlphaFold2 ensembles
AlphaFold3 ensembles
ESMFold ensembles
Model-level metadata
Curated structural-distance tables
Recycle-convergence manifests
Controlled-subset manifests
Updated dataset documentation
Links and citation information for the associated manuscript

Future structural models will retain the general organization:

channel → variant → prediction method and condition

Ethical and responsible use

This dataset contains computational predictions of protein structures and does not contain human participant information or personally identifiable information.

Researchers should avoid presenting predicted structural differences as experimentally confirmed mechanisms without appropriate validation.

Any biomedical interpretation should account for the limitations of computational structure prediction and the absence of direct functional evidence in the structural models alone.

License

A formal reuse license has not yet been finalized for this dataset.

Users interested in redistribution, incorporation into another dataset, or extensive reuse should review the current repository license status and contact the dataset maintainers when clarification is required.

The finalized license will be added to the Dataset Card metadata when available.

Citation

A formal manuscript citation will be added when the associated preprint or publication becomes publicly available.

Until then, users should cite:

VGIC Mutant Structural Ensembles. Dataset generated for the manuscript “Targeted MSA Masking Reshapes Structural Sampling in Mutant Voltage-Gated Ion Channels.” Hugging Face Datasets, repository adrishgz/vgic-mutant-structural-ensembles.

Users should include the access date and exact repository revision or commit used in their analysis.

Contact

Questions about the dataset, prediction conditions, structural measurements, or associated analyses may be submitted through the Hugging Face dataset discussion page or the associated project repository.

Downloads last month: 297

Total file size:

159 GB