An in-silico test of whether NKG7, a four-transmembrane granule protein of cytotoxic lymphocytes (NK and CD8⁺ T cells), forms an ion channel. The work combines structural-homology search, three independent structure predictors (AlphaFold3, Boltz-2, Chai-1) and reference-HOLE pore analysis, run as a tiered campaign where each stage gates the next.
NKG7 has no documented channel activity, yet loss of NKG7 reduces Ca²⁺ influx in CD8⁺ T cells, and its fold resembles the Ca_V-γ (CACNG) auxiliary subunits of calcium channels. Is NKG7 itself a Ca²⁺-permeable channel, or a channel-family lookalike that acts as a scaffold or auxiliary modulator? Both readings were fixed before any analysis was run, and every stage tests both.
NKG7 is a genuine member of the claudin / PMP-22 / Ca_V-γ tetraspan superfamily and it self-associates, but it does not form a self-contained ion channel. Across three independent structure predictors and five oligomeric states, no confident assembly encloses a conducting pore. The reproducible self-interface it does form is a lateral TM1–TM1 helix-packing contact, the hallmark of a scaffold or auxiliary modulator. This matches its closest relatives (non-conducting Ca_V/TARP-γ auxiliary subunits and GSG1L) and its known role inhibiting the v-ATPase to restrain mTORC1.
a) No oligomer order is confidently supported, and the three predictors disagree on the preferred stoichiometry. b) Pore-opening anti-correlates with confidence: every confident model is occluded (sub-water radius), and the only open pores appear in low-confidence models.
Stages run from cheap to expensive, and a negative result stops the next stage. Full rationale
is in docs/methods_and_decisions.md.
| Tier | Question | Methods | Outcome |
|---|---|---|---|
| 0. Triage | What is NKG7? Fold, relatives, transcripts, domains. Does the monomer conduct? | Foldseek, BLAST / Ensembl Compara / phmmer, GTEx, InterProScan, monomer pore scan | Escalate. A bona-fide tetraspan, but every analogue is non-conducting and the monomer is occluded (as expected). |
| 1. Oligomer | Does an oligomer enclose a pore? | Boltz-2, AlphaFold3 and Chai-1 for C2–C6, plus reference HOLE | Stop the channel track. Best interface ipTM is 0.63 (confident bar ≈ 0.8); every confident model is occluded. |
| Pivot: hetero-complex | What does NKG7 dock onto? | The same three predictors with ATP6AP2, ORAI1 and STIM1 | Inconclusive. No confident interface, and the ATP6AP2 positive control also fails, so binary co-folding is under-powered here. |
- No confident oligomer. The best interface confidence anywhere is ipTM 0.63 (Boltz-2 dimer). The predictors disagree on both stoichiometry and confidence, which is the signature of no well-defined assembly.
- Pore-opening anti-correlates with confidence. Every model a predictor is confident about is occluded. The only open models are the lowest-confidence ones, with 60–115 Å pore axes: splayed aggregates, not a compact ~40 Å transmembrane channel.
Wet-lab context (western-blot multimer bands, NKG7 self-coprecipitation) shows NKG7 does self-associate, and the predictors corroborate a reproducible dimer interface. But it is a lateral helix-packing contact, which is why the pore stays shut. Oligomerisation is supported; pore formation is not.
The best cross-chain ipTM is 0.59 (AlphaFold3, STIM1). The ATP6AP2 positive control, a known binder, scores only 0.10–0.33, so the ORAI1 and STIM1 negatives are inconclusive, not evidence of no interaction. Suggested follow-up: higher-order co-folds (NKG7 with an ORAI1 hexamer) and experimental partner discovery (co-IP or proximity-labelling mass spectrometry).
All inputs are public. Nothing heavy is stored in this repository.
| Resource | Used for |
|---|---|
| UniProt Q16617 (human) and Q99PA5 (mouse) | Target sequences (targets/) |
AlphaFold DB v6, AF-Q16617-F1, AF-Q99PA5-F1 |
Monomer models for Tier 0 |
| Foldseek against PDB100 and AFDB/SwissProt | Structural neighbours |
| NCBI BLAST, Ensembl Compara, EBI phmmer and InterPro | Sequence relatives, paralogs, domains |
| GTEx v8 | Tissue expression |
| UniProt O75787 (ATP6AP2), Q96D31 (ORAI1), Q13586 (STIM1) | Hetero-complex partners |
| Boltz-2, AlphaFold3 3.0.1, Chai-1 0.6.1 | Structure prediction |
| MDAnalysis 2.10 + mdahole2 0.5 (reference HOLE) | Pore-radius profiles |
- All structures are predictions. No experimental NKG7 structure exists.
- Co-folding models are trained largely on soluble complexes, and ipTM cut-offs (≈0.5 floor, ≈0.8 confident) are conventions, not calibrations for membrane homo-oligomers.
- Static models cannot capture gating, so the conclusion rests on low confidence and occlusion together, not on occlusion alone.
- Chai-1 ran without ESM embeddings (cached ColabFold MSA), which mainly affects per-residue accuracy, not interface confidence.
- The Tier-0 monomer scan used a geometric probe-sphere method equivalent to HOLE. All Tier-1 oligomers were scanned with reference HOLE.
- Results are computational. Channel function is defined electrophysiologically, so a positive result would need patch clamp, bilayer reconstitution or Ca²⁺-flux imaging.
Structure prediction ran on a SLURM GPU cluster (one A40 per predictor job). Tier 0 uses public
web services. See envs/README.md for setup and
pipeline/README.md for the exact commands and run order.
conda env create -f environment.yml # pore-scan environment
conda activate nkg7-structure
python pipeline/tier1/hole_scan.py --models-dir results/tier1_models --out-dir hole_outThe helper functions (symmetry-axis detection, pore verdict, AlphaFold3 input builder) have unit
tests that run without HOLE or a GPU: pip install pytest numpy MDAnalysis && pytest.
The scan re-runs reference HOLE on the 15 committed Tier-1 models. It needs the hole binary on
PATH (or --hole-exe), and its output corresponds to
reports/data/hole_results.json. Regenerating the models
themselves needs the three predictors and a GPU.
docs/ methods_and_decisions.md: design, gates and rationale
pipeline/ predictor inputs (tier1/, hetero/) and the HOLE scan and AF3 helper scripts
envs/ environment notes for each tool
targets/ NKG7 human and mouse sequences
reports/ three self-contained HTML reports, plus data/ (every table and figure behind them)
results/ the 15 Tier-1 oligomer models, hetero-complex models and the dimer-interface consensus
The three reports (open in any browser) are
Tier 0 triage,
Tier 1 oligomers and
hetero-complex screen.
nkg7-binder-design applies a
de-novo mini-binder design pipeline to the extracellular loops of the same target.
See CITATION.cff. Released under the MIT License.