Two cells can hold the same number of mitochondria for opposite reasons — one is building them, the other has stopped clearing them. Count how many are there and they look identical. Read the direction instead, and you can predict which cancers are addicted to their mitochondria, and which drugs will hit them.
▶ Open the interactive demo · Read the report · Full technical record
A standard organelle signature answers how much is there. But build-up and tear-down move many of the same genes, so a single-axis score cannot tell a cell that is proliferating mitochondria from one whose disposal machinery has stalled. To the cell those are opposite states. Worse, the genes that would separate them — the selective-autophagy receptors — are usually not in a conventional organelle signature at all.
So we score the two opposing programs separately and subtract them:
net_direction = biogenesis − selective degradation
One signed number per sample, read from ordinary transcriptomes, proteomes, or phosphoproteomes. Positive means net building; negative means net clearing.
It is a set-point — which program dominates now — validated against timecourses where the true direction is known. It is not an organelle headcount, and not a measured turnover rate.
| Result | Number | Data |
|---|---|---|
| Direction → genetic vulnerability | ρ = −0.35, p = 1.7×10⁻³², n = 1,066 cell lines | DepMap CRISPR |
| ↳ survives dependency-burden control | partial r = −0.32 | |
| ↳ lineage-robust | negative in 17/18 lineages | |
| ↳ mitochondria-specific | mito→mito −0.35 vs mito→ribosome −0.07 | |
| Direction → drug response | IACS-010759 at 3rd percentile of 1,514 drugs; MitoQ #1 | DepMap PRISM |
| ↳ survives a proliferation confound | mito-drug class MWU p = 0.017 | |
| Direction beats amount | AUROC 0.666 (direction) vs 0.613 (raw expression) | same vulnerability endpoint |
| Generalizes, graded by organelle | mito −0.35 / ER −0.15 (survives controls) / lysosome null | DepMap |
| Cross-modal agreement in human tumour | mito-biogenesis RNA g = −1.52 · protein g = −3.14 | CPTAC ccRCC |
Nothing told the score what those drugs or genes do. It ranked them from organelle biology alone.
Effect sizes are modest (ρ 0.15–0.35; AUROC 0.60–0.67) — as expected for a single-pathway expression score predicting a functional phenotype across highly heterogeneous cell lines. The value is a real, specific, lineage-robust, correctly-signed signal, not a high-accuracy point predictor.
Every headline above survived an adversarial check. The ones that didn't survive are reported too — that is the point.
- Reported a null at power. A weak survival hint in CPTAC kidney cancer (Cox p = 0.04, 21 deaths) did not replicate when retested at 8× the power in TCGA-KIRC (508 tumours, 168 deaths; Cox p = 0.21, log-rank p = 0.65). The flat survival curves are in the report; positive controls (stage HR = 1.93, p = 5.6×10⁻²¹) confirm the pipeline works.
- Recorded its predictions up front. Nine directional predictions were written down and
SHA-256-hashed at the start of the validation work (
data/PREREGISTRATION.md). 4/9 hit — including a high-confidence miss (COAD) that was recorded and reported as a miss, which is what makes the hits credible rather than fitted. The hash proves the prediction file was never edited afterward; it cannot by itself prove ordering, since no external timestamp was anchored during the hackathon. Read it as a documented commitment, not cryptographic pre-registration. - Refuted its own headline. A striking r = 0.88 organelle "coupling" was flagged as a shared-timecourse artefact and re-led with the mechanism that survived (TFEB-dependence).
- Walked back a replication claim. A second drug platform (GDSC) looked like independent replication; the confound test showed it was mostly a general drug-sensitivity axis, so it was downgraded to "weak directional consistency." PRISM survives the same test; GDSC does not.
- Demoted its own statistic. Swapped an over-generous equal-variance test (p = 0.0005) for the correct Welch test (p = 0.0094) — kept the weaker, right number.
Requires Python ≥ 3.9 with pandas and numpy.
git clone https://github.com/different-change/organelle-direction
cd organelle-direction/scorer
# 1. confirm every shipped gene module loads
python run_dynamics.py --self-check
# 2. score your own matrix (genes × samples, gene IDs in the first column)
python run_dynamics.py --expression your_matrix.csv \
--organelle mitochondrion --organism human --out scores.csv--self-check prints the full module grid:
organelle species biogenesis degradation selective
-------------------------------------------------------------
mitochondrion human 83 20 6
lysosome human 79 29 6
er human 73 25 7
peroxisome yeast 30 20 1
... (16 modules)
Or use it as a library:
from score_organelle_dynamics import score_organelle_dynamics
from run_dynamics import load_module
bio, deg, sel = load_module("human", "mitochondrion")
scores = score_organelle_dynamics(expr, bio, deg,
degradation_selective_ids=sel,
method="rank", n_perm=1000)
scores[["biogenesis_score", "degradation_score", "net_direction"]]The score is rank_pct(biogenesis genes).mean − rank_pct(selective-clearance genes).mean per sample,
with an empirical permutation null. method="rank" works on a single sample — no cohort required.
Seven organelles across four species, every gene resolver-backed with provenance in
data/genesets/:
| yeast | Arabidopsis | rice | human | |
|---|---|---|---|---|
| peroxisome | ✅ selective | ⬜ no receptor | ⬜ no receptor | |
| mitochondrion | ✅ selective | ⬜ no receptor | ⬜ no receptor | ✅ selective |
| ER | ✅ selective | ⬜ no receptor | ⬜ no receptor | ✅ selective |
| chloroplast | ⬜ effectively empty | ⬜ no receptor | ||
| lysosome | ✅ selective | |||
| lipid droplet | ⬜ no receptor | |||
| ferritin / iron | 🔶 partial |
The clearance arm is populated wherever a genuine selective-autophagy receptor exists (yeast Atg36,
Atg32/33, Atg39/40; human PINK1/Parkin/NIX/BNIP3/FUNDC1, the lysophagy galectins, the FAM134B ER-phagy
family) and empty where none is known — every plant organelle, and the human lipid droplet, where
lipophagy is bulk autophagy with no LD-selective receptor. That pattern is biology, not a scorer defect,
and the tool reports it instead of hiding it. Details and the honest partials in
scorer/README.md.
The engine is organelle- and trait-agnostic: drop in a new pair of gene-set CSVs with the same columns and it scores that instead.
index.html ← landing page (published via GitHub Pages)
demo/index.html ← interactive walkthrough; press P for Presenter Mode
report/
report.html ← the judging document: 7-beat spine, standalone-readable figures
full_technical_report.html ← complete technical record (35 figures, all analyses)
figures/ ← the six 300-DPI figures, each readable cold
scorer/
score_organelle_dynamics.py ← the organelle-agnostic scoring engine (pure function, permutation null)
run_dynamics.py ← module loader + command-line interface
README.md ← method, module grid, validation, honest limits
data/
genesets/ ← 16 curated biogenesis / selective-clearance modules + provenance
scores/ ← the result tables behind every figure
prospective_predictions.* ← the hash-locked blind predictions
PREREGISTRATION.md ← human-readable pre-registration record
- mRNA is a proxy, not an organelle count. Read a score as "consistent with a shift toward biogenesis / clearance," never as a turnover rate. Corroborate against proteome or imaging where possible.
- The biogenesis arm is high-confidence (conserved machinery, clean signal); the clearance arm is weaker, and in plants it is bulk-autophagy-confounded by construction.
- Ferritin's clearance arm is not mRNA-validatable — NCOA4 is controlled post-translationally, so its transcript stays flat even when ferritinophagy fires. Reported as a partial, not a result.
- Cross-taxa rule: machinery conserves, regulators diverge. Only conserved machinery was mapped by orthology; lineage-specific regulators were dropped.
- Kidney-cancer survival is a null, at power. It is in the report as prominently as the wins.
DepMap 24Q2 (expression, CRISPR gene-effect, PRISM drug response) · TCGA-KIRC (cBioPortal) · CPTAC ccRCC (RNA, protein, phosphoproteome) · MoTrPAC · AtGenExpress · NCBI GEO series (accessions in the technical report).
No proprietary or unpublished data is used anywhere in this repository.
The entire pipeline — gene-set curation, the scoring engine, every confound control, all figures, both reports, and the adversarial self-refereeing above — was built in Claude Science, with a scientist in the loop at every decision.
That last clause is the load-bearing one. The rigour is reproducible by anyone with the same tools; what is not is the biology judgment that shaped the method — splitting biogenesis from clearance in the first place, and knowing that the selective arm should only be populated where a real receptor exists. The tool works precisely where domain knowledge was injected, and returns honest emptiness where it wasn't.
Built for Built with Claude: Life Sciences (Anthropic × Cerebral Valley × Gladstone Institutes), Research track, 7–13 July 2026.
The method operationalises the organellomics framework — organelle abundance and turnover as a readout of cellular stress resilience:
Hickey K, Nazarov T, Smertenko A. (2023) Organellomic gradients in the fourth dimension. Plant Physiology 193(1):98–111. https://academic.oup.com/plphys/article/193/1/98/7181000
If you use the code, the gene modules, or the results, cite via CITATION.cff —
GitHub's "Cite this repository" button will format it for you.
Code is MIT. Gene modules, result tables, figures and reports are CC BY 4.0. Underlying primary data remain under their original providers' terms.