Organelle Direction

Identical counts, opposite states. The count is what a conventional signature reports; the direction is what the cell is actually doing — and in 1,066 cancer cell lines it is Cell A, the one still building, that dies when you switch its mitochondrial machinery off.

The one idea

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 — and the genes that would separate them, the selective-autophagy receptors, are usually not in the signature at all.

So we score the two opposing programs separately and subtract them:

biogenesis − selective degradation = net direction

One signed number per sample, from an ordinary transcriptome, proteome, or phosphoproteome. Positive means net building; negative means net clearing. It is a set-point — which program dominates right now — validated against timecourses where the true direction is known. It is not an organelle headcount and not a measured rate.

What direction buys you

ρ = −0.35

Predicts genetic vulnerability

1,066 cancer cell lines, DepMap CRISPR. p = 1.7×10⁻³²; negative in 17 of 18 lineages.

#1 of 1,514

Picks the drug, unsupervised

MitoQ ranks first across the PRISM library; IACS-010759, in clinical trials, lands in the top 3%.

0.666 vs 0.613

Direction beats amount

AUROC on the same vulnerability endpoint — signed direction against raw expression.

Scatter of net direction against CRISPR gene-effect across 1,066 DepMap cancer cell lines, with lineage and burden controls. — The more a cell line is building mitochondria, the harder it dies when you switch that machinery off. Nothing about drug or gene function was given to the score.

Where it holds, and where it doesn't

Claim	Result	Data
Direction → genetic vulnerability	ρ = −0.35, n = 1,066	DepMap CRISPR
↳ survives dependency-burden control	partial r = −0.32
Direction → drug response	mito-drug class MWU p = 0.017	DepMap PRISM
Cross-modal agreement in human tumour	RNA g = −1.52 · protein g = −3.14	CPTAC ccRCC
Graded by organelle, as biology predicts	mito −0.35 · ER −0.15 · lysosome null	DepMap
Survival in kidney cancer — reported null	Cox p = 0.21 at 8× the power	TCGA-KIRC, 508 tumours

Effect sizes are modest, as expected for a single-pathway expression score predicting a functional phenotype across heterogeneous cell lines. The value is a real, specific, lineage-robust, correctly-signed signal — not a high-accuracy point predictor.

Things we tried to break first

Reported a null at power. A weak survival hint in CPTAC kidney cancer did not replicate in TCGA-KIRC at eight times the power. The flat curves are in the report; positive controls (stage HR = 1.93) confirm the pipeline works.
Pre-registered the predictions. Nine blind directional calls were committed and SHA-256-hashed before the data was touched. Four hit — including a high-confidence miss locked in beforehand, which is what makes the hits credible.
Refuted its own headline. A striking r = 0.88 organelle “coupling” was flagged as a shared-timecourse artefact and replaced with the mechanism that survived.
Walked back a replication claim. A second drug platform looked like independent replication until the confound test showed it was mostly a general sensitivity axis. Downgraded to “weak directional consistency.”
Demoted its own statistic. Swapped an over-generous equal-variance test for the correct Welch test, and kept the weaker, right number.

Run it on your own data

git clone https://github.com/different-change/organelle-direction
cd organelle-direction/scorer
python run_dynamics.py --self-check
python run_dynamics.py --expression your_matrix.csv \
    --organelle mitochondrion --organism human

Sixteen curated modules ship with it — seven organelles across yeast, Arabidopsis, rice and human, each gene resolver-backed with provenance. The engine is organelle-agnostic: drop in a new pair of gene-set CSVs and it scores that.

The selective-clearance arm is populated wherever a genuine selective-autophagy receptor exists and empty where none is known — including every plant organelle. That pattern is biology, not a defect, and the tool reports it rather than hiding it.