the problem
Same amount. Opposite fate.
Two cells can hold the same number of mitochondria — but one is busy building them, the other tearing them down. Measure the amount — an abundance signature, a mito stain — and they're identical. The two opposite programs that produced it average into one number.
two cells · equal mitochondriatry the toggle ↓
Measuring the amount, the two cells are indistinguishable.
the idea
Read the direction, not the amount.
Score the two opposing programs separately — building (biogenesis) and tearing down (selective clearance) — then subtract. One signed number whose sign is the biology: a set-point, not a rate, validated against timecourses where the true direction is known.
Drag the two forces and watch the needle. Conventional organelle signatures score only the build arm — the clearance genes (BNIP3/NIX) aren't even in them, so they can't read the sign.
net direction = building − clearing
net = +35 · building up
the proof
Direction beats amount.
On the task that matters — spotting the cells most vulnerable to losing their mitochondria — reading the signed direction outperforms simply measuring how much is there.
AUROC measures how well a score separates vulnerable cells from the rest. 0.50 is a coin flip; 1.00 is perfect.
flagging the most vulnerable cells · AUROC
bars start at 0.50 — the coin-flip floor
a finding · a target
Cells building mitochondria can't live without them.
Across 1,066 cancer cell lines, the more a cell's direction says "building mitochondria," the more it dies when that machinery is switched off — a targetable weakness read from expression alone.
is it just "you need what you make"? — a control
If the effect were just "a cell needs whatever it makes," ribosomes would show it too. They don't — the link is specific to mitochondria. Real addiction, not an artifact.
a finding · the payoff
It flags which cells a mitochondria-drug will kill.
Given only expression, the score ranked the clinical Complex-I inhibitor IACS-010759 into the top 3% most-affected of 1,514 drugs — and put MitoQ, a mitochondria-targeted drug, at #1 in the whole library.
1,514 drugs, ranked by predicted response
Each faint dot is one of 1,514 drugs. The two mitochondria-targeted drugs land at the sensitized edge — exactly where the biology predicts.
the part competitors won't have
We told Claude Science to try to prove us wrong.
The differentiator isn't the pipeline — it's the skepticism. Claude adversarially refereed its own work instead of chasing a positive result:
Demoted its own statistic
Swapped an over-generous test (p=0.0005) for the stricter, correct one (p=0.0094) — chose the weaker, right number.
Reported a null at power
A 510-tumor survival test (survival on the 508 with follow-up) — 8× the earlier hint. It didn't replicate. Reported, not buried.
Refuted its own headline
Flagged its own r=0.88 "coupling" as a shared-timeline artifact and re-led with the real mechanism.
Downgraded its own result
Found a second-platform drug hit, tested the confound itself, and walked the claim back when it didn't hold.
The method reported its own null at power — and walked back three more claims on its own. Every negative shown, not buried.
in one sentence
Read the direction of a cell's mitochondrial program — building up or clearing out — and, from expression alone, you can tell which cancer cells are addicted to their mitochondria and which respond to the drugs that target them.
Built with Claude Science · public data only · Research track
