Reading the direction of a cell's machinery, not the amount
From gene expression alone, reading the direction of a cell's mitochondrial program — is it
building or clearing — predicts which cancers are addicted to their mitochondria
and which respond to OXPHOS drugs. Direction beats amount, and no single molecular layer sees it.
A reusable directional scorer, validated across 1,066 cancer cell lines, a 1,514-drug library,
and independent molecular layers — with every claim tested against its own confound, and every null reported at power.
The idea
Amount is blind to direction
Two drugs can raise the same organelle by the same amount — one by building more of it, one by
blocking its cleanup. A standard readout of how much is present calls them identical. They are opposites.
The directional score reads that difference. It combines a biogenesis signal and a selective-degradation
signal into a single vector — net direction = biogenesis − selective degradation — computed from
transcriptomes, proteomes, or phosphoproteomes. Positive means the cell is net-building an organelle; negative
means it is net-clearing it. The selective-degradation arm is populated only where the biology provides a
dedicated receptor (e.g. BNIP3/NIX for mitochondria), and is left empty otherwise.
Two points of precision, because they are what a careful reader asks first. First, the score is read from a
single expression snapshot — it is a signed set-point (which program dominates now), not a measured
rate; we validate that this static readout tracks true direction by sampling it along known timecourses
(lysosomal-damage washout, iPSC→cardiomyocyte differentiation, exercise). Second, the blind spot in
conventional analysis is not an inability to read genes up or down — it is that rolling per-gene changes up
into an organelle amount (an abundance signature, a mass stain, a pathway-enrichment score) averages
building-more and blocking-removal into the same total, and the selective-clearance genes that would separate
them are not in the standard organelle gene set at all. The clearest proof in our own data: in kidney cancer,
BNIP3 protein rises while its phospho-regulated occupancy collapses (g=−1.90) — amount says one
thing, direction says the opposite.
Figure 1.The metric reads a vector, not a level. Two perturbations that change an organelle's abundance in the same direction can be driven by opposite programs (build vs. block-cleanup); a single-layer abundance readout cannot separate them, whereas the composite direction score does.
The finding
Direction predicts genetic vulnerability
Across 1,066 cancer cell lines (DepMap CRISPR), the higher a cell line's mitochondrial
direction score, the more addicted it is to its own mitochondrial machinery — losing those genes
kills it.
Mitochondrial direction → mean CRISPR dependency on 83 machinery genes:
ρ = −0.35, p = 1.7×10⁻³², n = 1,066. Survives a global-dependency-burden control
(partial r = −0.32) and is negative in 17 of 18 lineages — strongest in solid tumors
(brain −0.57, kidney −0.51), weakest in blood. It is cell biology, not a tissue-type shortcut.
Figure 2.Higher direction, stronger mitochondrial addiction. Per-cell-line direction score vs. mean CRISPR gene-effect on mitochondrial machinery (a; lower gene-effect = more essential). Direction predicts dependency in a mitochondria-specific way (b): mitochondrial direction tracks mitochondrial dependency (−0.35) far more than ribosome dependency (−0.07), and ribosome direction does not predict its own dependency (−0.02). The −0.26 in the lower-left is the expected reflection that highly anabolic cells are broadly mitochondria-reliant — shown, not hidden — and it does not carry the diagonal. The top vulnerability genes are the correct biology — SDH complex, mtDNA replication (TWNK/POLG), OXPHOS assembly (c).Figure 3.Every practical claim, confound-controlled. Genetic dependency is lineage-robust (a). Mitochondria-targeting drugs stay at the sensitized tail after removing the general drug-sensitivity axis (b). The directional score beats raw expression on the vulnerability endpoint (c, AUROC 0.67 vs 0.61). Summary of the four tests (d).
The translational payoff
…and which drugs will work
Given nothing but expression, the score ranks the clinical Complex I inhibitor
IACS-010759 to the 3rd percentile of a 1,514-drug library, and puts
MitoQ — a mitochondria-targeted drug — at #1.
IACS-010759 sensitivity vs. direction: ρ = −0.15, p = 8×10⁻⁶, n = 872.
The signal survives the proliferation confound: after removing each cell line's general drug-sensitivity,
mitoquinone remains the single most direction-sensitized drug of 1,514, and the mito-drug class stays
enriched (Mann-Whitney p = 0.017).
Figure 4.The score ranks mitochondria-targeting drugs to the sensitized tail. Higher-direction cell lines are more sensitive to IACS-010759 (a). Across 1,514 drugs, the mitochondria-targeting compounds cluster at the sensitized extreme (b), with mitoquinone ranked first (c).
Direction beats amount — validated on the endpoint that matters: which cells die.
Generality
The same engine works across organelles and layers
The framework is not a mitochondrial one-off. At the genetic level it generalizes with a
graded, biologically-sensible pattern — and it corroborates across independent molecular layers.
Figure 5.Generalizes genetically, graded by organelle. Direction predicts genetic vulnerability strongly for mitochondria, modestly for ER (survives burden and lineage controls), and null for lysosome (a) — the buffered organelle correctly shows no coupling. At the drug level the cross-organelle test is null and reported as such (b): the honest boundary of the generalization.Figure 6.Cross-modal agreement in human tumor tissue. In kidney cancer (CPTAC ccRCC, 110 tumors within a 185-sample paired cohort), the mitochondrial direction signal agrees between RNA and protein — the same directional call from two independent molecular layers, retiring the mRNA-proxy caveat.
The part most teams skip
The method refereed its own science
The most important result is one we failed to find — and reported anyway. The analysis
adversarially audited itself at every turn.
What it does not do — at power
In kidney cancer, an underpowered hint suggested the direction score might predict survival
(CPTAC Cox p=0.04, 21 deaths). Tested at 8× the power in TCGA-KIRC (508 tumors, 168 deaths), it
does not — a clean null. We report it as a null, not a headline.
Figure 7.A clean null at power. Mitochondrial direction does not track tumor stage (a) and does not stratify overall survival (b) in 508 kidney-cancer patients — the survival curves are superimposed (log-rank p=0.65). Positive controls confirm the pipeline is sound (stage HR=1.93, p=5.6×10⁻²¹). The method reads biology, not prognosis.
This was not the only self-correction. The analysis also demoted an anti-conservative statistic it had
computed, refuted an earlier "coupling" headline as a shared-timecourse artifact, downgraded a promising
second-platform drug "replication" once it found the confound, and hash-locked nine blind cancer predictions
before seeing the data — reporting the misses. The tool tried to disprove itself, and said so.
Reproducibility. Directional scorer + gene modules, all datasets public (DepMap 24Q2 CRISPR + PRISM,
TCGA-KIRC via cBioPortal, CPTAC ccRCC). Every headline number traced to a single canonical source; every
confound control and null included by design, not omission.