A directional organelle-dynamics score distinguishes build-up from tear-down — across seven organelles

Organellomics as a resilience readout · anchored on the peroxisome, generalized to mitochondrion, ER, chloroplast, lysosome, lipid droplet and ferritin · a 16-cell organelle×species grid (yeast, Arabidopsis, rice, human) · both directional arms validated in human, read jointly across coupled organelles · public data, resolver-backed IDs · and the direction score predicts cancer mitochondrial vulnerability (DepMap, 1,066 lines) and OXPHOS-drug response (PRISM, IACS-010759 at the 3rd percentile), confound-controlled

This report grew from a single peroxisome study into a four-organelle platform; the peroxisome sections below are the anchor, and the later sections show the same engine generalizing unchanged.

What is load-bearing vs. exploratory: the engine transfers across organelles and species unchanged — that is the central claim — but statistical strength varies by cell. The defensible, well-powered results are the Arabidopsis 248-sample stress discrimination (Figs 3, 6), rice chloroplast dehydration (Fig 7), human exercise biogenesis (Fig 10), and the drug-screen de-circularization via urolithin A (Fig 16). The small-n yeast contrasts (Figs 1, 2, 4, 5; n = 2–4/group) are reproducible directions reported as screening results, and the Parkinson's single-cell cell (Fig 15) is exploratory/inconclusive. The honesty is applied consistently, not just in the robustness figure.

The idea in one paragraph

Organelle abundance and turnover are a readout of a cell's stress resilience — the organellomics framing (Hickey, Nazarov & Smertenko, Plant Physiology 2023, 193:98). The obvious way to probe it from expression data — take a textbook list of peroxisome "marker" genes and test whether it goes up under stress — is close to generic gene-set enrichment and, worse, cannot tell a cell that is building peroxisomes up (proliferation) from one that is tearing them down (pexophagy): both light up the same "peroxisome" list. This project scores the direction instead. Two separately-derived gene modules — biogenesis (peroxin / matrix-import / PEX11 proliferation machinery) and degradation (pexophagy / selective-autophagy machinery) — are scored independently per sample, and their difference, net_direction = biogenesis − degradation, gives a signed readout whose sign is the biology.

Design decisions (and why)

QuestionDecision
Which organelle?Peroxisome as the anchor (cleanest, most-conserved machinery; the lead's N-BODIPY abundance turf; independent grounding available), then generalized to mitochondrion, ER and chloroplast once the method was validated. Depth on one well-grounded organelle first, then breadth to test that the engine transfers.
What is the signal?Directional dynamics, not "activity up." A falsifiable, signed score (biogenesis − degradation) per sample, with a peroxisome-selective degradation sub-score kept separate from bulk-autophagy machinery.
How are gene sets defined?Ontology-derived, resolver-backed. GO/Reactome membership seeded and curated; every ID resolved through MyGene/UniProt/BioMart/Ensembl — none hand-typed. Symbol collisions caught and recorded (see provenance).
Built-in ground truth?Yeast positive controls with known direction: oleate → peroxisome proliferation (biogenesis must rise); rapamycin/starvation → autophagy (degradation must rise). Failure to recover these = method failure.
Cross-taxa?Yeast anchor → Arabidopsis transfer. Only conserved machinery mapped by orthology; lineage-specific regulators (Oaf1/Pip2/Adr1) dropped — machinery conserves, regulators diverge.
Deliverable?Research finding + reusable scorer. The validated finding is the credibility; the packaged score_organelle_dynamics makes it reusable.

Phase 0 — resolver-backed gene modules provenance

Four gene-set files (biogenesis + pexophagy × yeast + Arabidopsis), each row carrying the resolved systematic ID (SGD systematic name for yeast, AGI locus for Arabidopsis), UniProt ID, source GO terms, submodule, and a shared_with_bulk_autophagy flag. Collision handling was real, not decorative — examples recorded in geneset_provenance.json:

Phase 1 — the score recovers known yeast biology validation

Both positive controls passed on public, replicated data (n≥2 per condition).
Yeast validation: oleate raises biogenesis, rapamycin raises degradation
Figure 1. Directional score recovers known yeast biology. (a) Shifting yeast from glucose to oleate raises the biogenesis score (GSE5862). Two distinct tests: a gene-level competitive test (biogenesis-module genes vs background log2FC) is significant, Mann-Whitney p = 3.5×10−3, with canonical PEX5/PEX11/PEX25/PEX27 among the up genes; the sample-level score test shows a large effect (Cohen d = 1.17) in the correct direction but is underpowered at n = 4 (p ≈ 0.15). The two statistics come from separate tests and are not paired. (b) Rapamycin (TORC1 inhibition → autophagy) raises the degradation score 0.31→0.46 (Welch p = 0.005) and the peroxisome-selective ATG36 score 0.25→0.51 (p = 5×10−4); GSE149016. (n = 2–3/group: these p-values are screening statistics — the direction is reproducible and survives leave-one-driver-out, but defensible significance awaits higher n; see Fig. 19.)
ControlDataset (n)ExpectedResult
A — oleateGSE5862 (4 vs 4)biogenesis ↑PASS — gene-level competitive MWU p = 3.5e-3; sample-level effect d = 1.17 (same direction, underpowered at n=4, p≈0.15) — separate tests
B — rapamycinGSE149016 (3 vs 5)degradation ↑PASS — degradation p = 0.005; selective ATG36 p = 5e-4; net flips negative
B′ — N-starvation 1 hGSE151898 (2 vs 2)degradation ↑did not show it — 1 h is too early for transcriptional pexophagy; reported, not spun

Grounding — the direction holds at the protein level, cross-modality grounding

mRNA is a proxy. To move beyond it, the same generic scorer was applied to an independent quantitative proteome — a whole-yeast label-free mass-spectrometry dataset of oleate vs glucose (PRIDE PXD069526, Das et al. 2026, n = 4/4). This is a different molecule, a different lab, a different technology — a genuine cross-modality check, not a citation.

Protein-level corroboration: oleate raises the biogenesis protein score
Figure 2. Protein-level corroboration (PXD069526). (a) The biogenesis protein score separates glucose from oleate (0.326 → 0.434; one-sided MWU p = 0.014 at its n = 4 floor; the very large Cohen d = 22 reflects tight replicates, so the independent test in (b) carries the strength of evidence). (b) Competitive test: biogenesis-module proteins shift up against the whole proteome (median log2FC +0.95 vs +0.11 background; MWU p = 2.5×10−9 — a gene-level competitive test, so this is a screening statistic that inherits the same co-regulation anti-conservatism corrected in Fig. 19; the defensible claim is the coherent up-shift of the biogenesis proteins, not the exact p), with Pex11, Pex5, Pex15 and β-oxidation enzymes (Pox1/Fox2/Pot1) among the risers. The degradation score does not rise on oleate (0.362 → 0.343), so the shift is biogenesis-specific, not global up-regulation.

Honest limits: a single published study; proteomics under-samples low-abundance peroxins (the selective receptor Atg36 was not detected, and several peroxins rest on an imputation floor and are directional-only). Higher steady-state biogenesis-machinery protein abundance on oleate is "consistent with a shift toward biogenesis," not a measured turnover rate. Citations: Das et al., Histochem Cell Biol 2026 (PMID 41686312); Yifrach et al., J Cell Sci 2016 (PMID 27663510).

Phase 2 — transfer to Arabidopsis: the pattern separates stresses finding

The same, unchanged scorer and the orthology-mapped Arabidopsis modules were applied to the AtGenExpress abiotic-stress compendium (9 GEO series, 248 samples, all replicated, Affymetrix ATH1; probe→AGI mapping resolved from the platform annotation, not hand-typed). The real test is discrimination: does the directional pattern separate between stresses, or only stressed-from-unstressed?

Arabidopsis transfer: biogenesis score separates stresses
Figure 3. The conserved biogenesis arm discriminates plant stresses. (a) Biogenesis score across 9 stresses (roots; Kruskal-Wallis H = 42, p = 2×10−6). (b) Effect size vs control — heat and osmotic raise the biogenesis score, cold (strongest), salt and drought suppress it; rings mark p < 0.05, filled = roots, open = shoots. (c) The honesty panel: biogenesis (conserved) vs degradation (bulk-autophagy) side by side — the clearance arm rises under osmotic/salt but those are canonical bulk-autophagy inducers, so it cannot be claimed as peroxisome-selective.

The finding: different stresses move peroxisome biogenesis in opposite directions — this is a separating signature, not a stressed-vs-not detector.

Second vertical — mitochondria: closing the weak arm generalization

Is the platform reusable, and does the clearance weakness come from the method or from the biology? The same engine was pointed at a second organelle — the mitochondrion — in yeast, with resolver-backed biogenesis (n = 33: TOM/TIM import, MICOS, fusion/fission) and mitophagy (n = 10) modules. The decisive difference: mitophagy has a genuine selective receptor, Atg32 (with Atg33), whereas plant pexophagy has no ATG36 ortholog. So the mitochondrion is the test of whether the selective-degradation arm works when a real receptor exists.

Mitochondria validation: respiratory biogenesis up, Atg32-selective mitophagy up
Figure 4. The mitochondrion vertical closes the weak arm. (a) A fermentative→respiratory shift (glucose→glycerol) raises the mitochondrial biogenesis score (GSE301747, n = 3/3; Welch p = 0.031, Cohen d = 2.79). (b) Rapamycin-induced mitophagy raises the degradation module and, more sharply, the Atg32/Atg33-selective sub-score (Welch t = 7.38, d = 8.26) — a larger effect than the full module, because the selective arm is not diluted by shared bulk-autophagy machinery. The direction is unambiguous (every control replicate below every mitophagy replicate) but n = 2/3 leaves it formally underpowered (p = 0.058), stated plainly. (These yeast Welch p-values are screening statistics at n = 2–3/group; the reproducible directions, not the exact p, are the claim — Fig. 19.)

What this establishes: the platform is reusable across organelles with no change to the engine, and — critically — the peroxisome clearance weakness in plants is biological (a missing receptor), not a scorer defect. Where a genuine selective receptor exists (Atg32), the selective-degradation arm is populated and moves correctly, in fact more cleanly than the full module. The loader makes the contrast explicit: selective-marker count is 1 (peroxisome/yeast), 0 (peroxisome/Arabidopsis), 2 (mitochondrion/yeast).

Four organelles, one rule — the platform generalizes generalization

The engine was then applied unchanged to two more organelles and a crop. Across all four organelles the same rule holds: the selective-clearance arm works exactly where the biology provides a dedicated receptor, and is empty everywhere else.

Endoplasmic reticulum (yeast) — a second, cleaner closed arm

The ER has genuine selective ER-phagy receptors, Atg39 and Atg40 (Mochida et al. 2015). The scorer separates a biosynthetic ER-expansion program from selective clearance on independent datasets.

ER validation
Figure 5. ER vertical — the selective arm closes again. (a) Tunicamycin (UPR / ER stress) raises the ER biogenesis score (GSE260735, n = 2/2; Welch p = 0.026, driven by INO1/KAR2/PDI1). (b) Rapamycin raises the Atg39/Atg40-selective ER-phagy sub-score (GSE149016; Welch p = 8×10−4) while biogenesis stays flat (p = 0.37) — a clean directional separation. Effect sizes are very large but n is small (2–3), so these Welch p-values are screening statistics (Fig. 19); the reproducible directional separation, not the exact p, is the result. ER "abundance" is a continuous membrane network, so the score indexes the biosynthetic/clearance program, not a copy number.

Chloroplast (Arabidopsis) — the plant clearance limit again, and a granularity lesson

Chloroplast validation
Figure 6. Chloroplast vertical. (a) On de-etiolation (dark→light greening), the whole biogenesis module is flat because import and division machinery are light-neutral, but the photosynthesis-proliferation submodule rises sharply (GSE156677, n = 3/3; Welch p = 0.011, d = 5.9; RBCS/LHCB transcripts 2–3× up) — a reminder that the right granularity matters. (b) Dark-induced senescence raises bulk chlorophagy (GSE158053; MWU p = 0.032), but the selective arm stays flat (d = −0.26, n.s.): Arabidopsis has no clean chloroplast-selective receptor, exactly like plant pexophagy.

Crop demonstration — rice chloroplast biogenesis fingerprints stress crop

The conserved chloroplast biogenesis machinery was mapped to rice (Oryza sativa) by orthology and scored across a replicated four-condition stress panel. The biogenesis arm discriminates:

Rice crop transfer
Figure 7. Crop vertical — rice. (a) Chloroplast biogenesis score across control, drought, salt and cold (GSE6901, n = 3/condition). (b) Dehydration stresses suppress it — drought Hedges g = −2.27 (p = 0.028), salt g = −1.51, while cold maintains biogenesis (g = −0.19, n.s.); dehydration-vs-mild Welch p = 0.013. The 4-way omnibus is a trend (Kruskal-Wallis p = 0.14), power-limited at n = 3. (c) Honesty panel: the clearance arm rises under drought/salt but is bulk-shared (no plant chloroplast receptor), so it is not chloroplast-selective. The crop-resilience readout: dehydration shifts rice cells away from building chloroplasts; cold does not.

Completing the grid — universal organelles across kingdoms coverage

To test the platform at scale, the five remaining biologically-valid organelle×species cells were filled by orthology transfer onto expression data already in hand (yeast has no chloroplast, so the plant/yeast grid is 11 cells): peroxisome, mitochondrion and ER in rice; mitochondrion and ER in Arabidopsis. A twelfth cell — the human mitochondrion — is added below. In every plant cell the selective-clearance arm is empty by construction — plants carry no ortholog of the yeast receptors Atg32 (mitophagy) or Atg39/Atg40 (ER-phagy) — consistent with every earlier plant vertical.

Arabidopsis mitochondrion and ER fill
Figure 8. Mitochondrion and ER in Arabidopsis (AtGenExpress, 248 arrays). Both organelles' biogenesis scores discriminate across the nine stresses, tissue-stratified (mitochondrion: roots Kruskal-Wallis p = 1.4×10−8, shoots p = 2.0×10−6; ER: roots p = 1.3×10−2, shoots p = 4.1×10−4). Roots suppress biogenesis under salt/osmotic; shoots raise it under heat/UV-B. Selective-clearance arms empty (annotated).
Rice organelle fill
Figure 9. Peroxisome, mitochondrion and ER in rice (GSE6901, 4 stresses × 3 replicates). A deliberately-reported negative result: none of the three universal organelles' biogenesis scores discriminate the acute 3 h stresses (Kruskal-Wallis p = 0.83 for all three; within-replicate variance dominates the acute response). This contrasts with the rice chloroplast module on the same dataset, which did discriminate (dehydration Hedges g = −1.65, Fig. 7) — evidence that the chloroplast signal is organelle-specific (photosynthesis is acutely dehydration-sensitive), not a property of the dataset. Reported as-is, not hidden.

The human vertical — closing the clearance arm the plants left empty disease-relevant

Every plant cell left the selective-clearance arm empty by construction — plants lack orthologs of the yeast autophagy receptors. Humans are the opposite: a rich, well-characterized set of selective mitophagy receptors (PINK1/Parkin, NIX/BNIP3L, BNIP3, FUNDC1). The human mitochondrial vertical therefore does not just add an organism — it populates and validates the arm that was empty throughout the plant kingdom, in a disease-relevant species, on both directions.

Two design choices, both driven by review, make the human validation honest. First, the primary biogenesis score is machinery only (import/MICOS/OXPHOS-assembly/mtDNA apparatus); the master regulators (PGC1α, NRF1, TFAM) are scored separately, because PGC1α induction is the textbook exercise response and scoring it inside biogenesis would be near-tautological. Second, a general-anabolism control (ribosome biogenesis) tests whether the score is organelle-specific or merely a growth signal.

Human exercise validation
Figure 10. Biogenesis arm — human exercise positive control (GSE302582, vastus lateralis; parallel-group RCT, ~24 participants, ~12 per arm: SIE vs MICE; each is their own pre/post). (a) The machinery-only biogenesis score rises from baseline to +24 h (paired Cohen dz = 0.81, Wilcoxon p = 9×10−4, n = 24 paired pre/post samples), significant in both exercise arms (SIE and MICE). The pre/post pairing is within-participant even though the trial design is between-subjects across arms. Notably it dips at +2.5 h while the regulators peak — so the machinery rise is a distinct, delayed program, not an artifact of the coactivator burst. (b) The master regulators (PGC1α etc.), scored separately, peak at +2.5 h and relax by +24 h — expected coactivator kinetics, not counted as biogenesis proof. (c) Growth-confound control: the machinery score is essentially orthogonal to the ribosome-biogenesis anabolism score (r = 0.06); ribosome biogenesis itself rises strongly at +24 h, but after regressing it out the residual machinery score still rises (dz = 0.66, p = 0.007). The mitochondrial signal is not a generic anabolic signal.
Human mitophagy control
Figure 11. Clearance arm — human mitophagy positive control (GSE319526, Panc1, 1% O2 48 h). (a) The mito-specific selective sub-score (PINK1/PRKN/NIX/BNIP3/FUNDC1/FIS1) rises from normoxia to hypoxia (Welch p = 0.009, Cohen d = 8.2, n = 3 vs 3; the equal-variance Student's t gives 5×10−4 but normoxia has near-zero variance, so the variance-robust Welch value is reported). The induction is blunted in HIF1α-knockout cells — a CRISPR-KO mechanistic control confirming the receptors are HIF1α-dependent, not a nonspecific stress response. (b) The rise is driven by the canonical HIF targets BNIP3 (+2.4) and NIX (+1.7); FUNDC1 mRNA is flat (it is regulated post-translationally, an honest limitation of an mRNA score). Caveat — this arm is mildly circular: hypoxia→HIF transcriptionally induces BNIP3/NIX, which are two of the readout genes, so "does the score detect mitophagy" is here partly "does the score detect HIF targets." The HIF1α-KO control confirms the mechanism is real but does not de-circularize the readout. Both directional arms are now populated in human, but the biogenesis arm (which deliberately separates the stimulus-driven regulators from the machinery) is validated more cleanly than the mitophagy arm — see the tiering below.

Concrete utility — a drug mechanism-of-action screen application

The payoff of the human vertical: given any compound-perturbation transcriptome, does the directional score read whether a drug pushes mitochondria toward build-up or selective clearance? Five public human compound–RNA-seq studies were scored strictly within-dataset (compound vs its own vehicle; ≥3 replicates/arm). This is validation, not discovery: recovering the known mechanism of well-characterized compounds is the evidence the readout works.

Drug mechanism-of-action screen
Figure 12. Compounds land where their mechanism predicts. (a) Build-up (biogenesis, x) vs selective-clearance (mitophagy, y) plane. The clean known-answer anchors are urolithin A and CCCP (below); metformin (3 mM) does sit in the build-up quadrant (biogenesis Δz = +0.36, p = 0.008, dose-dependent — flat at 0.25 mM), but it is a contested anchor — metformin is both an AMPK activator and a complex-I inhibitor, and 3 mM is supraphysiological, so this is a consistent read rather than a clean mechanism recovery. (b) Per-arm within-dataset effects; the ISR-dominated poisons (CCCP, oligomycin, rotenone) are read only on the selective sub-score (they are stress poisons, so biogenesis/full-degradation calls are withheld). CCCP drives selective mitophagy hard (Δz = +0.83, p = 0.003); oligomycin trends up; rotenone does not recover, and AICAR does not recover on the transcriptional biogenesis arm (its AMPK action is largely post-transcriptional) — both reported openly as honest limits of an mRNA proxy.
De-circularization of the mitophagy arm — the key result. The hypoxia control (Fig. 11) was mildly circular: HIF directly induces the readout genes BNIP3/NIX. Urolithin A and CCCP trigger mitophagy post-translationally (PINK1 stabilization, Parkin recruitment), so a transcriptional selective-sub-score rise under them is a genuine downstream readout. Urolithin A is the cleanest instrument: its selective-sub-score rise is carried by PINK1 and FIS1 (Δz +1.5–1.9), not the HIF-target receptors BNIP3/NIX, and urolithin is not a HIF/ISR inducer — so there is no transactivation shortcut. This moves the mitophagy arm toward HIGH confidence in a way hypoxia alone could not. (CCCP corroborates but its rise is partly BNIP3/NIX-driven, carrying an ISR caveat.)

Closing the integration gap — net-direction within one human system integration

Until now the two human arms were validated in different systems — biogenesis in exercising muscle, mitophagy in hypoxic Panc1 — so the platform had never shown its defining quantity, the net directional score, integrating build-up against clearance inside a single system. This closes that gap.

Net-direction flips sign within one human system
Figure 13. The net directional score changes sign within one human system. (a) Across hiPSC→cardiomyocyte maturation (GSE273793, n = 3/state), the net score (biogenesis − selective mitophagy) flips from +0.82 in iPSCs (build-up) to −0.50 in contractile cardiomyocytes (clearance) — a monotone trend (slope −0.66/step, r = −0.78, p = 0.013; iPSC-vs-CM Δ = −1.32, p = 0.020). The flip is arm-decomposable: the selective-mitophagy arm rises faster (Δ = +1.64, p = 0.008) than biogenesis (Δ = +0.32, p = 0.12), recovering the documented perinatal cardiac mitophagy wave. (b) An independent HeLa-TFEB stress panel (GSE273702) reproduces net-direction separation: Torin1 drives the net score negative (Δ = −1.43, p = 0.021), arm-selectively (mitophagy p = 0.007, biogenesis p = 0.31). Both used the locked modules and the unmodified scorer. The flip arises from differential magnitude (clearance overtaking build-up), not a strictly opposite-direction perturbation; n = 3/arm, mRNA proxy — screening-grade, but the integrated directional read the two-system validation lacked.
What makes the two arms subtractable. net_direction = biogenesis − selective-degradation is only meaningful if the two arms share a scale. Under the scorer's method='zscore', each arm is the mean of per-gene z-scores across the study's samples, referenced to an empirical null of random same-size gene sets — so both arms are expressed in the same unit (within-study standard deviations vs chance), which is what licenses the subtraction. One honest consequence: the arms are not guaranteed to have equal dynamic range in a given dataset (module size and inter-gene correlation differ), so when both move the same way the sign of the net score is set by whichever arm has the larger within-study swing. That is exactly the situation here (clearance swings wider than build-up), so the net sign is a statement about relative dynamic range, not an absolute build-vs-clear balance — read it as "which process dominates the transcriptional response," not as an organelle headcount.

Coupled organelles — reading the balance across a QC network coupled organelles

The platform's deeper claim is not one more organelle but reading the directional balance across coupled organelles simultaneously — an organelle-QC fingerprint. Four human organelles were added scout-first (each required a public dataset with a known-direction positive control for the arm being built; no control = no vertical), then read jointly.

Mito–lysosome axis — the QC coupling of neurodegeneration and aging

Mitophagy delivers damaged mitochondria to lysosomes, and lysosomal failure (e.g. GBA in Parkinson's) blocks it — so mitochondria and lysosomes are a coupled quality-control unit. A two-armed lysosome vertical was built (79-gene TFEB/CLEAR biogenesis machinery; 29-gene lysophagy module with 6 lysosome-selective receptors — galectins LGALS3/8/9, TRIM16, UBE2QL1, FBXO27 — flagged against 23 bulk-shared genes).

Joint mito-lysosome directional fingerprint
Figure 14. The joint mito–lysosome fingerprint reveals coupling no single-organelle score sees. All controls sit in one dataset (GSE226743, HeLa, 56 samples): TFEB-overexpression raises lysosome biogenesis (Δ = +1.22, p = 0.005), LLOMe drives lysophagy, CCCP/valinomycin drive mitophagy — each arm recovers its known-direction control. (a) In the joint plane, mito uncouplers move mitochondria only, TFEB-OE moves lysosomes only — so across unrelated stressors the two are uncorrelated (r = −0.10, as expected). (b) Within a shared lysosomal-damage stress (LLOMe washout timecourse, WT n = 16), the strongest evidence of genuine coupling is the TFEB-dependence: both organelles swing to clearance together, then lysosome biogenesis recovers by 12 h while mitochondria stay negative — and that recovery is blunted in TFEB-KO, i.e. removing the shared regulator breaks the coordination. That regulator-dependence, not a raw correlation, is what makes this mechanistic coupling rather than coincidence. The two scores also co-vary across the timecourse (r = +0.88), but that number should be read with care: any two scores that both respond to a single temporal perturbation will correlate along the time axis, so the raw r partly reflects the shared trajectory, not coupling per se — the TFEB-KO contrast is the load-bearing evidence. This is a directional balance-reading no single-organelle analysis captures, and the natural home for the Parkinson's/aging question the single-cell track left inconclusive (Fig. 21).

A metabolic-disease fingerprint — lipid + mito + ER on existing data

NAFLD/NASH is lipid-droplet accumulation + ER stress + mitochondrial dysfunction. A lipid-droplet biogenesis module (31 machinery genes; SEIPIN/DGAT/perilipins) was built — biogenesis-only, since lipophagy is bulk autophagy with no clean LD-selective receptor (no fabricated clearance arm, the same limit the platform records for plants) — and a human ER vertical (below) was reused to score the triad on the metabolic-disease datasets already in the breadth panel.

Metabolic-disease organelle fingerprint
Figure 15. Organelles dissociate by tissue in metabolic disease. (a) Directional fingerprint (Cliff's δ, case vs own control). Liver (NAFLD/NASH) shows coordinated lipid-biogenesis UP + ER-stress UP + selective ER-phagy UP — the lipotoxic ER-stress axis; muscle (T2D) shows the classic mitochondrial-biogenesis deficit, with lipid/ER flat. The organelles are not one global disease axis; they dissociate by tissue. (b) Positive controls recover the mechanistically correct genes (tunicamycin: BiP/HSPA5, HYOU1, PDIA4, plus ER-phagy receptor CCPG1; adipogenesis: PLIN1, CIDEC, DGAT2). Marks (*) denote directional effect size at n = 3–6/arm, not powered significance (screening-grade). Honest method note: the pre-registered mito-down-in-liver is method-dependent — it holds under the rank method but flips under the mandated z-score method (absolute mito mRNA rises with the anabolic surge); reported, not method-switched. Because it is method-dependent it is tiered EXPLORATORY and is not a headline claim of this figure — the load-bearing result here is the tissue dissociation (liver lipid/ER vs muscle mito), which is robust to method. Marks (*) denote directional effect size at n = 3–6/arm, not powered significance (screening-grade); every biogenesis arm carries the anabolism/growth confound.

Human ER — extending the selective arm across species

The ER-phagy arm, validated in yeast (Atg39/40), was extended to human: 73 biogenesis-machinery genes and 7 genuine selective ER-phagy receptors (FAM134B/RETREG1, TEX264, CCPG1, RTN3, ATL3, RETREG2/3) flagged against bulk-autophagy genes.

Human ER two-arm validation
Figure 16. Human ER — the two arms move in opposite directions under the two stimuli. (a,b) Tunicamycin (UPR) drives the ER toward biogenesis (net_direction UP: HepG2 n = 5v5 p = 3×10−4; LN308 n = 3v3 p = 0.017, replication), while starvation drives it toward selective clearance (HAP1 EBSS net DOWN p = 0.007, with an ATG7-KO mechanistic control abolishing the rise; HeLa replicates the direction). (c) The ER-phagy receptor CCPG1 rises under UPR and FAM134B/RETREG1 under starvation. Biogenesis arm HIGH confidence; the selective-clearance arm is a lower bound — direction replicated across two cell lines + a KO control, but mRNA underestimates it because ER-phagy receptors are post-translationally controlled. An honest partial.

Ferritin–mito ferroptosis axis — a validated arm and an honest negative

Ferroptosis is iron-dependent, and NCOA4-ferritinophagy releases the iron that drives it. A ferritin/iron vertical was built (storage: FTH1/FTL/FTMT + PCBP1/2; ferritinophagy: NCOA4 + core machinery) and paired with mitochondria.

Ferritin-mito ferroptosis axis; ferritinophagy arm not mRNA-readable
Figure 17. The storage arm validates; the ferritinophagy arm is not mRNA-readable — reported as a partial. (a) Iron loading (ferric ammonium citrate) raises the ferritin-storage score (FTH1/FTL up), with the ground-truth marker TFRC moving oppositely (down), confirming genuine iron loading (GSE168534, replicated in HepG2). But under iron chelation (which should induce ferritinophagy), NCOA4 mRNA stays flat in every dataset — because NCOA4-ferritinophagy is controlled post-translationally, so the transcript cannot see it. (c) Built from only the two validated arms (iron storage × mito net-direction), the axis separates a-priori ferroptosis-prone (iron-loaded) from -resistant (chelated) states. This is the discipline working as designed: no clean mRNA positive control for the ferritinophagy arm means it is reported, never scored as validated — the same rule that keeps the plant clearance arms empty.

Breaking the mRNA ceiling in human — same-sample multi-omic grounding grounding

Every human arm to this point rested on mRNA alone, and mRNA-proxy is the caveat a reviewer leans on hardest. The wheat papers retired that caveat with a specific design: the same samples measured at several molecular layers, in a contrast. We replicated that design in human data — four datasets where the directional organelle score is computed independently on each layer of the same samples, then checked for agreement. The framing is cross-modal grounding, not disease discovery: the point is that the score reads consistently across transcript, protein, phospho-regulation, and function — agreeing cleanly where the biology is a mass shift, and decoupling in the published, informative way where regulation is the mechanism.

CPTAC ccRCC cross-modal grounding
Figure 18. Cross-modal agreement at scale — CPTAC clear-cell renal carcinoma, 185 same-patient tumor/normal pairs. ccRCC is a known-direction positive control: VHL loss shifts metabolism away from oxidative phosphorylation, so a mitochondrial-biogenesis decrease in tumor is textbook. (a) The machinery score is DOWN in tumor at both layers — RNA Hedges g = −1.5 (p = 4×10−21), protein g = −3.1 (p = 2×10−47). (b) Per-gene tumor effects coordinate across modalities (Spearman ρ = 0.50, p = 1×10−5; 94% of machinery proteins fall). (c) Specificity: the ribosome-anabolism control moves the opposite way (up in tumor, both layers), so the mitochondrial drop is organelle-specific, not a generic growth-state artifact. This is the breadth counterpart to the depth-in-one-disease result below.
Parkinson's paired omics
Figure 19. Informative decoupling in disease tissue — Parkinson's BA9 cortex, 12 vs 12 same-subject RNA + proteomics (GSE68719). Here the two abundance layers disagree in an informative way: mito-biogenesis machinery is significantly DOWN at the protein level (g = −1.0, p = 0.017; 93% of detected proteins fall) while mRNA is flat (g = −0.08, n.s.). The deficit is protein-specific — the published PD proteostasis-decoupling phenomenon. Reported as cross-modal consistency in direction at the group level (both ≤ 0), not a per-subject correlation, which sidesteps the neuron-loss composition confound. The selective-receptor arm was below the shallow-brain TMT detection floor here (0/5), motivating the deeper proteome and the phospho layer next.
CPTAC phospho selective arm
Figure 20. A third orthogonal layer resolves the selective-mitophagy arm — CPTAC phosphoproteomics. The selective-mitophagy arm had decoupled at abundance (mRNA up +0.85, protein down −0.86). Selective mitophagy is regulated post-translationally, so the signal should live in the phospho layer. (a) It does: the arm reads strongly DOWN in tumor at phospho (g = −1.5, p = 1×10−19) — two protein-level readouts now agree, and the mRNA-up was the outlier. (b,c) Phospho-occupancy (phospho normalized to protein) exposes regulation the abundance layers cannot see: PRKN loses occupancy beyond its protein loss (g = −1.0), and BNIP3 is the headline — its protein rises while its phosphorylation collapses (occupancy g = −1.9, p = 1×10−29), a pure regulatory signal invisible to both mRNA and protein. This converts the arm's weakness into a capability: it reads coherently once scored at the layer where its biology is regulated. (Honest limits: PINK1 undetected in the phospho matrix; FUNDC1's phospho drop merely tracks its protein loss.)
MoTrPAC multi-omic exercise
Figure 21. The flagship exercise control at full strength — MoTrPAC endurance training, rat skeletal muscle, four layers on the same animals. The human exercise control was previously the weak form (transcript and protein from different cohorts). MoTrPAC is the strong form: (a,b) 8-week training raises the mito-biogenesis machinery score at all three molecular layers of the same muscles — RNA g = +1.0, protein g = +1.7, phospho g = +1.6 (all significant). (c) The metabolome adds a functional readout, the closest analog to the wheat N-BODIPY organelle-abundance grounding: 9 of 10 TCA / oxidative-metabolism intermediates rise (Acetyl-CoA, isocitrate, aconitate, NAD+, carnitine, succinate, fumarate…; sign-test p = 0.021). Transcript → protein → phospho → function all read the same direction, in a canonical known-direction positive control. (Rat; the module is human-ortholog-mapped by symbol, and the supported metabolite claim is set-level direction, not any single metabolite.)

Together these four retire the mRNA-proxy caveat where it mattered most. Two same-sample human datasets show the score grounds at the protein level — agreeing at scale where the biology is a mass shift (ccRCC), decoupling in the published proteostasis way where that is the mechanism (PD). A third layer (phospho-occupancy) rescues the one arm that abundance could not read. And a four-layer atlas turns the flagship exercise control into a same-sample result with a functional metabolite anchor. The engine is unchanged throughout; only the molecular layer it reads differs.

Does it hold up? Benchmark, robustness, scale, and an honest null stress-tests

Four independent stress-tests probe whether the platform's claims survive scrutiny — the first is the one that matters most, because it tests the falsifiable premise of the whole project.

Naive baseline benchmark
Figure 22. The directional score beats single-axis enrichment — the core-thesis test, fought fair. The directional score is pitted against both a naive union-marker "activity" score and the competitive-standard method (ssGSEA, Barbie et al. 2009), so the win cannot be dismissed as beating a strawman. (a,b) In yeast both the naive flat score and ssGSEA rise under oleate-proliferation and rapamycin-clearance — neither single-axis method can tell build-up from tear-down. (c) The directional score gives them opposite signs (build-up vs tear-down). (d) In Arabidopsis, ssGSEA discriminates stresses (η2 = 0.36) but the directional axis carries η2 = 0.28 of discrimination it cannot recover (incremental F = 6.8, p = 6×10−8). (e) ssGSEA is orthogonal to direction (r = −0.17); osmotic stress ranks #1 of 9 on enrichment but last on net-direction. (f) The tell: ssGSEA and the naive flat score are 99% correlated (r = +0.99) — the field-standard method is just tracking the same single abundance axis. The failure to read direction is intrinsic to any single-axis enrichment, not an artifact of a weak baseline. This is the definitive answer to "isn't this just gene-set enrichment?"
Statistical robustness
Figure 23. Directions are robust; the competitive p-values are anti-conservative — stated openly. Every headline result re-tested under correlation-preserving nulls. (a) Because co-regulated module genes are not independent, a CAMERA-style variance-inflation correction (VIF up to 18×) inflates the naive competitive p-values by orders of magnitude — so the reported 10−6–10−9 values are screening statistics, not defensible significance. (b) Under a sample-label permutation null the directions survive; well-powered results (Arabidopsis 9-stress, rice dehydration) stay significant, while small-n yeast contrasts (n = 2–3/group) sit at their permutation floor — reproducible directions awaiting higher-n replication, and every result is single-driver-independent (leave-one-driver-out, right columns). (c) Growth-confound control for the most vulnerable plant claim: cold suppresses Arabidopsis peroxisome biogenesis raw (roots Hedges g = −1.76), and the suppression survives regressing out a ribosome-biogenesis anabolism control (residual g = −1.99 roots, −1.23 shoots; the anabolism control itself moves oppositely in shoots) — so cold-suppression of the organelle score is not a generic growth artifact. The platform's directional claims are the defensible product.
Human breadth map
Figure 24. The human score separates states at scale — platform demonstration (10 datasets). Scored strictly within-dataset across 10 public human RNA-seq studies (exercise, fasting, hypoxia, two senescence models, adipogenesis, cardiomyocyte differentiation, NASH, NAFLD, T2D). (a) Machinery-biogenesis rises in anabolic/adaptive states (adipogenesis, exercise, fasting) and falls in disease/senescence (NAFLD d = −0.47, p = 6×10−5); the selective-mitophagy arm rises where receptor induction is canonical (hypoxia, irradiation-senescence). (b) The two arms dissociate (opposite sign) in 7/10 datasets — concrete proof they are not two ends of one axis. Magnitudes are within-dataset only, never compared across studies. Scored against an a-priori directional expectation (anabolic/adaptive states → biogenesis up: fasting, exercise, adipogenesis; disease/senescence → biogenesis down / impaired: NAFLD, NASH, T2D, senescence): the machinery arm matches the expected sign in the large majority of datasets, with the disease-down cases carrying the strongest, best-powered effects. The growth-confound caveat travels with the score — senescence and differentiation move general anabolic state too, so these are read as directional consistency, not deconfounded per-organelle claims.
Parkinson's single-cell
Figure 25. Parkinson's surviving neurons — exploratory, hypothesis-generating only. Bulk substantia-nigra "mito biogenesis down" in PD mostly reflects dopaminergic-neuron loss, not a per-cell change. Using single-nucleus midbrain data (GSE157783, 5 PD + 6 controls) and pseudobulking per donor×cell-type, the directional question asked in surviving neurons returns neither a biogenesis collapse (d = −0.05, p = 0.89) nor a mitophagy rise (d = −0.16, p = 0.67). This is graded EXPLORATORY, not a result, and three confounds a batch analysis cannot self-catch all push toward the "biogenesis-down" answer we would otherwise want: (i) residual subtype composition — the comparison pools "neurons" broadly, but the vulnerable dopaminergic subtypes are precisely the ones lost, so a subtype-matched comparison is needed and was not done here; (ii) PD-vs-aging is not age-matched in this single cohort; (iii) ambient-RNA / low-count pseudobulk in degenerating tissue inflates variance. Dopaminergic neurons themselves are too sparse to test after loss (that sparsity is the composition confound), and receptor mRNA is a weak proxy for post-translational PINK1/Parkin flux. The honest verdict is inconclusive at this n and granularity — the clean test needs a larger, age-matched, subtype-resolved SN atlas.

The single-cell null asked for two things it could not itself provide: a confound-clean disease design, and a well-powered, population-scale test. We ran both.

The confound-clean disease test — isogenic Parkinson's iPSC models

The cleanest way to ask a disease question without the genetic-background, cell-composition, age and medication confounds that sank the single-cell analysis is an isogenic comparison: a mutant iPSC line versus its own gene-corrected control — same cells, one edited gene. We scouted for isogenic Parkinson's RNA-seq (GBA-PD is the natural fit — it is the mitochondria–lysosome-coupling disease, tying straight to Fig. 14) and pre-registered five hypotheses before scoring, with the hard rule that no clean isogenic dataset would mean reporting "no clean test," never a confounded cohort.

Mito-lysosome dynamics in isogenic Parkinson's iPSC models
Figure 26. In confound-clean isogenic PD models, the GBA lysosome deficit is confirmed — but the mito–lysosome coupling is a clean, pre-registered null. Primary dataset GSE315738 (a GBA1 allelic series in one isogenic background: gene-corrected +/+ vs IVS/+ vs IVS/IVS, iPSC-dopaminergic neurons), with an isogenic-KO panel (PRKN, ATP13A2, DJ-1) and LRRK2-G2019S-vs-corrected as support. (b) Lysosome biogenesis falls monotonically with GBA1 mutant-allele dose (corrected +1.33 → IVS/+ +0.21 → IVS/IVS −0.60; Spearman ρ = −0.74, p = 4×10−4), and the anabolism/growth control stays flat (ρ = −0.07, p = 0.77) — so it is a specific lysosomal deficit, exactly where glucocerebrosidase acts, not a general biosynthesis slowdown. (c) But the pre-registered prediction — that mitochondrial and lysosomal directions move together — fails: across the eight isogenic contrasts the two net-directions are uncorrelated (Pearson r = 0.05, p = 0.91), and mitochondrial biogenesis does not track the lysosomal deficit. This is an honest, well-powered refutation of the directional-coupling hypothesis in this system: the disease perturbs one arm (lysosome) cleanly without dragging the other. Robustness caveat stated plainly — the GBA effect is a genuine within-background allele dose-response but does not survive pooling against an unrelated-donor control (clone/background variation), so it reads as dose-dependent, not a background-independent classifier. iPSC-derived neurons are developmental models; mRNA proxy.

The population-scale test — organelle-QC decline with age across GTEx

Does the directional score index a real physiological axis at population scale? We scored 5,005 GTEx samples (recount3 GTEx v8, ~980 donors, ages 20–79) across 17 subtissues on every organelle arm, then regressed each directional score on donor age — raw, and after adjusting for the ribosome-biogenesis anabolism axis, RIN, and sex, so a generic age-related transcriptional decline cannot masquerade as organelle-specific.

Organelle-QC decline with age across GTEx tissues
Figure 27. Organelle quality-control declines with age — brain-specifically, and it survives the anabolism control. Age slope (Δz/yr) per subtissue × arm, raw (a) and anabolism+RIN+sex-adjusted (b). The selective-autophagy (QC) decline is brain-specific: 26/117 brain arm×region regressions show a significant negative adjusted age slope; 0/36 peripheral (heart/muscle/liver) tests do. Mitophagy is the earliest and steepest-failing QC arm (steepest single: amygdala, adjusted −0.0073 Δz/yr, q = 0.010), ahead of ER-phagy then lysophagy; cortical/limbic regions fail all three QC arms at once, while cerebellum and spinal cord are spared. Critically, the mito- and ER-QC declines survive the anabolism adjustment, whereas the lysosome- and ER-biogenesis raw declines vanish on adjustment — i.e. those were the generic anabolism axis, and the specificity control worked exactly as designed (the raw dark-red spinal-cord artifact in panel (a) also washes out). Cross-sectional, correlational, mRNA proxy — a machinery-capacity read, not a measured turnover rate; PMI not directly modeled.

Breaking the mRNA ceiling — protein-level grounding in human

The caveat stamped on every human result is that it reads mRNA, not the organelle. In yeast we answered that once, by corroborating the oleate direction against an independent proteome (Fig. 2). Here we do it in the species that matters for health: we scouted public quantitative proteomes (PRIDE/ProteomeXchange) for perturbations already scored transcriptomically, and scored the same locked modules on protein abundance.

Protein-level grounding of the mito-biogenesis score
Figure 28. The machinery-only mito-biogenesis score reproduces at the protein level in exercised human muscle. Flagship PXD026219 (HIIT training, 10 men, label-free): the biogenesis-machinery protein score rises through training to a peak post-HVT and tapers post-recovery — the same rise-then-taper the source paper reports for mitochondrial content — with post-HVT vs baseline paired dz = +2.4, p < 10−4, and critically the ribosome-biogenesis/anabolism control does not separate (p = 0.39): a mitochondria-specific protein signal, not generic growth. A second independent proteome (PXD023084, DIA) corroborates the direction (g = +1.15, p = 0.03), though there its anabolism control moves comparably — so specificity is established for the flagship, and the direction for both. This materially de-risks the mRNA-proxy caveat for the mitochondrial-biogenesis vertical — the health-relevant, transcriptomically-validated arm — mirroring the yeast cross-modality grounding in human. The caveat remains open for the mitophagy, lysosome and ER verticals: their receptors are undetectable in bulk proteomes, or no suitable public steady-state proteome exists (a documented data gap, not a null). Protein is still a proxy for turnover, not a rate.

The mRNA-proxy caveat as an auditable ledger — grounding status for all 16 cells

Rather than leave protein-grounding as one flagship, we swept the entire 16-cell grid: for every cell, is there a public quantitative proteome for a matching perturbation, and if so does the protein-level direction agree with the transcriptomic one? This converts the blanket "mRNA is a proxy" disclaimer into a per-cell ledger.

Protein-grounding status across the 16-cell grid
Figure 29. Protein-grounding status across the whole grid — and a rule for where it holds. (a) Each cell colored by status: corroborated at the protein level (3 cells), proteome found but direction fails (5 cells), or no matched/loadable public proteome (8 cells — all Arabidopsis, where PRIDE holds only PTM/interactome data, plus human cells whose abundance matrices are locked in 20 GB Spectronaut / multi-GB bundles). (b) The three corroborated cells — yeast peroxisome (oleate, d = 22), yeast mitochondrion (respiratory shift, d = 24), rice peroxisome (drought, direction DOWN matching mRNA) — show the protein effect meeting or exceeding the transcriptomic one. The pattern is interpretable, not random: grounding holds in the cells where the perturbation drives to a new steady-state organelle abundance (chronic metabolic shifts, d > 20), and fails in the acute-stress cells — both ER tunicamycin proteomes (yeast 4 h and human 8 h) show the protein biogenesis machinery still lagging the mRNA UPR at those early timepoints. So the score reads a genuine organelle-abundance program where abundance has had time to move, and reads transcriptional intent (not yet realized protein) in acute stress — a temporal caveat, stated per cell, not a blanket one.

Scaling to a drug atlas — L1000 compound × organelle-direction

Does the score extend from a handful of drugs to a library? We scored the LINCS L1000 corpus (GSE70138, 12,744 signatures, 1,713 compounds, 7 cell lines) on every directional arm and mined it as a repurposing resource — ranking which compounds push each organelle toward biogenesis vs selective clearance.

L1000 compound x organelle-direction atlas
Figure 30. A drug–organelle-direction atlas, validated by recovering known mechanisms. (a) Coverage gate: on the 978 directly-measured landmark genes every module fails (6–25% coverage); on the inferred BING space all arms pass (72–94%), so scoring is on BING space — an explicit imputation trade-off, reported not hidden. (d) Known-mechanism recovery validates the atlas: mTOR inhibitors are significantly biased toward selective clearance in mitochondria (Mann-Whitney p = 5×10−3) and ER (p = 3×10−4), while the lysosome arm is a pre-registered null (p = 0.40) because mTOR inhibition activates TFEB, which drives lysosome biogenesis and clearance at once; torin-2 is recovered blind as the strongest pan-organelle autophagy inducer. (b,c) Top ranked inducers per arm are the repurposing leads. Double proxy — transcript abundance (directional, not a rate) AND L1000 model-inferred genes; top leads are broad stress inducers (HDAC/proteasome), pathway-level hits not organelle-selective drugs. A validated resource with its limitations on the label.

Tiered confidence — the honest arm

High confidence — biogenesis arm. Conserved machinery, clean signal, validated in yeast and transferring to plants. The Arabidopsis discrimination finding (Fig. 3) rests on this arm, and it is corroborated at the protein level (Fig. 2).
Lower confidence — clearance arm. Pexophagy machinery overlaps bulk autophagy. In yeast there is a genuine selective marker (ATG36) that behaved correctly. In Arabidopsis there is no clean ATG36 ortholog, so the selective-degradation score is empty by construction (0/0 coverage, NaN every sample) and the full degradation score is bulk-autophagy-confounded — a rising clearance score under osmotic/salt is most parsimoniously bulk autophagy, not peroxisome-selective pexophagy. Reported, tiered down, not overclaimed. The mitochondrion (Fig. 4) and ER (Fig. 5) verticals show this weakness is biological, not a scorer defect: where a genuine selective receptor exists (Atg32 mitophagy, Atg39/Atg40 ER-phagy), the selective arm is populated and moves correctly. The rule is consistent across the whole grid — the selective arm works in every yeast case with a known receptor and in human (mitophagy PINK1/Parkin/NIX/BNIP3/FUNDC1, Fig. 11), and is empty in every plant case (peroxisome, chloroplast, rice), where none exists. The human vertical closes the arm the plants left empty.
High confidence — human biogenesis arm. The exercise validation (Fig. 10) is non-circular by design: the machinery-only score deliberately excludes the master regulators (PGC1α etc.), and it rises on a delayed program (up at +24 h, actually down at +2.5 h when the regulators peak) — so it is not riding the coactivator. The growth-confound defense holds: near-zero machinery↔anabolism correlation (r = 0.06) and a significant residual after regressing out ribosome biogenesis.
Moderate confidence — human mitophagy arm (wears its asterisk). The hypoxia validation (Fig. 11) is mildly circular: the stimulus (hypoxia→HIF) transcriptionally induces BNIP3/NIX, which are themselves readout genes — the same circularity the biogenesis arm avoids by holding regulators out. The HIF1α-KO control shows the mechanism is real but does not de-circularize the readout. The clean path to HIGH is a post-translational mitophagy trigger (urolithin A, CCCP: PINK1 stabilization / Parkin recruitment), where a transcriptional selective-sub-score rise is a genuine downstream readout rather than the stimulus acting directly on the module genes. The drug screen (Fig. 12) delivered exactly this: under urolithin A the selective sub-score rises through PINK1/FIS1, not the HIF-target genes BNIP3/NIX — a de-circularized positive that upgrades this arm from moderate toward high confidence.
Cross-system gap — now closed (screening-grade). The two human arms were first validated in different systems (biogenesis in muscle/exercise, mitophagy in Panc1/hypoxia). The net directional score — biogenesis minus degradation, the method's whole premise — is now shown flipping within one human system (Fig. 13): iPSC→cardiomyocyte maturation flips net-positive to net-negative (trend p = 0.013), reproduced in an independent HeLa-TFEB panel (Torin1 p = 0.021). Screening-grade (n = 3/arm, mRNA proxy) and driven by differential arm magnitude rather than a strictly opposite-direction perturbation, but the integrated directional read is demonstrated, not just asserted.
Coupled organelles & the human expansion. Four human organelles were added scout-first, each gated on a known-direction positive control. Lysosome (Fig. 14) validates both arms and the mito–lysosome coupling — evidenced by its TFEB-dependence (coordination breaks in TFEB-KO), not the raw timecourse correlation — is the coupled-organelle novelty. Human ER (Fig. 16): biogenesis HIGH (two cell lines + a KO control), clearance a partial (mRNA lower-bound). The metabolic triad (Fig. 15) reads a tissue-dissociating liver-vs-muscle fingerprint. Ferritin (Fig. 17): storage arm validated, but the ferritinophagy arm is not mRNA-readable (NCOA4 post-translational) — reported, never scored, exactly as the discipline requires. Lipid droplet is biogenesis-only (no LD-selective receptor), the same empty-arm rule as plants. The pattern holds across all seven organelles: the selective arm is populated where, and only where, a genuine receptor exists.
Disease test — honest split (Fig. 22). In confound-clean isogenic PD models, the mechanistically predicted GBA lysosome-biogenesis deficit is confirmed as a specific, allele-dose-dependent effect (ρ = −0.74, anabolism control flat) — but the platform's own mito–lysosome coupling prediction is a clean, well-powered null (r = 0.05, p = 0.91). We report the null as a null: the disease perturbs one arm without dragging the other, and the coupled-imbalance headline does not generalize to this system. The GBA effect itself is a within-background dose-response, not a background-independent classifier — stated plainly.
Population-scale physiological axis (Fig. 23). Across 5,005 GTEx samples the organelle-QC score declines with age brain-specifically (26/117 brain tests significant; 0/36 peripheral), mitophagy the earliest and steepest arm, and the mito/ER-QC declines survive the anabolism+RIN+sex adjustment while the biogenesis raw declines wash out — the specificity control confirming the signal is quality-control, not generic decline. Correlational and mRNA-proxy, but the score indexes a real physiological axis at population scale.
mRNA is a proxy, not an organelle count. All statements read as "consistent with a shift toward biogenesis / clearance," never as turnover rates. The oleate direction is corroborated at the protein level (Fig. 2); tying the score to actual organelle counts via imaging (e.g. N-BODIPY peroxisome abundance) is the natural next step.

The practical payoff — direction predicts cancer vulnerability and drug response disease application

Everything to this point establishes that the directional score reads real organelle biology across species and molecular layers. The decisive question for human impact is different: does the direction a cell is driving its mitochondria predict something a clinician or drug-hunter would act on? Using three large public human/cancer resources — DepMap CRISPR (1,066 cancer cell lines), the PRISM 1,514-drug viability screen, and TCGA-KIRC survival — the answer is yes for vulnerability and drug response, and an honest no for survival. Every claim below is reported with the confound control that tests it.

The anchor result: across 1,066 cancer cell lines, the higher a line's mitochondrial direction score, the more essential its mitochondrial machinery genes are in a genome-wide CRISPR knockout screen (Spearman ρ = −0.35, p = 1.7×10−³²). The coupling survives a global-dependency-burden control (partial r = −0.32), is specific to mitochondria (the ribosome arm shows no self-coupling), and the top driver genes are the correct biology — SDH complex, mtDNA replication (TWNK/POLG), OXPHOS assembly.

Figure 31
Figure 31. Direction predicts mitochondrial genetic vulnerability across 1,066 cell lines. Per-cell-line direction vs. mean CRISPR gene-effect on machinery (a); specificity matrix showing mito-for-mito coupling with a null ribosome self-arm (b); top vulnerability genes (c). Higher direction → more essential mitochondrial machinery.

The result is robust to the sharpest confound a reviewer raises — that it merely reflects tissue lineage. It does not: the coupling is negative in 17 of 18 lineages tested individually and survives lineage-residualization (ρ = −0.29), and it is strongest in solid tumors (brain −0.57, kidney −0.51), weakest in blood — the opposite of the "blood cancers are OXPHOS-addicted" shortcut. And it outperforms the obvious baseline: reading direction beats reading raw machinery expression ("amount") on the vulnerability endpoint (AUROC 0.67 vs 0.61) — the project's core thesis, demonstrated on a functional readout.

Figure 32
Figure 32. The practical axis, every claim confound-controlled. Lineage-robustness (a); PRISM drug specificity surviving the proliferation control (b); direction beats amount on the vulnerability endpoint (c, AUROC 0.67 vs 0.61); four-test summary (d).

From expression alone, the score also ranks drugs. Against the PRISM viability screen, the clinical Complex I inhibitor IACS-010759 sits at the 3rd percentile of all 1,514 named compounds for direction-predicted sensitivity (ρ = −0.15, p = 8×10−&sup6;, n = 872), and the mitochondria-targeted drug mitoquinone (MitoQ) ranks first in the entire library. Crucially, this survives the proliferation confound: after removing each cell line's general drug-sensitivity, MitoQ is still the most direction-sensitized drug and the mito-drug class stays enriched (Mann-Whitney p = 0.017). The second-platform GDSC check was honestly downgraded — its raw signal proved to be mostly a general-sensitivity axis, and we report it as weak directional consistency, not replication.

Figure 33
Figure 33. Direction ranks mitochondria-targeting drugs to the sensitized tail. IACS-010759 sensitivity vs. direction (a); the full 1,514-drug distribution with mito-drugs at the sensitized extreme (b); top direction-sensitized drugs, MitoQ first (c).

The framework generalizes across organelles at the genetic level, with a graded and biologically-sensible pattern: direction predicts genetic vulnerability strongly for mitochondria, modestly for the ER (surviving both burden and lineage controls), and null for the lysosome — the buffered/degradative organelle correctly shows no coupling. The drug-level cross-organelle test was null and is reported as such, marking the honest boundary of the generalization.

Figure 34
Figure 34. Genetic generalization is graded by organelle; the drug-level test is a reported null. Direction→dependency for mitochondria (strong), ER (modest, survives controls), lysosome (null) (a); the null drug-level cross-organelle matrix (b).

Finally, the honest boundary of the clinical claim. An underpowered survival hint in CPTAC kidney cancer (Cox p=0.04, 21 deaths) suggested direction might be prognostic. Tested at 8× the power in TCGA-KIRC (508 tumors, 168 deaths), it is not — a clean null (Cox p=0.21, log-rank p=0.65), with positive controls confirming the pipeline is sound (stage HR=1.93, p=5.6×10−²¹). The direction score reads mitochondrial biology and predicts genetic and pharmacological vulnerability; it does not read patient survival, and we say so.

Figure 35
Figure 35. A clean survival null at power. Direction does not track stage (a) or stratify overall survival (b) in 508 kidney-cancer patients; superimposed KM curves, log-rank p=0.65. The underpowered CPTAC hint does not replicate at 8× the events.

Reproducibility

All inputs are public (GEO); all gene IDs are resolver-backed with provenance. The packaged organelle_dynamics_kit.zip contains the scoring engine, all sixteen gene-set cells (seven organelles across yeast, Arabidopsis, rice and human — five organelles built out in human), the provenance JSONs, and a multi-organelle loader/demo.

from score_organelle_dynamics import score_organelle_dynamics
from run_dynamics import load_module
bio, deg, deg_sel = load_module("arabidopsis")        # or "yeast"
scores = score_organelle_dynamics(expr, bio, deg,
             degradation_selective_ids=deg_sel, method="rank", n_perm=1000)
scores[["biogenesis_score","degradation_score","net_direction"]]

Engine is organelle- and trait-agnostic: swap the gene-set CSVs (same columns) to score a different organelle, trait, or taxon. The kit now ships 16 gene-set cells across 7 organelles and 4 species (peroxisome, mitochondrion, ER, chloroplast, lysosome, lipid droplet, ferritin; yeast, Arabidopsis, rice, human — five organelles built out in human), each resolver-backed with provenance, plus the scoring engine and a multi-organelle loader/demo. Representative datasets: GSE5862, GSE149016, GSE151898, GSE301747 (yeast expression); PXD069526 (yeast proteome); GSE5620–5628 (AtGenExpress, 248 samples); GSE302582 (human exercise); GSE319526 (human hypoxia mitophagy); GSE226743 (human lysosome/mito coupling); GSE273793 (human net-direction); plus the drug-screen, ER, lipid, and ferritin accessions listed per figure. All inputs public; full per-module provenance in the kit's genesets/*_provenance files.

Built with Claude: Life Sciences · Research track · public / open-source data only. The directional framing, the built-in yeast ground truth, and the cross-taxa transfer are the design's core; the tiered clearance arm is its honest limit.