One lipid clock, two tissues: kinetic pathway enrichment across fruit senescence and immune ferroptosis

When a banana is wounded or a CD8+ T cell enters a tumor, the same chemistry unfolds: polyunsaturated fatty acids (PUFAs) — the fatty acids carrying several carbon-carbon double bonds — lose the weakly bound hydrogens at their bis-allylic positions, and oxygen-driven radical chains convert membrane lipids into hydroperoxides and then into reactive aldehydes like malondialdehyde and hexanal. In fruit, this cascade softens tissue and ruins shelf life; in the tumor microenvironment, it drives ferroptosis — a form of iron-dependent cell death caused by membrane lipid peroxidation — in the very T cells that should be killing the tumor. Two literatures study this cascade in near-total isolation: post-harvest physiologists measure bulk lipid markers in fruit over days, while cancer immunologists profile single cells in patients. Both would benefit from one thing neither has: a way to read pathway activity as a function of where a tissue sits on the lipid-peroxidation timeline, rather than as a static snapshot.

Functional enrichment analysis — the family of methods that turns lists of perturbed genes into pathway-level summaries — is the natural bridge, but every existing tool is static. Overrepresentation tests score a single gene list; LIPEA does the same for lipid lists at coarse class resolution; pmVAE and DeepGSEA learn rich pathway representations from single-cell RNA-seq but treat each sample as an unpositioned point. This proposal builds a kinetic enrichment model in which all pathway modules are indexed by a shared, monotone lipid-oxidation clock estimated from measurable lipid markers, and tests whether that clock — and the map from clock position to pathway activity — transfers between a plant senescence regime and a mammalian immune regime.

Core Claim

The claim is that a one-dimensional latent clock, fit jointly to lipid degradation markers (peroxide value, malondialdehyde, phosphatidylcholine-to-phosphatidic-acid ratio) and to transcriptomes, is a sufficient shared axis for pathway activity in both post-harvest senescence and T-cell ferroptosis. Concretely, we extend the pathway-module architecture of pmVAE — a variational autoencoder whose subnetworks are restricted to annotated pathway gene sets — so that each pathway module produces a latent state conditioned on the clock:

$$z_{p,i} = f_p(x_i, c_i), \qquad c_i = g\big(\text{lipid markers of sample } i\big),$$

where $x_i$ is the expression profile, $z_{p,i}$ is a multi-dimensional state for pathway $p$, and $c_i$ is the clock value, constrained to be monotone in time within each experiment. The scientifically loaded part is not the architecture but the transfer test: whether $f_p$ for the lipid-peroxidation and glutathione modules, learned on a fifteen-day banana wounding time-course, ranks ferroptosis-susceptible CD8+ T cells in a human head-and-neck tumor dataset without retraining. If it does, post-harvest biology becomes a cheap, ethically simple, kinetically clean in vivo model for lipid-peroxidation-driven immune collapse. If it does not, we will have cleanly bounded how far the cross-kingdom analogy carries.

Research Program

Three components get built. First, the clock estimator: a monotone scalar model fit to lipid marker trajectories, calibrated on the purest available chemistry — the Schaal oven benchmark of microalgal DHA-rich oil oxidation, where control oil reaches an oxidation induction period of 9.74 days at 60 °C and a vitamin E / rosemary extract / ascorbyl palmitate mixture extends it to 20.21 days while retaining 42.44% DHA content. These curves, with and without antioxidants, pin down how the clock should respond to both spontaneous and suppressed oxidation. Second, the kinetic pathway modules: pmVAE-style subnetworks over Reactome and KEGG pathway masks, plus a new lipid module whose inputs are acyl-chain-resolved lipid species — the granularity that LIPEA explicitly cannot handle because KEGG annotates lipid classes, not chain length, unsaturation count, or hydroperoxide status. Third, the cross-kingdom transfer harness, which aligns banana and tumor samples by clock position rather than by time, tissue, or species.

The biological motivation is documented on both sides. In wounded banana fruit, phospholipase D inhibitor treatment shows that phospholipase D and lipoxygenase dismantle membrane phospholipids, that 0.1% hexanal — itself an omega-6 PUFA cleavage product — suppresses this cascade and preserves firmness through 15 days, and that the physiological inflection arrives sharply after day 9. On the mammalian side, SLC2A3 in head and neck cancer shows that glucose-transporter overexpression in tumor-infiltrating CD8+ T cells suppresses the GCL/GSS/GPX4 antioxidant axis, precipitating ferroptosis in those T cells, with high SLC2A3 predicting poor overall survival ($p = 0.0075$, TCGA cohort, $n = 504$); and the pan-cancer MBOAT2 analysis shows a membrane-remodeling, ferroptosis-suppressing enzyme correlating with immunosuppressive infiltration. Notably, the banana paper leaves the downstream transcriptional network unmapped and uses no enrichment analysis at all, while both tumor papers rely on static enrichment — each side holds exactly the piece the other lacks.

Proposed Experiments

The banana arm extends the published 0–3–6–9–12–15 day wounding design (25 °C, 90–95% relative humidity, hexanal / n-butanol / 2-butanol / untreated arms) with transcriptome and acyl-chain-resolved lipid sampling at each time point. The headline metric is early warning: using only days 0–6 data, predict the day-9 physiological inflection (firmness loss, respiration rise, malondialdehyde surge) and report how many days in advance the prediction fires, at a fixed false-alarm rate, against the eventual measured outcome. The file reports the banana paper's effects qualitatively; no reported result found on Lacuna for exact numerical firmness or malondialdehyde values, so the day-9 threshold serves as the event label and the model's detection horizon as the primary number.

The tumor arm uses the GSE103322 single-cell dataset of head-and-neck tumors already exploited by the SLC2A3 study. Cells are labeled ferroptosis-susceptible by low GPX4/GCL/GSS and high SLC2A3 expression, and the task is to rank cells by ferroptosis-trajectory membership, scored by AUROC. The transfer test applies the banana-trained lipid-peroxidation and glutathione modules directly. A secondary validation checks whether module activity in bulk TCGA HNSC tumors recovers the published survival separation.

Two benchmark comparisons anchor the method in the enrichment literature. On the Kang et al. PBMC interferon-β stimulation dataset, pmVAE reports 0.97 classification accuracy for identifying stimulated cells, against 0.76 for Interpretable AE and 0.44 for f-scLVM; our model must stay within noise of 0.97 there, demonstrating that adding lipid and clock conditioning costs nothing on a standard non-lipid task. On the LIPEA validation panels — a six-lipid major depressive disorder panel that recovered glycerophospholipid metabolism (KEGG hsa00564, corrected $p < 0.05$) and a rat triacylglycerol panel that recovered insulin resistance (KEGG rno04931) — we run LIPEA and our model on degraded, class-collapsed versions of the same lipids plus synthetic acyl-chain-resolved variants, to quantify how much early signal class-level annotation erases.

Baselines

The naive reference point is bulk biochemical monitoring alone — peroxide value and malondialdehyde assays as used in the DHA oil screening and the banana wounding study. This is the incumbent practice in both fields: it is cheap and directly interpretable, but it detects collapse only as it happens and cannot say which cells are dying versus which are surviving at a peroxidizing interface — the exact blind spot the SLC2A3 paper exposed, where bulk measurements hid T-cell-specific ferroptosis inside a viable tumor. Beating it means issuing a correct inflection warning at least three days before the measured day-9 event at a pre-registered false-alarm rate.

The strongest existing method is pmVAE (pmVAE), which learns multi-dimensional pathway-module representations and currently defines the accuracy ceiling on pathway-level single-cell interpretation; it is the direct architectural ancestor of our model, which makes the comparison both fair and diagnostic. We will run pmVAE with identical pathway masks, identical training splits, and a matched tuning budget (nested cross-validation over the same hyperparameter ranges and the same GPU-hours), feeding it the same expression matrices but without lipid inputs or clock conditioning. DeepGSEA (DeepGSEA), which reported higher sensitivity than ssGSEA, GSVA, and AUCell in Splatter simulations, runs as a second competitor under the same protocol; no reported result found on Lacuna for its absolute Splatter sensitivity values, so we regenerate its numbers on our splits rather than quoting them. We count our method as winning only if it significantly exceeds both on the two kinetic tasks — banana inflection prediction and ferroptosis-cell ranking in GSE103322 — while staying within one standard error of pmVAE's 0.97 on the PBMC benchmark. For lipid-side analysis, LIPEA (LIPEA) is the incumbent; beating it means detecting the phospholipase D / lipoxygenase cascade from lipid data before day 9 and resolving hydroperoxide species its KEGG mapping collapses.

Finally, the key-ingredient ablation is our own model with the clock removed — pathway modules receiving expression but no $c_i$, trained on the same data. This isolates whether kinetic conditioning, rather than merely more inputs or a bigger network, is what buys the early warning; if the static ablation matches the full model, the clock is decoration and the proposal's core claim fails.

Why not just…?

Why not just run standard overrepresentation analysis on the differential genes? Standard Fisher's-exact testing over GO and KEGG — as unified at scale by funkea — is fast and familiar, but it collapses a pathway to one p-value per gene list and needs a hard cutoff on which genes count. The Splatter results in the DeepGSEA paper show that such cutoff-and-scalar methods lose the subtle, coordinated shifts that dominate early senescence and pre-ferroptotic states, and the pmVAE paper shows quantitatively what one-dimensional pathway scores cost: f-scLVM's 0.44 and Interpretable AE's 0.76 versus pmVAE's 0.97 on the same task, precisely because scalar scores conflate cell identity with stimulus response. ORA also cannot, even in principle, output a time or a per-cell risk, which is the entire deliverable here.

Why not just measure malondialdehyde and peroxide value and skip the modeling? These assays are the workhorses of both the banana and oil-oxidation literatures, and they remain part of this proposal as clock inputs — so the objection is partially right, and we adopt it. Where they fall short is prediction and attribution: in the banana system the aldehyde surge is a consequence of the day-9 collapse, not a forecast of it, and in the head-and-neck tumors the SLC2A3 study demonstrated that bulk lipid-peroxidation readouts actively mislead, because the dying cells are a minority inside a viable mass. The decisive experiment is simple: if peroxide value at day 3 predicts the day-9 inflection as well as our model, the modeling layer is unjustified, and we will report that comparison head-to-head.

Why not keep plant post-harvest science and tumor immunology separate? The honest answer is that the transfer is a hypothesis, not a fact, and this proposal is structured so that its failure is informative rather than embarrassing. What justifies testing it is the shared biochemistry: the same phospholipase D and lipoxygenase enzymes, the same hydroperoxide and aldehyde intermediates (hexanal is both a fruit preservative and an endogenous PUFA cleavage product), and the same glutathione-peroxidase defense axis appears in the banana experiments and in the ferroptosis literature. What makes the test worth running is an asymmetry of experimental access: fruit offers a clean, reproducible, fully sampled in vivo time-course from health to collapse — days 0 through 15 with interventions — that no tumor microenvironment study in the file achieves, since even the MBOAT2 pan-cancer work is static bulk correlation without in vivo kinetic resolution. The cross-kingdom AUROC on GSE103322, with a banana-trained module and no retraining, settles the question either way.

Decisive Checks

Three outcomes would decide the proposal. First, the early-warning number on banana: a correct inflection prediction at day 6 or earlier, with the static ablation failing to match it, establishes that kinetic pathway conditioning extracts lead time that neither bulk assays nor static enrichment provides. Second, the transfer AUROC on GSE103322: a banana-trained lipid-peroxidation module ranking GPX4-low CD8+ T cells above chance would open a genuinely new experimental economy — perishable produce as a standing model system for immune lipid collapse — while a null result cleanly bounds the analogy to the chemistry level, still a useful negative finding given that no cited work has tested it. Third, the non-regression check on the PBMC interferon-β benchmark at pmVAE's reported 0.97: confirms the added machinery is not buying kinetic signal by degrading the general interpretation pipeline. All three checks run on datasets and benchmarks already documented above; none requires data that the existing literature has not shown to be collectable.