Localizing Metabolic Pacemakers for Cancer Drug Stratification

Neuroscientists have spent decades solving a problem that sounds strange outside their field: when many brain areas oscillate together in a rhythm, which one is actually causing it? A rhythm can be driven by one autonomous pacemaker — the lateral geniculate nucleus of the thalamus was shown to push a ~60 Hz rhythm into visual cortex, because silencing the thalamus, but not the cortex, abolished it (Subcortical Source and Modulation of the Narrowband Gamma Oscillation in Mouse Visual Cortex) — or it can emerge from a network of mutually coupled cells with no leader (A neural circuit for gamma-band coherence across the retinotopic map in mouse visual cortex). The distinction matters: a rhythm with a focal driver dies when you silence the driver, while an emergent rhythm survives and reroutes. Cancer metabolism has an exactly analogous unsolved problem. Tumor cells exhibit dynamic metabolic behavior — glycolytic flux, proton extrusion, lactate export — and therapy stratification schemes classify tumors by metabolic state, yet every deployed scheme uses static snapshots: bulk gene-expression panels that run tournaments of machine-learning algorithms to pick a handful of metabolic genes (LAC-TME classifier, MLLS breast-cancer signature).

This proposal ports the perturbation-and-phase logic of pacemaker localization from neuroscience to single-cell cancer metabolism, and tests whether where a metabolic rhythm comes from — a focal, autonomous subpopulation versus an emergent network property — is a better stratifier of drug response than any static metabolic gene signature. The premise is biologically motivated: metabolic constraints demonstrably produce self-sustained oscillations in coupled cell networks (Energetic Constraints Produce Self-sustained Oscillatory Dynamics in Neuronal Networks), and metabolic perturbation can flip a network into a pathological oscillatory regime that matches clinical recordings (Network oscillations in a neural mass model induced by metabolic modulation). No one has asked whether a tumor's metabolic rhythm has an identifiable pacemaker, or whether that pacemaker's existence predicts therapy outcome.

Core Claim

Within a heterogeneous tumor or co-culture, individual cells differ not only in mean metabolic rate but in whether their metabolic dynamics oscillate and whether those oscillations are self-generated or entrained by neighbors. We claim that a cell's pacemaker status — autonomous oscillation that persists under decoupling and phase-leads coupled neighbors — can be measured with existing microfluidic time-lapse technology, computed with a localization score adapted from traveling-wave and perturbation analyses in neuroscience, and that the fraction and identity of metabolic pacemakers in a sample improves prediction of drug response beyond static multi-omics signatures. The central mechanism is that metabolic inflexibility, the phenotype stratification actually cares about, is a dynamic failure — a cell's inability to shift flux regime under challenge — and pacemaker dynamics expose that failure in a way a snapshot cannot: a high lactate level in a static assay is ambiguous between fast turnover and stagnation, whereas the oscillation's persistence, frequency, and response to a metabolic insult are not.

Research Program

Three components. First, an instrument extension: the SCO-pH platform (SCO‐ph: Microfluidic Dynamic Phenotyping Platform for High‐Throughput Screening of Single Cell Acidification) already tracks extracellular pH in ~40,000 droplets with a ratiometric dye, but it isolates each cell and treats acidification as a monotonic ramp over 3 hours. We will build coupled-droplet variants — paired chambers and small clusters connected by diffusion channels — and extend recording to 6–12 hours at ~30-second sampling, since intercellular metabolic entrainment requires both coupling and time. Its reported 16.5% single-cell identification error rate and 22.2% post-correction single-cell yield set the throughput we must engineer around, e.g. with optical barcoding of chambers.

Second, a localization algorithm with two prongs, mirroring how neuroscience separates driver from resonator. The observational prong scores each cell by phase relationships, in the spirit of the traveling theta wave analysis in medial entorhinal cortex (Medial entorhinal cortex activates in a traveling wave in the rat): extract each cell's acidification phase, compute pairwise phase differences across coupled pairs, and look for consistent phase leads. The causal prong is the decisive one, modeled directly on the thalamic-silencing experiment: candidate pacemaker cells are selectively removed (photo-ablation of a chamber) or metabolically silenced (oligomycin A, which the SCO-pH work already shows pushes cells into hyperglycolysis), and we test whether neighbors' rhythms collapse — the tumor analog of suppressing cortical gamma by silencing the LGN. We summarize the result as a pacemaker index,

$$\mathrm{PI}_i = A_i^{\text{iso}} \cdot \frac{1}{|N_i|}\sum_{j \in N_i} \Delta\phi_{ij},$$

where $A_i^{\text{iso}}$ is cell $i$'s normalized oscillation amplitude measured in isolation and $\Delta\phi_{ij}$ is its mean phase lead over coupled neighbor $j$; a true pacemaker has both a persistent isolated rhythm and a positive, causal phase lead that disappears when it is silenced.

Third, a stratification model: pacemaker fraction, oscillator frequency distribution, and the change in these quantities under metabolic challenge (oligomycin A, glucose withdrawal) become covariates alongside standard static features, feeding a survival or drug-response predictor. The design choice this addresses is documented in the clinical record: static average dose metrics in osteosarcoma completely missed the biological signal, while only the cycle-by-cycle dynamic trajectory (regulated received dose intensity) separated the >60% from the <50% three-year event-free survival groups (A novel method to address the association between received dose intensity and survival outcome) — evidence that in this domain the dynamic trajectory, not the static summary, carries the prognostic content.

Proposed Experiments

Synthetic ground truth comes first, because we need known pacemakers to validate against. We will generate coupled-oscillator networks using the mean-field architecture of The Generation and Modulation of Distinct Gamma Oscillations with Local, Horizontal, and Feedback Connections in the Primary Visual Cortex — which shows how local recurrent microcircuits versus long-range coupling produce distinct oscillatory regimes — re-parameterized so that oscillating variables are metabolite concentrations and proton flux, with energetic constraint terms following Energetic Constraints Produce Self-sustained Oscillatory Dynamics in Neuronal Networks. Networks with an implanted autonomous driver, with no driver, and with Rocchi-style regimes where silencing a hub unmasks a distant pacemaker (Increased fMRI connectivity upon chemogenetic inhibition of the mouse prefrontal cortex) provide the test cases; the metric is pacemaker-recovery rate and false-positive rate of the index, with the perturbation prong as the arbiter whenever the observational prong is ambiguous.

The biological experiments use EL4 lymphoma cells (the SCO-pH model system) and a panel of cancer cell lines spanning the GDSC2/CCLE resource — 123,760 IC50 pairs across 204 drugs and 684 cell lines as preprocessed by DrugVNN — selecting lines whose static CCLE features suggest distinct metabolic wiring (glycolytic versus oxidative). For each line we measure pacemaker structure at baseline and under oligomycin A challenge, then test whether oscillatory-pacemaker features improve IC50 prediction for held-out drugs, with particular attention to metabolic-acting compounds, including MEK-inhibitor sensitivity previously linked to a metabolic-state signature in lung squamous carcinoma (disulfidptosis-related mRNA signature).

For the clinical endpoint, we reanalyze the TCGA cohorts used by the static signatures — COAD with the 9-gene mitochondrial signature (Comprehensive analysis of mitochondrial-related gene signature for prognosis), which reported time-dependent AUCs of 0.740, 0.723, and 0.763 for 1-, 3-, and 5-year survival in training — asking whether a pacemaker-derived risk score, mapped onto tumors through matched cell-line models, adds prognostic information out-of-cohort. No reported result found on Lacuna for that signature's validation-cohort AUC on GSE39582/GSE29623, so we will establish that number ourselves as part of the comparison. Prospective validation on a trial cohort analogous to the MRC BO06 / EORTC 80931 design is the stated end goal but lies beyond this proposal's scope.

Baselines

The naive reference point is the monotonic-ramp summary: fit each cell's acidification trajectory to a line, as SCO-pH effectively does, and stratify by slope and slope heterogeneity. It uses the same data and instrument, so it isolates exactly what oscillation analysis adds over rate analysis. The strongest static competitor is the tournament-style signature pipeline exemplified by the LAC-TME classifier (LAC-TME classifier), which screened 101 algorithm combinations with the Mime1 workflow and combined a 9-gene lactate signature with immune-cell fractions; it is the right competitor because it is the current best practice in metabolic stratification and it reports a C-index of 0.92 in training falling to 0.73 in external validation — we will retrain its pipeline on our matched cohorts under the same tuning budget and beat it only if our external-validation C-index exceeds 0.73 with non-overlapping confidence intervals, and, more informatively, if adding pacemaker features to its own signature raises its external C-index. The strongest dynamic-data competitor is the deep drug-response model tCNNS (Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network), which reports a mean $R^2 = 0.826$ on known drug–cell-line pairs in GDSC but a severe drop on blind novel drugs; no reported result found on Lacuna for its blind-test $R^2$ value, so we will measure the blind-test number ourselves on a held-out-drug split and compare our model's blind-drug $R^2$ against both tCNNS and its reported in-distribution 0.826 as context. The key ablation runs the full pipeline with the pacemaker-index features deleted — keeping oscillation amplitude and frequency but removing persistence-under-decoupling and causal phase-lead — which isolates whether localization, rather than merely having dynamic data, is the ingredient that carries the gain.

Why not just…?

Why not just use static multi-omics signatures? Because their own numbers show the ceiling: LAC-TME drops from C-index 0.92 to 0.73 out-of-cohort, and the COAD mitochondrial signature plateaus at AUC 0.740–0.763 in training, leaving substantial unexplained variance exactly where dynamic metabolic state should matter. More decisively, the osteosarcoma trial analysis showed that the static summary of the same patients (average dose intensity) carried no signal while the dynamic trajectory did. Static signatures are retained as baselines, not replaced by assertion.

Why not just run flux balance analysis on a genome-scale metabolic network? The canonical reduction machinery of Metabolic network reductions is elegant and fast — shrinking the mouse iMM1415 model from 3,726 to 1,171 reactions — but it is built on the steady-state axiom $S \cdot v = 0$, under which metabolite concentrations do not change over time; by construction it cannot represent a limit cycle, a phase lead, or the bifurcation that occurs when a drug attacks a pathway. It answers a different question (which fluxes are possible) than ours (which cells drive the dynamics).

Why not just do single-cell RNA sequencing and pseudotime? The SCO-pH authors state the failure mode directly: flow cytometry and bulk assays offer snapshots unsuitable for dynamic measurement, and a snapshot cannot distinguish rapid flux turnover from metabolic stagnation — both look like high lactate. Pseudotime infers an assumed ordering; it does not observe an oscillation, and it cannot run the perturbation test that distinguishes a driver from a resonator. scRNA-seq remains useful here as a complementary readout of which genes mark pacemaker-identified cells, and we use it for that.

Why not just train an end-to-end deep model on all available data? This is the most serious threat, and tCNNS is the cautionary evidence: it reached $R^2 = 0.826$ in-distribution, but its authors showed that shuffling SMILES symbols barely hurt it — it had memorized statistical artifacts — and blind performance on novel drugs collapsed. Our held-out-drug experiment is designed to settle this directly: if a purely data-driven model given the same dynamic features matches our mechanistically-constrained model on blind drugs, the localization machinery is unnecessary and we will say so. Our bet, based on tCNNS's artifact-memorization finding, is that the perturbation-derived pacemaker label — a causal quantity no correlational model can extract from unperturbed data — is what transfers.

Decisive Checks

Three findings would make or break the claim. The synthetic benchmark must show the pacemaker index recovers implanted drivers at rates well above chance, with the perturbation prong resolving cases the phase analysis mislabels. The co-culture experiments must demonstrate at least one biological instance where silencing a small identified subpopulation abolishes oscillatory proton dynamics in coupled neighbors — the tumor-scale replication of the thalamic-silencing result. And the stratification test must show that pacemaker features improve external-cohort C-index beyond the 0.73 static ceiling and blind-drug $R^2$ beyond what tCNNS achieves on the same split, with the ablation confirming that localization rather than mere dynamics is responsible. Failure of the third check with success of the first two would still yield a usable result: a validated method for finding metabolic pacemakers in cell populations, and an honest negative on their clinical value.

References

  1. Subcortical Source and Modulation of the Narrowband Gamma Oscillation in Mouse Visual Cortex
  2. A neural circuit for gamma-band coherence across the retinotopic map in mouse visual cortex
  3. LAC-TME classifier: machine learning-driven model predicts survival and prioritizes targeted therapy in clear cell renal cell carcinoma
  4. Personalized treatment decision-making using a machine learning-derived lactylation signature for breast cancer prognosis
  5. Energetic Constraints Produce Self-sustained Oscillatory Dynamics in Neuronal Networks
  6. Network oscillations in a neural mass model induced by metabolic modulation are consistent with EEG data of neocortical epileptic seizure onset
  7. SCO‐ph: Microfluidic Dynamic Phenotyping Platform for High‐Throughput Screening of Single Cell Acidification
  8. Medial entorhinal cortex activates in a traveling wave in the rat
  9. A novel method to address the association between received dose intensity and survival outcome: benefits of approaching treatment intensification at a more individualised level in a trial of the European Osteosarcoma Intergroup
  10. The Generation and Modulation of Distinct Gamma Oscillations with Local, Horizontal, and Feedback Connections in the Primary Visual Cortex: A Model Study on Large-Scale Networks
  11. Increased fMRI connectivity upon chemogenetic inhibition of the mouse prefrontal cortex
  12. Interpretable Drug Response Prediction through Molecule Structure-aware and Knowledge-Guided Visible Neural Network
  13. A novel disulfidptosis-related mRNA signature predicts prognosis and therapeutic response in lung squamous cell carcinoma
  14. Comprehensive analysis of mitochondrial-related gene signature for prognosis, tumor immune microenvironment evaluation, and candidate drug development in colon cancer
  15. Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network
  16. Metabolic network reductions