Metabolic flux analysis tells us how fast chemical reactions actually run inside a cell: feed a tumor nutrient labeled with heavy carbon and watch where the label travels, and you can read off pathway traffic rather than just pool sizes. In cancer, this technique has produced striking mechanistic findings — for example, that glucose routed through the pentose phosphate pathway makes pancreatic tumors resist chemotherapy Metabolic classification suggests the GLUT1/ALDOB/G6PD axis as a therapeutic target in chemotherapy-resistant pancreatic cancer. But these measurements can only be taken on tumor material that surgeons removed, from patients selected for resection, and the claimed link to patient survival is then validated in retrospective registries whose follow-up is short and whose patients were filtered by the same surgical gate. Every published statement of the form "this metabolic flux phenotype predicts worse survival" therefore carries two unexamined distortions: the sample only contains patients healthy enough to be operated on, and the survival data is cut off by censoring whose mechanisms nobody has checked.
This proposal builds and tests a pipeline that takes those distortions seriously. The core move is to learn, on a small set of patient-derived organoids that have both gene expression and isotope-tracing measurements, a calibration map from static expression to posterior flux distributions; apply that map to large retrospective cohorts where only expression exists; feed the calibrated flux posteriors into a survival model that handles right-censoring explicitly; and then quantify — and bound the damage from — the resection selection filter using population registry data. The deliverable is twofold: an honest, uncertainty-aware estimate of whether flux phenotypes genuinely predict long-term survival, and a reusable method for connecting short-term ex vivo kinetics to censored clinical trajectories.
Core Claim
I claim two things. First, methodological: the uncertainty in a flux estimate — which Bayesian Metabolic Flux Analysis reveals intracellular flux couplings shows how to compute as a posterior distribution over reaction rates — can and should be propagated through to survival inference, because the quantity of interest (a hazard ratio attached to a pathway) inherits that uncertainty, and collapsing it to a point estimate silently overstates confidence. Second, empirical: in pancreatic ductal adenocarcinoma (PDAC), the widely cited association between pentose-phosphate flux and overall survival has never been evaluated with censoring-aware metrics or with any correction for resection selection, and its true magnitude in the general (not surgically selected) patient population is unknown. Testing this is feasible with existing public data: the paired organoid biobank, the TCGA and Bailey PDAC cohorts, and the SEER registry, which together span both the selected and broader populations.
PDAC is the right testbed for a specific reason: conventional clinical staging is weak there. Predictive value of current nodal staging systems and development of machine learning nomogram for resectable pancreatic head cancer: a population-based study and multicenter validation reports that AJCC 8th edition staging achieves a concordance index of only 0.518 to 0.581, while its best machine-learning nomogram reaches 0.675 in training and 0.664 on external validation. This is unlike lung or kidney cancer, where simple clinical anchors beat advanced models — Robustness of Multimodal Foundation-Model Forecasting for Postoperative Cancer Outcomes found pathologic stage alone at C-index 0.837 outperforming the best multimodal AI at 0.799. In a disease where the clinical anchor is weak, molecular features have room to add value, and event scarcity (the same paper shows learned models collapse with only 56 recurrence events) argues for features with strong biophysical priors rather than high-capacity learned representations.
Research Program
The pipeline has four components, each reusing an existing technique in a new position rather than inventing one from scratch.
The first component is the calibration map. The 28-organoid PDAC biobank in Metabolic classification suggests the GLUT1/ALDOB/G6PD axis as a therapeutic target in chemotherapy-resistant pancreatic cancer is unusual: each organoid has RNA-seq, metabolomics, and $[\text{U-}^{13}\text{C}_6]$ glucose tracing on the same sample. I will run the BAMFA Gibbs sampler (released at github.com/markusheinonen/bamfa) on the tracer data within a mammalian genome-scale model to obtain, per organoid, a posterior distribution over pathway fluxes — not a single optimum. On top of those posteriors I will train a map from expression features to flux posteriors, validated by holding out organoids. The precedent for this step is direct: BMFA achieved an average F1 of 41% in predicting unobserved intracellular fluxes from exchange data in C. acetobutylicum, versus 9% for flux variability analysis, and handled models with over 4,000 reactions Bayesian Metabolic Flux Analysis reveals intracellular flux couplings. The map is deliberately low-dimensional — a handful of pathway-level flux summaries — because the downstream cohorts are small (TCGA PDAC has 156 patients) and overfitting a flexible map on 28 organoids would be self-defeating.
The second component is cohort transport. The calibrated map is applied to the TCGA PDAC cohort ($n=156$), the Bailey cohort ($n=55$), and the two published validation cohorts ($n=283$ and $n=285$), assigning each patient not a flux value but a flux posterior. This is where the proposal differs from the existing PDAC work, which used static GLUT1/ALDOB/G6PD expression as the survival marker, and from MIRTH: Metabolite Imputation via Rank-Transformation and Harmonization, which imputes static metabolite abundances across cohorts but explicitly loses kinetic information — rank-transformed pool sizes cannot distinguish active flux from a downstream enzymatic blockage.
The third component is the survival model. Each patient contributes a marginal likelihood that integrates over their flux posterior:
$$\ell = \sum_i \log \int p(\mathbf{v} \mid \mathbf{x}_i, \mathcal{D}_{\text{tracer}})\; p(T_i, \delta_i \mid \mathbf{v})\; d\mathbf{v}$$
where $\mathbf{v}$ is the pathway-flux vector, $\mathbf{x}_i$ is patient $i$'s expression profile, $\mathcal{D}_{\text{tracer}}$ is the organoid tracing data, $T_i$ is the observed time, and $\delta_i$ is the censoring indicator. The survival term uses cluster-specific Weibull likelihoods in the style of VaDeSC A Deep Variational Approach to Clustering Survival Data, whose code is public VaDeSC codebase, but with two departures: the latent structure is informed by the flux posterior rather than a generic encoder, and the likelihood is reweighted by an inverse probability of resection, estimated from registry data, so that the fitted flux–hazard association targets the broader patient population rather than the surgically selected one. Discrete-time extensions can draw on PyDTS.
The fourth component is the bias audit, run in both directions. Censoring is stress-tested by administratively truncating follow-up at 12, 24, and 36 months and measuring how the apparent flux–survival association and metric scores distort, following the synthetic-censoring methodology of Position: Stop Chasing the C-index when Evaluating Survival Analysis Models, which showed that the C-index reports illusory gains under dependent censoring and that roughly 72% of recent survival papers mismatch their metrics to their censoring assumptions. Selection is addressed by modeling the probability of resection from registry covariates and running a tipping-point analysis: how strong would an unobserved confounder (frailty, physician intuition) have to be to erase the flux–survival association entirely? The honest expectation, grounded in the retrospective-bias literature — A novel computed tomography radiomic nomogram for early evaluation of small airway dysfunction development shows that requiring complete future testing filters out exactly the patients who resolved or deteriorated, and the SEER-based nomogram work documents registry omissions of R0 margins and CA19-9 — is that unobserved confounding cannot be fully corrected, only bounded, and the proposal's contribution is to make the bound explicit rather than pretend it away.
Baselines
The "do nothing" reference point is conventional staging: in resectable pancreatic head cancer, AJCC 8th edition staging yields a C-index of 0.518 to 0.581 Predictive value of current nodal staging systems and development of machine learning nomogram for resectable pancreatic head cancer: a population-based study and multicenter validation. This anchors how much headroom any molecular model has; if a flux model cannot clear roughly 0.6, it is not clinically interesting.
The standard-practice baseline is the current published approach: the GLUT1/ALDOB/G6PD expression signature scored via gene-set methods and fed to a Cox model with Kaplan-Meier stratification, exactly as in the PDAC metabolic-classification paper, which reports significantly worse survival for the glucomet subtype but no numeric concordance statistic — no reported result found on Lacuna for a numeric C-index of the glucomet-PDAC expression signature. I will therefore run it myself, on the same cohorts, with the same compute and tuning budget as my model, so the comparison is like-for-like rather than against an unreported number.
The strongest existing competitor is the machine-learning nomogram pipeline from the SEER pancreatic study — Stepwise Cox feature selection combined with Survival-SVM, selected from 101 algorithm combinations under leave-one-out cross-validation — which reports C-index 0.675 in SEER training and 0.664 on a 670-patient external Chinese multicenter cohort Predictive value of current nodal staging systems and development of machine learning nomogram for resectable pancreatic head cancer. I will fit the same algorithm family on the clinical and pathological variables available in the PDAC cohorts, under the same cross-validation and tuning budget as my pipeline. Beating it means a higher concordance under a censoring-robust metric, with non-overlapping confidence intervals across the held-out cohorts, plus better survival-curve calibration than either baseline.
Two ablations isolate the proposal's key ingredients. Replacing the BMFA flux posteriors with point estimates from flux variability analysis removes the uncertainty propagation and tests whether the Bayesian ingredient matters or a cheap point estimate suffices; the BMFA-versus-FVA gap of 41% versus 9% F1 on unobserved flux prediction suggests it will, but that evidence comes from microbial systems and must be checked in this setting. Replacing the calibrated flux features with raw expression features of the same genes, while keeping the survival machinery identical, isolates whether the organoid calibration adds anything beyond what a clinician already has in a transcriptomic cohort. Finally, running the full pipeline unweighted versus resection-weighted measures how much of the flux–survival association is an artifact of who gets surgery.
Why not just…?
Why not just use the expression signature? This is the most serious alternative, and it may well match the flux-calibrated model in predictive accuracy — that possibility is exactly what the expression-versus-calibration ablation is designed to detect, and if the ablation shows parity, the honest conclusion is that calibration adds interpretation (which pathway is actually misrouted) rather than prediction, and the paper will say so. The counter-evidence is mechanistic, not yet predictive: the PDAC organoid work itself found that expression levels of GLUT1 and G6PD did not reveal the resistance mechanism until $^{13}$C tracing exposed the carbon flow into pyrimidine pools, and Bayesian Metabolic Flux Analysis reveals intracellular flux couplings shows static measurements cannot infer intracellular reaction rates while tracer-constrained posteriors can. Whether that mechanistic resolution translates into better survival discrimination is an open empirical question this proposal settles rather than assumes.
Why not just apply propensity scores or inverse probability weighting and be done? These adjust only for observed covariates. The retrospective-bias literature is explicit that physician intuition, frailty, and subtle symptom presentations — the very things that drove surgical selection — are unobserved in the records A novel computed tomography radiomic nomogram for early evaluation of small airway dysfunction development, and registries omit known confounders such as resection margin status and CA19-9 Predictive value of current nodal staging systems. My pipeline does use inverse-probability weighting for the observed part, but the real contribution is the tipping-point analysis that quantifies how much unobserved confounding the reported association can survive. A reviewer who believes weighting solves the problem is invited to read the sensitivity result: if a modest unobserved confounder erases the association, both my model and the naive one were equally wrong, and now we know it.
Why not just report the C-index like everyone else? Because Position: Stop Chasing the C-index when Evaluating Survival Analysis Models demonstrates through synthetic experiments that under non-random or dependent censoring — precisely what truncated follow-up produces when sicker patients drop out — the C-index shows apparent improvement while true accuracy degrades, and its literature survey found roughly 72% of recent survival papers misaligned between metrics and censoring assumptions. The fix costs little: evaluate with metrics satisfying that paper's five desiderata (proper scoring, interpretability, model-agnosticism, calibration sensitivity, censoring robustness), and additionally measure how the C-index itself inflates as I administratively truncate follow-up, so the censoring stress test doubles as a demonstration of the metric failure on real data.
Proposed Experiments
Experiment one validates the calibration map on the 28 organoids: with a rotating held-out set, predict held-out flux posteriors from expression alone and compare against flux-variability point estimates, reporting the same F1-style coverage metrics used in the BMFA paper (41% versus 9% precedent). Success is prediction of held-out pathway fluxes materially above the point-estimate baseline.
Experiment two runs the cohort predictions: apply the map to TCGA PDAC ($n=156$), Bailey ($n=55$), and the two validation cohorts ($n=283$, $n=285$), fit the censored survival model, and compare against staging, the expression signature, and the Stepwise-Cox/Survival-SVM nomogram family, using censoring-robust metrics alongside C-index for reference. All models get the same folds, tuning budget, and feature counts.
Experiment three is the censoring stress test: administratively censor at 12, 24, and 36 months, and track how each model's apparent performance and the flux–hazard estimate change; a model whose ranking and calibration are stable under truncation, where competitors' C-index inflates, is the decisive positive signal.
Experiment four is the selection audit: estimate resection probability from registry covariates, re-weight the cohort association, and run the tipping-point analysis over unobserved confounding. The headline output is a range: the flux–survival hazard ratio is at least X and at most Y in the general patient population, versus the uncorrected point estimate Z.
Experiment five is a synthetic ground truth: simulate patients with known flux-to-hazard relationships, assemble a retrospective cohort through a resection filter and truncated follow-up, and verify the pipeline recovers the true association while the naive expression Cox and unweighted models do not. If the pipeline cannot recover a known effect in simulation, the real-data results should not be trusted — and the paper will report that first.
References
- Metabolic classification suggests the GLUT1/ALDOB/G6PD axis as a therapeutic target in chemotherapy-resistant pancreatic cancer
- Bayesian metabolic flux analysis reveals intracellular flux couplings
- Predictive value of current nodal staging systems and development of machine learning nomogram for resectable pancreatic head cancer: a population-based study and multicenter validation
- Robustness of Multimodal Foundation-Model Forecasting for Postoperative Cancer Outcomes
- MIRTH: Metabolite Imputation via Rank-Transformation and Harmonization
- A Deep Variational Approach to Clustering Survival Data
- Position: Stop Chasing the C-index when Evaluating Survival Analysis Models
- A novel computed tomography radiomic nomogram for early evaluation of small airway dysfunction development