Metabolic Reserve as a Modifier of Lung Dose Tolerance

Radiotherapy planners estimate the risk of damaging healthy organs with a normal tissue complication probability (NTCP) model. This is a dose-response curve that maps the radiation dose an organ receives to the probability of a complication such as radiation pneumonitis, an inflammation of the lung. These curves treat every patient as biologically identical, so two people with the same dose map get the same predicted risk. Clinical data say otherwise. In thoracic radiotherapy, pre-existing chronic obstructive pulmonary disease (COPD) and poor baseline lung function predict severe pneumonitis on top of dose, and a single variant in a folate-metabolism gene splits pneumonitis incidence into 19.7% versus 6.7% at matched dose. Separately, blood metabolomics (measuring hundreds of small molecules in serum by mass spectrometry) can read out a person's metabolic state. Failing mitochondria, accumulated energy-cycle intermediates and amino-acid wasting mark the COPD patients who later die.

We propose to test one specific idea. A small, pre-specified blood panel of metabolic reserve, measured before and early during treatment, shifts where a patient sits on the lung dose-response curve: it moves the dose at which half of patients develop pneumonitis. It does not simply add a fixed amount of risk. The distinction matters for planning. If metabolism only adds risk, it changes counseling but not the plan. If it changes dose tolerance, then low-reserve patients gain more from sparing lung dose, and they are the ones who should be prioritized for lung-sparing techniques. Asking "who benefits from dose reduction" is a causal question about heterogeneous effects. That forces us to handle the confounding that, according to a review of 98 NTCP studies, no causal recommendation in that literature had adjusted for.

Core Claim

Our hypothesis is that the patient-specific tolerance dose for grade ≥2 radiation pneumonitis depends on a low-dimensional metabolic reserve score. The score shifts tolerance more than any equally simple clinical surrogate does (a COPD diagnosis, FEV1, smoking), and more than a germline folate-metabolism genotype does. We would encode this in the standard Lyman-Kutcher-Burman (LKB) model, the classical NTCP curve. The LKB model summarizes a non-uniform dose map as one equivalent uniform dose and passes it through a normal-distribution curve centered at $TD_{50}$, the dose giving 50% complication probability. We make $TD_{50}$ patient-specific:

$$NTCP_i = \Phi\left(\frac{gEUD_i - TD_{50}\, e^{\gamma^\top z_i + \delta^\top c_i}}{m\, TD_{50}\, e^{\gamma^\top z_i + \delta^\top c_i}}\right)$$

Here $\Phi$ is the standard normal cumulative distribution, $gEUD_i$ is patient $i$'s generalized equivalent uniform lung dose, $m$ is the slope parameter, $z_i$ is the metabolic reserve score (a few components), $c_i$ are measured clinical confounders, and $\gamma$ is the quantity of interest. Because $z_i$ enters through the tolerance dose, a nonzero $\gamma$ means the absolute benefit of lowering $gEUD_i$ differs between patients. A purely additive logistic covariate would imply something different. The two forms are directly testable against each other.

The novelty is narrow and checkable. The closest prior work, the MTHFR study by Genetic Variants in MTHFR Gene Predict ≥ 2 Radiation Pneumonitis in Esophageal Squamous Cell Carcinoma Patients Treated with Thoracic Radiotherapy, found that rs1801131 predicts pneumonitis after adjusting for mean lung dose and V20 (HR 0.339). It used a static germline marker in an additive Cox model, never placed the marker inside a dose-response curve, and never measured the functional metabolite pool. The serum work of Serum metabolic profiling enables diagnosis, prognosis, and monitoring for brainstem gliomas supplies the method we reuse. It ran fast, low-volume serum mass spectrometry across radiotherapy and found eight temporal metabolite patterns that separate shrinking from growing tumors (AUC 0.818). It studied only the tumor, never normal-tissue injury or dose.

Research Program

The motivating evidence comes from two papers that have never been connected. Analysis of clinical and dosimetric factors associated with severe acute radiation pneumonitis in patients with locally advanced non-small cell lung cancer treated with concurrent chemotherapy and intensity-modulated radiotherapy showed that severe pneumonitis tracks both the low-dose lung volume V10 (the fraction of lung receiving at least 10 Gy) and clinical COPD. Plasma metabolomics and clinical predictors of survival differences in COPD patients showed that COPD patients with poor outcomes carry a characteristic plasma signature. It includes raised succinate, alpha-ketoglutarate and lactate, pentose-phosphate activity, and depleted branched-chain amino acids, and it classified survivors at 85% accuracy with an SVM. Our bet is that this signature, not the COPD label, is what lowers lung tolerance to the low-dose bath.

To avoid the discovery-by-overfitting problem common in metabolic gene-signature work, we fix the score before seeing outcomes. It has two components: a mitochondrial-stress component built from the COPD signature metabolites, and a one-carbon component of homocysteine, folate and S-adenosylmethionine, motivated by the MTHFR result. We compute it from a targeted LC-MS assay with standardized fasting morning draws, because the COPD paper itself warns that diet, smoking and inflammation move plasma metabolites. Untargeted serum profiling with the brainstem-glioma platform runs in parallel on the same draws, at baseline, around fraction 10, around fraction 20 and at the end of treatment. It is used only in a separately labeled exploratory analysis of early trajectories.

The cohort is prospective: about 400 patients receiving definitive thoracic radiotherapy for locally advanced non-small cell lung cancer or esophageal squamous cell carcinoma. The 15.8% grade ≥2 pneumonitis incidence reported in the MTHFR cohort implies roughly 60 events. That supports a model with a handful of interaction parameters but nothing high-dimensional, which is why the score is pre-specified. We collect full 3D dose grids, MTHFR genotype, spirometry, smoking status, COPD diagnosis, body composition, tumor volume and the chemotherapy regimen. Chemotherapy is included because Impact of Chemotherapy Regimens on Normal Tissue Complication Probability Models of Acute Hematologic Toxicity in Rectal Cancer Patients Receiving Intensity Modulated Radiation Therapy With Concurrent Chemotherapy From a Prospective Phase III Clinical Trial showed that regimen alters normal-tissue dose response.

Confounding is handled explicitly, in the spirit of Considering causality in normal tissue complication probability model development: a literature review. Before analysis we write down a causal diagram. In it, tumor size and location drive lung dose, while tumor burden, cachexia and smoking drive both metabolic state and pneumonitis. The metabolic score must therefore be shown to act on tolerance after conditioning on tumor volume and cachexia. Otherwise it is just a proxy for advanced disease.

Baselines

The naive reference point is the conventional LKB model on lung gEUD, together with logistic regression on mean lung dose. On a 70-patient VMAT cohort, Dosiomics: Extracting 3D Spatial Features From Dose Distribution to Predict Incidence of Radiation Pneumonitis reported AUC 0.665 for mean dose and 0.744 for a multivariate NTCP model. We refit these on our cohort rather than borrowing those numbers.

The strongest dose-only competitors are spatial models. Dosiomics, meaning texture and gradient features computed from the 3D dose grid, reached multivariate AUC 0.782 in that same paper. The 3D convolutional network of Prediction of Radiation Pneumonitis With Dose Distribution: A Convolutional Neural Network (CNN) Based Model, pre-trained on video and fine-tuned on dose grids, reached AUC 0.842, the highest figure reported in the file for this endpoint. Because that number comes from 70 patients, we expect optimism. We will reimplement both on our data with identical cross-validation folds, the same feature-selection or tuning budget, and nested selection, and report their AUCs alongside ours.

The strongest biology-aware competitor is a clinical-plus-genotype additive model: dose plus COPD, FEV1, smoking and MTHFR rs1801131, as in the MTHFR study.

Two ablations isolate the key ingredient. The first enters the same metabolic score additively on the logit scale instead of through $TD_{50}$, which tests the tolerance-shift claim directly. The second swaps the score for a size-matched random set of measured metabolites, which tests whether the pre-specified biology matters at all.

Beating the baselines has two conditions. First, the tolerance-shift model must improve out-of-sample log-likelihood and calibration over the clinical-plus-genotype model, with $\gamma$ excluding zero after confounder adjustment. Second, added to the best dosiomic or CNN model, it must improve AUC and net benefit on decision curves. Discrimination alone is not the target. A model that matches the CNN's AUC but correctly identifies who gains from dose sparing is the result we care about.

Why not just…?

Why not just use COPD diagnosis, FEV1 and smoking history? This is the most serious threat. These variables are free and already significant in the severe-pneumonitis cohort. The COPD metabolomics paper argues that clinical indices miss metabolic heterogeneity, but it did not test radiotherapy toxicity, so that is motivation rather than evidence. We settle the question with two tests. The first asks whether the metabolic score retains a nonzero $\gamma$ with all three clinical variables in $c_i$. The second asks whether it still stratifies tolerance inside the COPD-positive and COPD-negative strata separately. If it adds nothing in either test, we will report that the cheap surrogate suffices.

Why not just add metabolomics features to a dosiomics or deep-learning model? With roughly 60 events, hundreds of serum features plus dose textures will overfit. More fundamentally, an additive black box predicts risk but cannot say whether lowering dose helps one patient more than another, and that is the planning question. We still run this comparison, using untargeted features stacked onto dosiomics under the same nested cross-validation. If it beats the pre-specified score out of sample, we would adopt its signal while keeping the tolerance-shift structure.

Why not just genotype MTHFR and other variants instead of measuring metabolites? Genotyping is stable, cheap and immune to fasting effects. That is a real advantage, and the HR of 0.339 shows the folate pathway matters. A germline variant, however, fixes predisposition and cannot track the functional folate pool or within-treatment change. Our test is whether the one-carbon metabolite component carries signal after conditioning on rs1801131 genotype. If it does not, the genotype is the better clinical tool for that axis, and we will say so.

Proposed Experiments

The primary analysis estimates $\gamma$ in the tolerance-shift LKB model. It conditions on the pre-registered confounder set, uses a bootstrap for uncertainty, and is compared against the additive ablation by cross-validated likelihood.

The decisive check is a dose-by-reserve interaction visible without the model. We split patients by tertile of the reserve score and plot empirical pneumonitis rates against V10 and mean lung dose. Low-reserve patients should show a curve shifted to lower doses, not just a parallel lift. An important falsification test follows from this. If the curves are parallel on the logit scale, the tolerance claim fails and metabolism becomes a counseling covariate only.

A negative-control check guards against the score being a tumor-burden proxy. It tests whether the score predicts tumor response, a process driven by the tumor-side metabolic reprogramming described in Proteomic and metabolomic signatures of rectal tumor discriminate patients with different responses to preoperative radiotherapy. It also tests whether the score loses its effect after adjusting for tumor volume and cachexia.

A planning consequence analysis uses each patient's clinical plan and a re-optimized plan that lowers V10. We compute the predicted risk reduction under the dose-only and tolerance-shift models and ask whether the two models disagree on which patients clear a meaningful benefit threshold. This is a hypothesis-generating result: it shows what would change in practice, not that the change helps.

Finally, an exploratory longitudinal analysis asks whether the change in serum profile between baseline and fraction 10 or 20 predicts pneumonitis before symptoms, beyond the baseline score. It reuses the high-throughput serum platform and keeps that result clearly separate from the confirmatory test.