Blood tests for circulating tumor DNA (ctDNA) sometimes disagree sharply with what actually happens to the patient. A lesion that looks ominous on imaging may shed almost nothing into the bloodstream, while a small, quietly progressing lesion dominates the blood signal. This proposal argues that the discordance is not noise to be averaged away but structure to be modeled: the journey of molecular debris from tissue to blood is a physical communication channel, and tumor hypoxia is what sets that channel's properties. Dense, oxygen-starved, unperfused tumor cores are strong generators of DNA but terrible transmitters — the debris they release is attenuated, fragmented, and chemically degraded before reaching plasma. Well-perfused tumor rims are weak generators but clean channels. The same mathematics that information theorists built for "degraded channels" — channels whose output is a strictly noisier copy of another receiver's output — tells us exactly when a pooled blood draw is a faithful, invertible summary of the tissue, and when it cannot be, no matter how sensitive the assay.
Two clinical phenomena motivate this framing. In glioblastoma, post-treatment MRI often shows enlarging enhancing lesions that mimic recurrence while the patient is actually responding — pseudoprogression — and current imaging criteria resolve this only by waiting out a six-month observation window (Advanced Imaging Techniques for Differentiating Pseudoprogression and Tumor Recurrence After Immunotherapy for Glioblastoma). In recovered COVID-19 patients, ultrasensitive RT-PCR keeps detecting viral RNA for weeks after patients are well and non-infectious, because the assay reads inert fragments that carry no information about viable virus (Positive RT-PCR detected in patients recovered from COVID-19); these re-detectable positive patients in fact have milder lung involvement and shorter readmissions than typical patients (Atypical presentations of coronavirus disease 2019 (COVID-19) from onset to readmission). Both are the same failure mode: an over-sensitive peripheral assay detecting debris whose mutual information with the clinically relevant variable — viable, progressing burden — is near zero. We propose to quantify that information directly.
Core Claim
Per-lesion ctDNA shedding is not a free statistical weight, as the Lesion Shedding Model treats it (Lesion Shedding Model: unraveling site-specific contributions to ctDNA), but a physically structured quantity determined by hypoxia and vascular failure. Multiparametric MRI already measures the channel's ingredients: the volume transfer constant $K^{trans}$ and flux rate $K_{ep}$ capture vascular permeability, while the apparent diffusion coefficient (ADC) and extracellular volume fraction $V_e$ capture cell density and stromal barrier — the very axes along which breast tumors partition into a hypoxic, unperfused, cellular core (the low-vascularity/high-cellularity habitat) and a perfused rim (Study on heterogeneity of vascularity and cellularity via multiparametric MRI habitat imaging in breast cancer). Writing the plasma observation as
$$Y \;=\; \sum_{k=1}^{K} h_k\, S_k \;+\; N, \qquad h_k \;\propto\; K^{trans}_k \cdot \phi(\text{ADC}_k, V_{e,k}),$$
where $S_k$ is the DNA generation rate of lesion or habitat $k$ (rising with cellularity and necrosis), $h_k$ is its transport gain to blood, and $N$ is sampling noise, turns the shedding problem into a channel-coding problem. The key information quantity is the bound on what plasma reveals about viable progressing burden $V$:
$$I(V; Y) \;\le\; \sum_{k=1}^{K} \log\!\Big(1 + \frac{h_k^{2}\, \sigma^{2}_{V,k}}{\sigma^{2}_{N} + h_k^{2}\, \sigma^{2}_{D,k}}\Big),$$
where $\sigma^{2}_{D,k}$ is the variance of inert debris generated by habitat $k$. A hypoxic core has large $\sigma_D$ and small $h$: it contributes detectable signal but almost no information. This is the biological counterpart of the wiretap regime in which the eavesdropper's channel is stronger than the legitimate receiver's, where the optimal strategy qualitatively changes (Secure Broadcasting Using Independent Secret Keys) — here, the assay "hears" everything and learns little, while the slowly-evolving clinical trajectory is the receiver that actually matters.
When a single lesion dominates, its channel gains are ordered and the plasma observation is a degraded — strictly filtered, invertible — copy of the interstitial signal; attribution and viability inference are then tractable. When multiple lesions or habitats cross in channel quality (one hypoxic and cellular, another sparse but hyper-perfused), no scalar ordering exists. This is precisely the situation in which Multi-Antenna Gaussian Broadcast Channels with Confidential Messages showed that naive layering fails and capacity is governed by generalized eigenvalues of the channel matrices, and in which standard inner and outer bounds diverge for non-degraded channels (Generalized Multiple Access Channels with Confidential Messages). We will import those tools to derive hard, assay-independent limits on per-lesion attribution from plasma, using the matrix entropy-power machinery of A Vector Generalization of Costa's Entropy-Power Inequality with Applications to handle the directional, non-isotropic noise structure of multi-habitat transport.
Research Program
The program has three parts. First, theory: characterize when the habitat-to-plasma channel is stochastically degraded (ordered $h_k$, single dominant source) and derive Sato-type genie-aided outer bounds on attribution accuracy in the non-degraded multi-lesion case, in the spirit of the converse techniques in Multi-Antenna Gaussian Broadcast Channels with Confidential Messages. These bounds are assay-independent — they cap what any sequencing depth can achieve. Second, estimation: replace the Lesion Shedding Model's Dirichlet-sampled, Chebyshev-fitted scalar weights with a plug-in channel model whose $h_k$ are read directly from quantitative MRI or, where MRI is unavailable, from lesion size, anatomical site, and necrosis fraction. The model has two to three physical parameters per habitat rather than unconstrained weights, so it should generalize where the black-box fit cannot. Third, cross-domain validation: the same multi-compartment channel structure should explain the viral post-recovery data — an infant shedding SARS-CoV-2 RNA in feces for 100 days while nasopharyngeal swabs stayed negative (An Asymptomatic SARS-CoV-2-Infected Infant With Persistent Fecal Viral RNA Shedding in a Family Cluster: A Rare Case Report) is exactly two reservoirs with wildly asymmetric channel gains, and the impaired immune activation signature of retest-positive patients (Transcriptomic characteristics and impaired immune function of patients who retest positive for SARS-CoV-2 RNA) supplies a mechanistic knob ($S_k$ clearance) the tumor case lacks. Treating the viral setting as a controlled testbed for the mathematics, not as an equal-weight application, keeps the proposal focused.
A known limitation we will confront directly: the Lesion Shedding Model cohorts (metastatic gastrointestinal cancer patients TPS037 and TPS177) have multi-region sequencing but no paired mpMRI, so the channel inputs there must come from pathological proxies. The breast habitat cohort has mpMRI but no ctDNA. Part of the deliverable is therefore a prospective paired mpMRI–liquid-biopsy collection protocol, designed so the theory says in advance which patients should show assay–imaging discordance (those with large hypoxic-core fractions, which in the breast cohort were 56.9% of tumor volume in HER2-negative versus 34.5% in HER2-positive disease, $p = 0.005$).
Baselines
The naive reference point is guideline-based observation: the iRANO six-month window in neuro-oncology (Advanced Imaging Techniques for Differentiating Pseudoprogression and Tumor Recurrence After Immunotherapy for Glioblastoma) and, on the PET side, the conventional tumor-to-background ratio $\text{TBR}_{max}$, which achieved 74% accuracy, 81% sensitivity, and 67% specificity on the post-chemoradiation glioma cohort (FET PET Radiomics for Differentiating Pseudoprogression from Early Tumor Progression in Glioma Patients Post-Chemoradiation). The strongest existing methods are supervised: the four-feature Random Forest from the same FET PET study, which reached 100% sensitivity and 100% negative predictive value for pseudoprogression but only 70% overall accuracy on its independent test set; the Random Forest radiomics-plus-clinical model of Radiogenomic analysis of hypoxia pathway is predictive of overall survival in Glioblastoma, which achieved a Concordance Index of 0.83 for short- versus long-term glioblastoma survival on a 30-case validation set; and the scalar-weight Lesion Shedding Model itself, whose assigned weights correlated 0.97 with predicted network edges in the four-lesion patient TPS037 but only 0.54 in the seventeen-lesion patient TPS177 (Lesion Shedding Model: unraveling site-specific contributions to ctDNA). Each baseline will be run on the identical cohort and split with the same feature and tuning budget as our model, and each comparison has a pre-stated success criterion: on the 34-patient FET PET cohort we must exceed 74% accuracy (the better of TBRmax and the Random Forest's 70%) using a mechanistic statistic with no more free parameters than the Random Forest's four features; on the TCIA glioblastoma cohort (115 studies, 85 training with 3-fold cross-validation, 30 independent validation) we must improve on the C-index of 0.83; on the shedding cohorts we must beat the scalar model's multi-lesion correlation of 0.54 while fitting fewer parameters. The critical ablation is the same estimator with the habitat structure collapsed to a single scalar channel gain — a well-mixed one-compartment model — which isolates whether the matrix, hypoxia-resolved channel is what earns the performance, rather than generic smoothing or regularization.
Why not just…?
Why not just train a Random Forest on radiomics? This is the most serious alternative, and the evidence for it is strong: C-index 0.83 in glioblastoma survival prediction and 100% sensitivity for pseudoprogression are hard numbers to argue with (Radiogenomic analysis of hypoxia pathway is predictive of overall survival in Glioblastoma; FET PET Radiomics for Differentiating Pseudoprogression from Early Tumor Progression in Glioma Patients Post-Chemoradiation). But the FET PET result rests on 34 patients inflated to 102 datasets by repeated manual segmentations, with 70% overall accuracy, and supervised models are uncalibrated mappings that cannot extrapolate to lesion counts or habitat configurations absent from training. The decisive experiments are twofold: a head-to-head on the same data with matched budgets, which we may lose and will report honestly; and a transfer test — fit the channel parameters on one cohort, apply them unchanged to another — which a Random Forest cannot attempt without retraining. If the mechanistic model matches the Random Forest on-cohort and transfers where the Random Forest cannot, the mechanism is paying for itself.
Why not just extend the observation window? The iRANO six-month window is clinically safe and requires no modeling (Advanced Imaging Techniques for Differentiating Pseudoprogression and Tumor Recurrence After Immunotherapy for Glioblastoma), and simply waiting out re-detectable positive patients is likewise benign (Atypical presentations of coronavirus disease 2019 (COVID-19) from onset to readmission). The cost is delay: true progressors wait six months for second-line therapy, and TBRmax misclassifies roughly one in three non-progressing patients at 67% specificity (FET PET Radiomics...). Our claim is that a calibrated early discriminator is possible; if the channel model cannot beat 74% accuracy on the same cohort, the observation window remains the right clinical answer and our contribution collapses to the theory, which we would say plainly.
Why not just use compartmental ODEs? Standard one- or two-compartment exponential-decay models are the obvious first tool for clearance kinetics. They fail here for documented reasons: shedding showed no correlation with tumor size or genetic divergence, and the multi-lesion correlation collapsed from 0.97 to 0.54 as lesion count grew (Lesion Shedding Model: unraveling site-specific contributions to ctDNA); and 100 days of fecal viral shedding alongside repeatedly negative nasal swabs (An Asymptomatic SARS-CoV-2-Infected Infant With Persistent Fecal Viral RNA Shedding in a Family Cluster: A Rare Case Report) cannot be produced by any well-mixed scalar decay. Our scalar-channel ablation is precisely this alternative run on the same data; if a two-compartment ODE matches the habitat-resolved model on every benchmark, the extra structure is not earning its keep and we would report the simpler model as sufficient.
Proposed Experiments and Decisive Checks
Five experiments, each tied to a benchmark in the file. First, attribution on the Lesion Shedding Model cohorts: predict per-lesion plasma contributions from channel-model weights using pathological proxies in place of MRI, and test whether the multi-lesion case (TPS177, 17 lesions) beats the scalar model's correlation of 0.54 with fewer fitted parameters; a companion simulation study, parameterized by the habitat distributions measured in the 76-patient breast cohort (Study on heterogeneity of vascularity and cellularity via multiparametric MRI habitat imaging in breast cancer), establishes the theoretical recovery limits that the Sato-type bounds predict. Second, pseudoprogression discrimination on the 34-patient FET PET dataset using its published 70/30 split, targeting accuracy above 74% with a two-parameter mechanistic viability index. Third, survival stratification on the TCIA glioblastoma cohort of 115 studies with its 85/30 train–validation structure, testing whether the channel-degradation statistic adds to the reported C-index of 0.83 (Radiogenomic analysis of hypoxia pathway is predictive of overall survival in Glioblastoma). Fourth, response prediction in the breast habitat cohort: does the channel model's effective informative-shedding fraction predict Miller-Payne pathological response beyond what the raw hypoxic-core fraction already gives? Fifth, the cross-domain check on the viral side, using the public transcriptomic dataset GSE166253 and the documented asymmetric-clearance cases to fit the multi-reservoir channel and confirm it reproduces negative-swab, positive-stool dynamics. The single most decisive check across all of them: a pre-registered prediction of which patients will show assay–imaging discordance, based solely on hypoxic-core fraction, tested on the prospective paired mpMRI–ctDNA collection. If hypoxia-structured channel gains do not predict discordance better than tumor size does — and size is documented not to predict shedding — the core claim is wrong and the theory, however elegant, describes a channel biology does not use.
References
- Advanced Imaging Techniques for Differentiating Pseudoprogression and Tumor Recurrence After Immunotherapy for Glioblastoma
- Positive RT-PCR detected in patients recovered from COVID-19
- Atypical presentations of coronavirus disease 2019 (COVID-19) from onset to readmission
- Lesion Shedding Model: unraveling site-specific contributions to ctDNA
- Study on heterogeneity of vascularity and cellularity via multiparametric MRI habitat imaging in breast cancer
- Secure Broadcasting Using Independent Secret Keys
- Multi-antenna Gaussian broadcast channels with confidential messages
- Generalized Multiple Access Channels with Confidential Messages
- A vector generalization of costa's entropy-power inequality with applications
- An Asymptomatic SARS-CoV-2-Infected Infant With Persistent Fecal Viral RNA Shedding in a Family Cluster: A Rare Case Report
- Transcriptomic characteristics and impaired immune function of patients who retest positive for SARS-CoV-2 RNA
- FET PET Radiomics for Differentiating Pseudoprogression from Early Tumor Progression in Glioma Patients Post-Chemoradiation
- Radiogenomic analysis of hypoxia pathway is predictive of overall survival in Glioblastoma