A Dose-Rate-Aware Gene Signature for FLASH Radiotherapy

FLASH radiotherapy delivers a full therapeutic radiation dose in under a second, at dose rates above 40 Gy/s, instead of over minutes at conventional rates. Animal and radiochemical evidence suggests this spares healthy tissue — probably by draining local oxygen faster than it can be replaced, so that normal cells transiently behave as if hypoxic — while tumors, many of which are already oxygen-starved, keep their sensitivity. At the same time, a separate literature builds prognostic gene signatures: weighted sums of tumor gene-expression measurements that stratify patients by survival risk, and in one glioblastoma study even identify which patients fail to benefit from radiotherapy. These two literatures have never touched. No transcriptomic predictor has ever been evaluated against, or built for, ultra-high dose rate delivery, and no FLASH planning system uses any molecular information about the patient.

This proposal argues the connection is not decorative. The single variable that FLASH biology says matters most — baseline tissue oxygenation and radical-handling capacity — is exactly what the transcriptome measures, and what no imaging or physical planning metric captures. The core deliverable is a gene signature augmented with a dose-rate interaction term, embedded as a patient-specific prior inside an existing Bayesian digital-twin optimizer, so that the model can recommend not just how much radiation a patient should receive but at what dose rate.

Core Claim

A static prognostic signature predicts outcome under one assumed treatment. A dose-rate-aware signature predicts how the difference between conventional and ultra-high dose rate delivery depends on the tumor's baseline molecular state. Concretely, we extend the linear radiosensitivity parameter $\alpha_{\text{RT}}$ used in a published glioma digital twin — currently calibrated from MRI tumor volumes alone — to depend on both a gene-expression risk score $S$ (a $\beta$-weighted sum of expression values, the standard signature form) and the mean dose rate $R$:

$$\alpha_{\text{RT}}(S, R) \;=\; \alpha_0 \exp\big(\gamma_1 S + \gamma_2\, S \log R\big)$$

where $\alpha_0$ is a population baseline, $\gamma_1$ captures conventional radiosensitivity, and $\gamma_2$ is the new quantity: whether a patient's molecular profile changes the benefit of dose rate itself. The logarithm in $R$ is not arbitrary — radiochemical measurements across 0.14 to 1500 Gy/s show oxygen consumption and peroxide yield regress linearly against log mean dose rate (coefficients −0.094, $p=0.0009$ and −0.46, $p=0.001$), with per-Gy oxygen consumption about 1.4× higher and peroxide yield about 3.3× lower at conventional versus ultra-high rates (Mean dose rate in ultra-high dose rate electron irradiation is a significant predictor for O2 consumption and H2O2 yield). If tissue transcriptional state sets the starting point for those kinetics — which that same paper shows, since oxygen depletion depended on initial oxygen concentration — then $\gamma_2$ should be estimable and nonzero.

Research Program

Three components, in increasing order of experimental risk.

First, a reanalysis of existing glioblastoma cohorts to calibrate the conventional arm. The published 5-gene immune signature (ARPC1B, FCGR2B, NCF2, PLAUR, S100A11) was derived on 166 TCGA training cases and validated on three CGGA cohorts of 325, 693, and 301 cases, where it predicted survival specifically among irradiated patients, with 1–4 year AUCs of 0.63–0.73, outperforming IDH status, age, and molecular subtype (Novel Immune-Related Gene-Based Signature Characterizing an Inflamed Microenvironment Predicts Prognosis and Radiotherapy Efficacy in Glioblastoma). That paper also found its high-risk group was severely hypoxic — the exact phenotype FLASH theory predicts is unaffected by dose-rate sparing (already oxygen-depleted) yet radioresistant under conventional delivery. We will test, in those same cohorts, whether hypoxia and oxygen-metabolism module scores moderate the signature's association with radiotherapy benefit, which calibrates $\gamma_1$ and bounds plausible $\gamma_2$ values. An honest limitation: no FLASH-treated patient with matched transcriptomics exists anywhere in the literature — this absence is the gap, cataloged plainly by the Radiotherapy Prognosis Database, whose 775 studies and 115,516 patients contain no dose-rate annotation at all (RTPDB: a database providing associations between genetic variation or expression and cancer prognosis with radiotherapy-based treatment). The human data can only calibrate; the FLASH side must be tested in vitro and in silico.

Second, an in vitro dose-rate-by-genotype matrix. Published work on HeLa cells at 2, 6, and 10 Gy/min found survival fractions differed significantly across dose rates ($p=0.04$) while every classical radiobiological parameter — $\alpha/\beta$, $D_0$, $D_q$, SF$_2$ (0.79, 0.78, 0.75) — showed no significant difference ($p>0.05$), meaning the standard phenomenological models are blind to dose rate even in a simple cell line (Effect of High Energy X-rays at Different Dose Rates on Biological Effects of Cervical Cancer HeLa Cells). We will extend this design: a panel of glioma and lung cell lines chosen to span hypoxia-signature and ferroptosis-pathway expression, irradiated across mean dose rates from conventional to 1500 Gy/s on the modified clinical linac configuration described in the radiochemical study, with clonogenic survival and matched RNA-seq at each point. The decisive readout is the slope of survival against log dose rate per line, and whether baseline expression of oxygen-metabolism, DNA-repair, and regulated-cell-death modules predicts that slope.

Third, embedding the signature as a molecular prior inside the digital twin. The published twin couples a tumor-growth ODE with linear-quadratic cell kill, calibrates proliferation and radiosensitivity by Bayesian MCMC on serial MRI, and optimizes schedules under uncertainty; on its in-silico 100-patient glioma cohort it raised median time to progression by about 6 days at standard dose and found 16.7% (10 Gy) median dose reductions for responsive tumors — but its $\alpha_{\text{RT}}$ is a blind parameter, and it never modeled dose rate (Predictive Digital Twin for Optimizing Patient-Specific Radiotherapy Regimens under Uncertainty in High-Grade Gliomas). We replace the literature prior on $\alpha_{\text{RT}}$ with the signature-informed prior above, add dose rate as a decision variable, and measure whether the twin's recommendations change and improve. The downstream proof of integration is on the 10-patient matRad proton transmission FLASH planning cohort, where 5-field FLASH plans already matched conventional proton SBRT on target uniformity ($D_{2\%}$ 105.3% versus 105.6%) with about 80% of target volume above 40 Gy/s (Dose rate and dose robustness for proton transmission FLASH-RT treatment in lung cancer); we will verify that signature-weighted biological objectives do not degrade these physical metrics.

Baselines

The do-nothing reference point is the uniform standard prescription — 60 Gy in 30 fractions with population-averaged radiosensitivity — which the digital-twin paper showed to be suboptimal for many simulated patients. Any personalized model must beat it by improving time to progression or reducing dose without losing tumor control, under the twin's own 100-patient simulated cohort.

The strongest existing competitor is the imaging-only digital twin itself, since it is the only published system that personalizes fractionation at all and already demonstrates gains: a median time-to-progression increase of approximately 6 days ($p<0.05$) and a 16.7% (10 Gy) median dose reduction in responsive tumors, with reductions up to 46.7 Gy in late progressors (Predictive Digital Twin for Optimizing Patient-Specific Radiotherapy Regimens under Uncertainty in High-Grade Gliomas). We will rerun it on the identical in-silico cohort, compute budget, and optimizer settings, changing only the prior on $\alpha_{\text{RT}}$. Beating it means the transcriptomic prior either extends the dose-reduction benefit to patients the imaging-only model classifies as unresponsive, or improves median time to progression beyond the 6-day gain, without any loss of tumor control in low-risk patients. If the molecular prior merely reshuffles recommendations without moving these endpoints, the claim fails.

On the prognostic side, the strongest competitor is the 5-gene GBM immune signature, reporting 1–4 year survival AUCs of 0.63 to 0.73 on its TCGA training and CGGA validation cohorts (Novel Immune-Related Gene-Based Signature Characterizing an Inflamed Microenvironment Predicts Prognosis and Radiotherapy Efficacy in Glioblastoma). We will evaluate on the same 166-case training and 325/693/301-case validation cohorts with the same gene-discovery budget, but the bar is different from matching its AUC: our model must additionally show a significant dose-rate interaction (nonzero slope of survival versus log dose rate predicted by baseline expression), which a static signature cannot express by construction.

The key ablation sets $\gamma_2 = 0$ in the equation above: same signature, same twin, same data, no dose-rate term. This isolates whether dose-rate awareness adds anything beyond a conventional molecularly informed radiosensitivity prior, and it is the ablation a reviewer should demand first.

Why not just…?

Why not just use clinical and pathological markers instead of transcriptomics? For conventional-risk stratification this is a fair instinct, but the published signature already beat IDH status, age, and molecular subtype on its own cohorts (Novel Immune-Related Gene-Based Signature Characterizing an Inflamed Microenvironment Predicts Prognosis and Radiotherapy Efficacy in Glioblastoma), and, more decisively for this proposal, no clinical marker has any plausible mechanism for predicting a dose-rate interaction, since none measures tissue oxygen handling. We will still include IDH, age, and subtype as covariates in every model so the comparison is direct.

Why not just fit linear-quadratic parameters at each dose rate? This is the strongest objection, and the HeLa result is the evidence against it in its naive form: survival differed across dose rates ($p=0.04$) while fitted LQ parameters did not ($p>0.05$), so per-dose-rate fitting of the same parameters absorbs nothing (Effect of High Energy X-rays at Different Dose Rates on Biological Effects of Cervical Cancer HeLa Cells). The honest version of the objection is fitting a free dose-rate correction term per patient, which is mathematically what our $\gamma_2$ does — the difference is that we constrain it with a measurable baseline (gene expression) rather than leaving it a free parameter, which is what makes it predictive rather than descriptive. The in vitro matrix is the experiment that settles this: if the free-parameter fit and the signature-constrained fit perform equally out-of-sample, the transcriptome adds cost without signal, and we would drop the molecular claim.

Why not just treat FLASH as pure physics, since radiolysis happens in microseconds? At the microsecond scale, genes indeed change nothing. But the same radiochemical study shows the outcome depends on the initial oxygen concentration of the medium — the starting state the tissue brings to the beam (Mean dose rate in ultra-high dose rate electron irradiation is a significant predictor for O2 consumption and H2O2 yield), and reviews of the mechanism emphasize downstream biological cascades — mitochondrial preservation, immune and inflammatory reprogramming — not just radiolysis (FLASH-RT for pulmonary protection: a comprehensive review of mechanisms, technological advances, and clinical translation). The claim is precisely bounded: the transcriptome predicts the starting state and the downstream response, not the radiolysis. If the in vitro matrix shows no correlation between baseline expression and dose-rate survival slope, this objection wins and the proposal's core is falsified.

A fourth alternative — adaptive anatomical planning on MR-LINAC-type systems — is addressed by the twin paper itself: imaging measures net tumor volume and misses the molecular drivers of resistance (New Approaches in Radiotherapy, Predictive Digital Twin for Optimizing Patient-Specific Radiotherapy Regimens under Uncertainty in High-Grade Gliomas); anatomical adaptation and molecular priors also answer different questions, so we treat it as complementary rather than competing.

Proposed Experiments and Decisive Checks

The falsifiable predictions are, in order: first, in the in vitro matrix, the slope of clonogenic survival against log mean dose rate varies across cell lines and is significantly predicted by baseline hypoxia and oxygen-metabolism module scores (with classical LQ parameters flat, replicating the HeLa finding); second, in the 100-patient in-silico twin, the signature-informed prior on $\alpha_{\text{RT}}$ improves on the imaging-only twin's reported ~6-day median time-to-progression gain or extends its 10 Gy dose-reduction benefit to high-risk patients; third, on the matRad 10-patient lung cohort, signature-weighted objectives preserve the physical benchmarks of the published 5-field FLASH plan ($D_{2\%}$ near 105.3%, roughly 80% of volume above 40 Gy/s). The clinical-trial landscape (FAST-01, FAST-02, IMPULSE) is moving toward FLASH in patients — no reported result found on Lacuna for their outcomes — which means the first FLASH-treated transcriptomic cohorts will appear within trial pipelines; a prospective archival-biopsy add-on to such a trial is the natural translation path for a validated $\gamma_2$.