Radiation biology

What this is

Radiation biology in this corpus spans (a) quantitative imaging biomarkers derived from radiotherapy planning or diagnostic CT/MRI (radiomics, radiogenomics, deep-learning prognostic models), (b) molecular/genomic correlates of radiation response and radioresistance (DNA-damage response (DDR), hypoxia, IFN-related DNA-damage-resistance (IRDS), late genomic sequelae such as radiation-associated sarcoma), (c) biomarker-designed radiotherapy clinical trials (ROBIN consortium, nivolumab + IMRT reirradiation) and the datasets/infrastructure that support them (RADCURE, TCIA HNSCC, NSCLC-Radiogenomics, TCGA-GBM/LGG MRI), and (d) preclinical radiosensitization strategies. The corpus has a particularly deep imaging-biomarker substrate in head-and-neck cancer and NSCLC, and offers explicit contradictions about whether radiomic signatures generalize across institutions.

What the corpus says

Imaging biomarkers & radiomics (signatures, datasets, deep learning)

  • A fixed four-feature CT radiomic signature (Statistics Energy, Shape Compactness, Grey Level Nonuniformity, wavelet Grey Level Nonuniformity HLH) trained on 422 curative-intent NSCLC patients (Lung1/MAASTRO) generalized without retraining to independent NSCLC (Lung2, CI=0.65, P=2.91×10⁻⁹), HNSCC (H&N1, CI=0.69; H&N2, CI=0.69), and retained performance in HPV-negative HNSCC (CI=0.66) and within both radiation-only and concurrent chemoradiation subgroups PMID:24892406.
  • Two of the four Aerts signature features (intratumour-heterogeneity descriptors Grey Level Nonuniformity and wavelet Grey Level Nonuniformity HLH) were significantly correlated with cell-cycling/proliferation gene sets by GSEA on the 89-patient Lung3 NSCLC radiogenomics cohort, providing an initial biological anchor for CT-derived heterogeneity PMID:24892406.
  • In a crowdsourced benchmark on 2,552 Princess Margaret HNSCC patients (RADCURE training cohort) with external validation on 873 patients (HN1 MAASTRO, MDACC, GPCCHN Poznań), the best-performing 2-year OS model was a multitask logistic regression over EMR features (age, T/N/overall stage, disease site, performance status, HPV status, dose, systemic therapy) plus primary tumor volume (AUROC=0.823, C-index=0.801); no radiomics-only model beat any EMR-only model, and replacing tumor volume with a deep-CNN image embedding lowered AUROC from 0.823 to 0.766 PMID:37397861.
  • The RADCURE dataset (3,346 HNSCC planning CTs with manual target and 19 organ-at-risk contours, standardized RT-STRUCT nomenclature, median 5-year follow-up, 60% surviving) is now the largest public resource for head-and-neck RT modelling PMID:38362943.
  • The TCIA HNSCC collection from MD Anderson (215 patients, 433,384 DICOM files, full RTPLAN/RTSTRUCT/RTDOSE, paired pre- and post-RT PET-CT, L3 skeletal-muscle/adipose body-composition measures) is the first publicly shared HNSCC RT archive that supports body composition as either a risk factor or endpoint PMID:30179230.
  • The NSCLC Radiogenomics Stanford dataset pairs preoperative CT/FDG-PET in 211 surgical NSCLC patients with EGFR/KRAS/ALK clinical calls and RNA-seq (n=130) or Illumina HT-12 microarrays (n=26), explicitly designed to support CT-to-molecular biomarker discovery PMID:30325352.
  • The TCGA-GBM/LGG MRI collection on TCIA was released with expert-revised segmentation labels (enhancing tumor, non-enhancing tumor core, peritumoral edema) generated by GLISTRboost and >700 radiomic features per case (n=243), enabling glioma radiogenomics against matched gbm_tcga and lgg_tcga cBioPortal studies PMID:28872634.
  • In high-grade serous ovarian cancer (n=444), a multimodal late-fusion Cox model combining an omental-implant CT radiomic feature (wavelet-HLL gray-level-co-occurrence autocorrelation, log(HR)=1.68, test c=0.53) with an H&E tumor-nuclear-area feature (test c=0.54) and HRD status/PARP exposure reached test c=0.61–0.62, significantly beating HRD status alone (c=0.52) and every unimodal model; modality risk scores were orthogonal (|Kendall|<0.14) PMID:35764743.

Molecular correlates of radiation response and resistance

  • In pediatric neuroblastoma single-nuclei transcriptomes (n=11 tumors; validated in the 498-sample SEQC cohort), the high-risk-enriched undifferentiated tumor cluster nC3 significantly over-expresses BRCA1, BRCA2, MYCN, and ALK; gene-ontology analysis of nC3 genes correlated with poor survival was enriched for DNA damage / repair and cell cycle pathways PMID:34493726.
  • In a prostate-cancer PDX series (n=44 PDXs, 38 patients; cross-referenced with SU2C), unsupervised clustering of DDR-associated gene expression segregated neuroendocrine (NEPC) from adenocarcinoma PDXs concordant with morphology; NEPC samples showed DDR-pathway upregulation plus elevated MYCN/AURKA expression without gene amplification PMID:38488813.
  • Radiation-associated sarcomas (n=82 MSK-IMPACT-profiled cases across 4 histotypes) show histotype-specific genomic landscapes: RT-angiosarcoma is dominated by MYC amplification (75%, exclusively in breast/chest-wall cases, 87% of 38 breast cases) with co-occurring FLT4/CRKL/HRAS/KMT2D alterations; MYC-amplified RT-AS had significantly shorter RT-to-sarcoma latency than non-MYC-amplified cases (log-rank P=0.0083, medians 8.0 vs 12.5–18.5 years across other histotypes); predominant mutational signatures across all four RT-sarcoma histotypes were associated with errors in DNA repair and replication PMID:37350195.
  • In matched primary-recurrent glioma (GLASS consortium, n=132), IDH-mutant gliomas treated with temozolomide and/or radiotherapy undergo progressive genome-wide loss of DNA methylation at recurrence (620 differentially methylated probes Kruskal-Wallis FDR<0.01, |Δβ|>20%); 34% of GCIMP-high tumors progressed to GCIMP-low in the treatment group vs 4% untreated (Fisher P=0.005), with CDKN2A homozygous deletion associated with radiotherapy-related acquisition and worse survival from second surgery (log-rank P=0.0001) PMID:38117484.
  • In fallopian-tube HGSOC precursors, an IFN-related DNA-damage-resistance (IRDS) signature (STAT1, MX1, MCL1) emerges at the co-occurring STIC stage, and cGAS-STING is activated through BANF1+ ruptured micronuclei with gamma-H2AX co-localization as early as incidental STIC lesions — linking chronic IFN activation to a candidate treatment-resistance axis PMID:39386723.
  • The ROBIN consortium reports preclinical evidence that dietary interventions modulate radiosensitivity: low-fat and calorie-restricted regimens enhanced radiation response, whereas ketogenic diets unexpectedly promoted radio-resistance (OligoMET Project 2) PMID:41941260.
  • ROBIN-GenRad preliminary data implicate intratumoral bacteria in the Javelin HN100 chemoradiation-immunotherapy expansion cohort as a major source of therapy resistance in HNSCC PMID:41941260.

Radiosensitization (preclinical mechanism)

  • Conjugating potent anti-tubulin warheads (auristatins, maytansinoids) to anti-ErbB antibodies (cetuximab→EGFR, trastuzumab→ERBB2) restricts radiosensitization to receptor-expressing tumors: free MMAE indiscriminately radiosensitized both EGFR+ CAL-27 and EGFR− LN229, but C-MMAE radiosensitized CAL-27 and not LN229; T-DM1 IC50 was <1 nM in HER2+ OE19/NCI-N87 vs >100 nM in HER2− HCT116/CAL-27; a single 0.25 nmol T-DM1 dose + 2.5 Gy × 3 produced tumor doubling-time increases from 7→66 days (OE19) and 9→113 days (NCI-N87) with no benefit in HER2− HCT116 PMID:27698471.
  • Mechanistically, ADC radiosensitization tracked G2/M accumulation (pS10 Histone H3) and DNA double-strand breaks (neutral comet assay tail length), and was abolished when the antibody-receptor pairing was mismatched, implicating receptor-mediated endocytosis and intracellular release of the anti-tubulin warhead as the gating step PMID:27698471.

Clinical biomarker and combination trials

  • A single-arm phase 2 trial of IMRT reirradiation + concurrent and maintenance nivolumab in 51 recurrent/second-primary HNSCC patients met its primary endpoint (1-year PFS 61.7%, log-rank P=0.002 vs MIRI historical 43.8%), with manageable toxicity (grade ≥3 TRAEs 11.8%, no grade 5); any-grade radiation dermatitis occurred in 58.8% and oral mucositis in 49.0% PMID:38780927.
  • In that trial, PD-L1 combined positive score (22C3, <20 vs ≥20) did not stratify PFS (68.8% vs 59.2% at 1 year, P=0.86) or OS (P=0.74) — divergent from PD-L1’s role in the recurrent/metastatic HNSCC setting with single-agent PD-1 blockade PMID:38780927.
  • Also in that trial, an early ≥1.5-fold peripheral-blood surge of CD4+PD-1+Ki-67+ T cells trended toward worse PFS (HR 2.09, 95% CI 0.77–5.66; median PFS 16.0 vs 27.1 months) — opposite in direction to what has been reported in lung cancer and melanoma under PD-1 blockade alone PMID:38780927.
  • The ROBIN U54 consortium is running five disease-focused Molecular Characterization Trials (OligoMET/prostate, ImmunoRad/rectal, GenRad/bladder+HNSCC, METEOR/cervix+ pancreas, KIDSROBIN/DMG+neuroblastoma) with longitudinal pre-/on-/post-RT biospecimens and imaging, and has already produced preliminary signals including RT-induced MYC-driven CD8+ T-cell activation in rectal-cancer PBMCs, p53/MDM2 upregulation (without PD-L1/PD-1 axis induction) in cervical tumors after chemoradiation, and RT-altered cDC1 infiltration in pancreatic tumors PMID:41941260.

Normal-tissue toxicity and late genomic effects

  • The MIRT + nivolumab HNSCC reirradiation trial documented any-grade radiation dermatitis 58.8%, dysphagia 54.9%, oral mucositis 49.0%, dry mouth 47.1%, and fatigue 82.4%, with only 6 (11.8%) grade ≥3 TRAEs and no grade 5 events — establishing tolerability of PD-1 blockade in a previously irradiated field PMID:38780927.
  • Radiation-associated sarcoma latency varies by histotype (median RT→sarcoma interval: RT-AS 8.0 y, RT-MPNST 12.5 y, RT-UPS 18.5 y, RT-OS 9.0 y; global log-rank P=0.0047), and the radiation-specific angiosarcoma genome (MYC amplification, low FGA, translocation-free) differs markedly from other radiation-associated sarcomas and from sporadic angiosarcoma PMID:37350195.

Where papers agree

  • Tumor volume is a hard-to-beat baseline. Aerts et al. explicitly report that tumor volume alone performed well and that the four-feature radiomic signature combined with volume beat volume alone in every NSCLC/HNSCC validation cohort PMID:24892406; Kazmierski et al. find the best 12-model HNSCC submission against RADCURE was an EMR + primary-tumor-volume MTLR model that no radiomics-only submission outperformed, and that on the MDACC external subset only the top model beat tumor volume alone and by a small margin PMID:37397861. Both literatures converge on volume as a strong anchor in thoracic and head-and-neck imaging.
  • Intratumor heterogeneity on imaging has biological correlates. Aerts et al. link the two intratumor-heterogeneity features in their signature to cell-cycling pathways by GSEA on Lung3 PMID:24892406; Boehm et al. identify an autocorrelation feature of omental implants on wavelet-transformed CT as an independent prognostic factor in HGSOC, interpretable as lesion-density / intratumoral heterogeneity PMID:35764743; Bakas et al. release >700 radiomic features on TCGA-GBM/LGG MRI precisely to support heterogeneity-to-molecular mapping PMID:28872634.
  • Public, curated RT-planning datasets are the substrate. Welch et al. (RADCURE, n=3,346), Grossberg et al. (TCIA HNSCC, n=215), Bakas et al. (TCGA-GBM/LGG MRI, n=243), and Bakr et al. (NSCLC Radiogenomics Stanford, n=211) all explicitly frame their contribution as infrastructure for radiomics, radiogenomics, and RT-response modeling PMID:38362943 PMID:30179230 PMID:28872634 PMID:30325352.
  • DNA-damage-response biology drives radiation-relevant phenotypes. The high-risk neuroblastoma undifferentiated cluster nC3 up-regulates BRCA1/BRCA2 and is enriched for DDR/cell-cycle GO categories among poor-survival genes PMID:34493726; NEPC PDXs separate from adenocarcinoma PDXs on DDR-gene-expression clustering PMID:38488813; RT-associated sarcomas share mutational signatures of defective DNA mismatch repair and replication errors across histotypes PMID:37350195; an IRDS signature (STAT1, MX1, MCL1) emerges early in HGSOC precursors via cGAS-STING activation on BANF1+ ruptured micronuclei PMID:39386723.
  • Treatment with radiotherapy leaves measurable longitudinal molecular scars. Treated IDH-mutant gliomas show genome-wide DNA hypomethylation, GCIMP-high→low transitions (10/29 vs 1/27 untreated, P=0.005), CDKN2A loss, and increased CD31+ endothelial / CD8+ T-cell infiltration at recurrence PMID:38117484; RT-associated sarcomas carry histotype- specific driver landscapes distinct from sporadic counterparts PMID:37350195.

Where papers disagree

  • Does a single CT radiomic signature generalize across institutions in head-and-neck cancer? Aerts et al. report that a fixed four-feature signature trained on 422 NSCLC patients validated without retraining in HNSCC H&N1 (CI=0.69, P=7.99×10⁻⁷) and H&N2 (CI=0.69, P=3.53×10⁻⁶), and retained prognostic performance in HPV-negative HNSCC (CI=0.66) — framed as cross-disease transferability of a “general prognostic tumour phenotype” PMID:24892406. Kazmierski et al., working on a larger and more recent HNSCC benchmark (RADCURE n=2,552 + 873 external), report that performance of top models dropped in 2 of 3 external cohorts (HN1, MDACC, GPCCHN), that on the MDACC subset only the top model beat tumor volume alone and by a small margin, and that distribution shifts in HPV status, disease site, and outcome prevalence were statistically significant — concluding that any such model requires population-specific validation before deployment and that deep radiomics added to volume does not improve 2-year OS over EMR+volume PMID:37397861. The papers disagree on the strength of the generalizability claim for CT-derived HNSCC radiomic prognosis; Kazmierski explicitly frames this as echoing earlier negative findings by Ger and Vallières.
  • Does volume-independent image information add prognostic value beyond volume+EMR? Aerts et al. state that combining the radiomic signature with tumour volume outperformed volume alone in every NSCLC/HNSCC validation cohort, and that signature + TNM beat TNM alone in all three validation cohorts PMID:24892406. Kazmierski et al. report, in the RADCURE challenge, that (i) replacing tumour volume in the winning MTLR model with deep image representations from a 3D convnet dropped AUROC from 0.823 to 0.766,
    1. no radiomics-only model outperformed any EMR-only model, and (iii) the best radiomics model learned features only weakly correlated with volume (ρ=0.22) but still did not improve over the EMR+volume baseline PMID:37397861. Boehm et al. land in between: omental-implant autocorrelation radiomics did add prognostic value in HGSOC independent of HRD and clinical features, but the individual radiomic Cox submodel test-set c-index was only 0.53 PMID:35764743.
  • Does PD-L1 (CD274) combined positive score predict benefit of PD-1 blockade when combined with radiotherapy? In the MIRT + nivolumab trial, PD-L1 CPS <20 vs ≥20 showed no correlation with PFS (P=0.86) or OS (P=0.74) in previously irradiated HNSCC fields, contradicting the recurrent/metastatic HNSCC setting where PD-L1 CPS stratifies benefit of single-agent PD-1 blockade PMID:38780927. The corpus does not contain a primary RT + PD-1 HNSCC trial that shows a positive PD-L1 CPS effect, but the authors frame their result explicitly as divergent from the broader PD-L1 literature — leaving open whether the discordance reflects RT-induced PD-L1 changes, a previously-irradiated-field effect, or small-sample artifact (n=51).
  • Does an early PD-1+Ki-67+ CD4+ T-cell surge in peripheral blood mark benefit or harm under PD-1 blockade? Saba et al. report a trend toward worse PFS (HR 2.09, 95% CI 0.77–5.66; median 16.0 vs 27.1 months) in HNSCC reirradiation + nivolumab patients with a ≥1.5-fold surge at week 2–4, explicitly noting this is opposite in direction to reports in lung cancer and melanoma PMID:38780927. The corpus does not contain the lung/melanoma references as primary papers, but the paper records the disagreement against that external literature; the unresolved question is whether radiation reshapes the interpretation of peripheral PD-1+Ki-67+ CD4 dynamics, or whether the HNSCC result is a small-sample HR artifact.
  • Do ketogenic dietary interventions enhance or undermine radiation response? The ROBIN OligoMET preclinical arm reports that low-fat and calorie-restricted regimens enhanced radiation response whereas ketogenic diets unexpectedly promoted radio-resistance — a finding the authors flag as contrary to prior expectation that ketogenic diets would potentiate RT PMID:41941260. The corpus does not contain a counter-paper showing ketogenic-diet-induced radiosensitization, but the ROBIN white paper itself records the disagreement against the prior preclinical literature; the direction of the true effect remains open.

Open questions

  • Which imaging-biomarker-plus-molecular combinations actually add prognostic information beyond tumor volume, EMR (stage, HPV, performance status), and HRD/driver-gene status in curative-intent RT cohorts? The corpus shows both strong positive (Aerts signature across NSCLC/HNSCC) and negative (RADCURE deep radiomics, MDACC external validation) results and does not resolve the question. PMID:24892406 PMID:37397861 PMID:35764743
  • What is the mechanism of the observed peripheral PD-1+Ki-67+ CD4+ T-cell surge tracking worse PFS specifically in previously-irradiated HNSCC fields? Candidates enumerated by the authors — regulatory skew, exhaustion phenotype, lymphatic drainage effects from prior radiation — remain uninterrogated in the corpus. PMID:38780927
  • Is the RT-induced intratumoral-microbiome shift (HNSCC) causal for chemoradiation- immunotherapy resistance, or a correlate of an already-resistant TME? The ROBIN consortium flags this as a major forward-looking mechanistic question. PMID:41941260
  • Do the DDR-gene-expression signatures observed in high-risk neuroblastoma nC3 and in NEPC PDXs translate to differential radiosensitivity in patients? The corpus documents the signatures but does not test radiation-response endpoints. PMID:34493726 PMID:38488813
  • Does treatment-induced IDH-mutant glioma hypomethylation confer actionable therapeutic vulnerability at recurrence, or merely mark worse prognosis? The corpus demonstrates the epigenetic drift and worse post-second-surgery survival (log-rank P=0.0001) but does not provide a radiation-independent intervention that reverses it. PMID:38117484
  • Why is MYC amplification exclusively observed in breast/chest-wall radiation-associated angiosarcoma (87% of 38 breast cases) and not in radiation-associated sarcoma from other anatomic sites? The corpus records the association but not the mechanism. PMID:37350195
  • Do ketogenic diets enhance or attenuate radiation response in humans? The ROBIN preclinical signal disagrees with the prior preclinical literature; no human data are in the corpus. PMID:41941260
  • Can ErbB-directed antibody-drug-conjugate radiosensitization (T-DM1 + IR, C-MMAE + IR) be translated to a therapeutic index advantage over free anti-tubulin + IR in patient cohorts? No human trial data are in the corpus; HER2/EGFR expression cut-offs predictive of ADC radiosensitization in human tumors remain undefined. PMID:27698471
  • What is the optimal sequencing, duration, and fractionation of PD-1 blockade with reirradiation? The MIRT + nivolumab trial met its primary PFS endpoint but cannot attribute the benefit to radiation vs PD-1 vs combination in a single-arm design. PMID:38780927

This page was processed by crosslinker on 2026-05-04.