S2596
Radiobiology - Preclinical biomarkers
ESTRO 2026
Conclusion: We report the development of a ligand–receptor index quantifying tumour–muscle signalling, demonstrating prognostic association in RNAseq datasets from pancreatic and liver cancer and showing disease specificity. External microarray datasets did not reproduce these findings, suggesting possible measurement-platform effects. Further evaluation using platform-consistent cohorts is required to define the model’s wider applicability. Keywords: cachexia biomarker, tumour–muscle signalling
receptor availability, and quantified its prognostic performance and external reproducibility as well as disease specificity across gastrointestinal cancer types. Material/Methods: We constructed a two-compartment model using tumour ligand expression from TCGA RNAseq data and skeletal muscle receptor expression from GTEx v8. A focused panel of up to 25 cachexia- related ligands underwent LASSO-Cox selection to derive prognostic coefficients ( λ ). Receptor weights were generated using a Hill- equation transformation with the GTEx median receptor expression as the fixed . The composite index was defined as: Multivariable Cox models adjusted for age, sex, stage, and NR3C1 (Glucocorticoid Receptor) expression, which was included as an adjusted covariate to control for systemic catabolic confounding and ensure the zCCI score reflects the specific, tumor-driven signaling axis rather than general hormonal-induced wasting. Predefined QC excluded cohorts with N <60 patients. Specificity was assessed in pancreatic adenocarcinoma (PAAD), liver hepatocellular carcinoma (LIHC), stomach adenocarcinoma (STAD), esophageal carcinoma (ESCA), and cholangiocarcinoma (CHOL). External validation used microarray datasets (GSE62452 for PAAD; GSE76427 for LIHC) with z-score normalisation. Results: The zCCI demonstrated prognostic association in RNAseq cohorts: pancreatic adenocarcinoma (PAAD), N=176: 9-ligand signature; HR=1.78, p<0.0001; C-index=0.676. Hepatocellular carcinoma (LIHC), N=365: 2-ligand signature (CXCL8, CCL2); HR=1.42, p=0.0011; C- index=0.660. Specificity analysis showed no selected ligands in STAD, a statistically unstable solution in ESCA (lacking sufficient degrees of freedom). External validation did not replicate the RNA- seq findings. The PAAD microarray cohort (N=66) showed a non- significant trend (p=0.068), and the LIHC cohort (N=115) showed no significant association (p=0.625) with wide, unstable confidence intervals.
Digital Poster 4750
Dynamic optical coherence tomography radiomics for high- throughput, non-invasive assessment of radiation damage in spheroids Mark Daniel Arndt 1 , Rico Hansler 2 , Luca Tirinato 3 , Andrey Tkachenko 1 , Joao Seco 4,5 , Ute Schepers 2 , Maria Francesca Spadea 1 1 IBT, KIT, Karlsruhe, Germany. 2 IFG, KIT, Karlsruhe, Germany. 3 Dipartimento di Scienze Mediche e Chirurgiche, Magna Graecia University, Catanzaro, Italy. 4 Biomedical Physics in Radiation Oncology, DKFZ, Heidelberg, Germany. 5 Department of Physics and Astronomy, Universität Heidelberg, Heidelberg, Germany Purpose/Objective: We developed and validated a fast, label free, non destructive method to quantify radiation induced damage in multicellular spheroids. We hypothesized that dynamic optical coherence tomography (dOCT), which provides rapid 3D tomographic images together with time variance maps related to intracellular activity [1], combined with radiomics [2], would predict radiation damage more accurately than brightfield imaging size metrics or destructive live dead staining. A more accurate assessment of radiation damage can enhance experimental power, enable robust longitudinal studies, and ultimately inform better radiobiology workflows. Material/Methods: Spheroids were irradiated and then imaged using brightfield imaging and dOCT on several different days following the irradiation, yielding volumetric intensity data and time variance images acquired in less than 2 minutes per spheroid. From these data, radiomic features were extracted, spanning morphology, first order intensity, and texture and non local descriptors sensitive to spatial heterogeneity and dynamic metabolic activity. Feature subsets were selected and used as inputs to classical machine learning models to predict radiation damage, and performance was compared against a brightfield baseline that used spheroid size as the predictor. Model development and evaluation followed a standardized pipeline with checks for feature intercorrelation, redundancy reduction, and repeated held out validation to assess robustness. Statistical comparisons tested whether dOCT radiomics outperformed the brightfield baseline. Results: For different models, such as support vector machines and random forests, and for different numbers of features, dOCT-derived radiomic features achieved significantly higher predictive performance than brightfield size measures when estimating radiation damage (p < 0.05). For the support vector machine this can be seen in the attached figure. The joint use of 3D tomographic images and time variance information enabled accurate discrimination of damage levels while preserving the ability for longitudinal follow up, since no staining or destructive preparation was required. Acquisition times supported high throughput workflows.
Made with FlippingBook - Share PDF online