Radiogenomics Overview

• Radiogenomics is an integrative field that fuses imaging-derived features with genomic data to noninvasively characterize disease and predict clinical outcomes.
• It employs mechanistic models and machine learning to link radiomics with genetic markers, enhancing personalized therapy stratification.
• The discipline has practical applications in oncology and neurology, supporting virtual biopsies, improved radiosensitivity assessments, and precision medicine.

Radiogenomics is the integrative field at the interface of biomedical imaging, genomics, and clinical outcome modeling. It systematically links imaging-derived quantitative phenotypes—radiomics—with germline and somatic genomic information to identify molecular correlates of imaging features, predict outcome, and enable both noninvasive biomarker discovery and individualized therapy stratification. Radiogenomics has matured over the past decade into a highly multidisciplinary endeavor encompassing mechanistic radiobiology, feature engineering, biostatistics, Bayesian inference, and deep multimodal representation learning. Research in this domain spans applications in oncology (e.g., glioblastoma, NSCLC, ccRCC, hepatocellular carcinoma), neurology (e.g., Alzheimer’s disease), and extends toward precision medicine, particularly in radiotherapy adaptation and immunotherapy response prediction.

1. Core Objectives and Scope

Radiogenomics aims to characterise and predict biological, prognostic, or therapeutic phenotypes by fusing imaging and genomic/layered -omics data. The field encompasses two principal axes (Kang et al., 2019):

• Normal-tissue radiosensitivity: Identification of patients at high risk of radiation-induced toxicity through germline polymorphisms (e.g., SNPs, CNVs) mapped to normal-tissue complication probability (NTCP).
• Tumor radiosensitivity and biology: Identification of tumors deriving maximal benefit from local therapy and stratification for chemo/radiotherapy via gene expression, miRNA, and mutational panels, linking to outcomes such as tumor control probability (TCP), survival, or recurrence.

By embedding genomic features into classical dose–response models or associating imaging features directly with molecular subtypes, radiogenomics enables individualized risk modeling, noninvasive genotyping (“virtual biopsy”), and the development of both prognostic and predictive biomarkers (Feng et al., 15 Oct 2025, Ismail et al., 2020, Wijethilake et al., 2020).

2. Mechanistic, Statistical, and Machine Learning Frameworks

Radiogenomic methodologies are structured around two main model classes: mechanistic augmentation of established radiobiological formalisms and high-dimensional statistical learning (supervised, unsupervised, or Bayesian).

Mechanistic Integration: Radiogenomics augments dose–response models such as the linear–quadratic (LQ) model:

$S(D) = \exp\bigl(-\alpha D - \beta D^2\bigr)$

by modulating cell-kill parameters ($\alpha$, $\beta$) or defining genome-adjusted radiation dose (GARD). For instance, the radiosensitivity index (RSI) is a 10-gene expression signature used in

$$\text{GARD} = D \, (\alpha + \beta D)\, \text{RSI}^{-1}$$

to personalize dosing (Kang et al., 2019). NTCP models incorporate dose-modifying factors (DMFs) tied to the presence of allelic risk variants.

High-Dimensional Machine Learning: Large-scale feature matrices (imaging: radiomics; genetics: microarrays, RNA-seq) are analyzed via penalized regression (lasso, elastic net), kernel SVMs, tree ensembles, and deep NNs. Feature selection is critical given the $p \gg n$ regime, employing filters (univariate tests), wrappers (e.g., genetic/search algorithms), embedded regularization, and manifold approaches (PCA, t-SNE). Performance is reported using cross-validated ROC AUC, c-index (for time-to-event), calibration, and external validation (Shiri et al., 2019, Aonpong et al., 2021, Navarrete et al., 2022, Zeng et al., 2022).

Bayesian and post-selection frameworks explicitly address the inference after variable selection, correcting bias from $\ell_1$ screening using selection-aware posteriors or hierarchical spike-and-slab/group-slab priors (Panigrahi et al., 2020, Mohammed et al., 2021, Mohammed et al., 2021). Multitask and joint models couple imaging–genomic and clinical–genomic regressions to exploit cross-domain signal and improve selection stability (Zeng et al., 2022).

3. Radiomics Feature Engineering and Predictive Biomarker Development

Radiomics refers to the extraction of quantitative descriptors (intensity, shape, texture, wavelet, fractal, contextual) from standard clinical images (MRI, CT, PET). Imaging phenotypes are engineered via:

• First-order features: Histogram mean, variance, skewness, kurtosis, entropy
• Texture features: GLCM (e.g., contrast, correlation, homogeneity), GLRLM, GLSZM, GLDM, NGTDM
• Shape & geometric measures: Volume, sphericity, bounding-box axes, fractal dimension (notably effective for heterogeneity quantification (Wijethilake et al., 2020, Feng et al., 15 Oct 2025))
• Context-aware features: Population atlases for spatial priors; features from co-localized biopsy samples for “virtual biopsy” mapping (Ismail et al., 2020)
• Deep learning features: CNN-based high-dimensional image embeddings, possibly fused with handcrafted features (Navarrete et al., 2022, Aonpong et al., 2021)

Stability analysis of candidate radiomics (e.g., via Overall Concordance Correlation Coefficient, OCCC ≥ 0.95) is increasingly deployed to mitigate sensitivity to segmentation variability and ensure reproducibility, especially in resource-constrained sites (Nadeem et al., 5 Jun 2024).

Validated biomarker pipelines—e.g., genotype-guided radiomics (GGR) for NSCLC recurrence (Aonpong et al., 2021), multiplexed signatures for EGFR/KRAS in CT/PET (Shiri et al., 2019), MCMO models for multi-gene mutation prediction in ccRCC (Chen et al., 2018)—demonstrate significant gains in predictive accuracy and clinical potential.

4. Multimodal and Graph-Based Fusion: Integration of Imaging and Omics

Recent work emphasizes the necessity and complexity of multimodal fusion architectures for radiogenomics. Linear fusion (e.g., elastic-net Cox models) may offer robustness with small cohort sizes, as demonstrated in lung cancer recurrence (Subramanian et al., 2020), while nonlinear and neural architectures (MLPs, attention modules) often require larger datasets (Oghenekaro, 30 Nov 2025). Graph representation learning—e.g., heterogeneous bipartite GNNs linking subject-wise imaging and gene nodes—enables more flexible, biologically plausible mapping of cross-domain dependencies, supporting disease classification and gene-importance ranking (Raj et al., 14 May 2025).

Cross-modal learning frameworks for mutation prediction in HCC exploit multiphase CT, concatenated ResNet-derived vectors, and auxiliary clinical biomarkers to achieve robust predictions and handle intra-tumoral heterogeneity by sector-specific gene labeling (Gu et al., 2020). These architectures generalize to multipartite graphs capturing imaging, genomics, proteomics, and metabolomics (Raj et al., 14 May 2025).

In brain tumor radiogenomics, alignment of feature extraction with biological structure (e.g., spherical radiomics on concentric tumor shells (Feng et al., 15 Oct 2025), layered Fisher–Rao density PCA (Mohammed et al., 2021)) enhances the detection of spatially-driven molecular heterogeneity.

5. Clinical and Translational Applications

Key translational advances include:

• Noninvasive genotyping ("virtual biopsy"): Voxel-wise probability maps of mutation status (e.g., SpACe maps for EGFR/MGMT in GBM) support precision neurosurgical sampling and reflect spatial genomic heterogeneity (Ismail et al., 2020).
• Therapy stratification and predictive modeling: Genome-informed radiotherapy protocols (e.g., GARD, RSI) adjust therapeutic intensity to optimize the TCP/NTCP balance (Kang et al., 2019).
• Immunotherapy radiogenomics: MRI radiomic signatures of diffusion, perfusion, and texture are linked to immune infiltration genes and response, establishing candidates for immunotherapy stratification though rigorous prospective validation remains lacking (Ghimire et al., 13 May 2024).
• Phenotype-to-genotype mapping: Imaging-derived features act as surrogates for gene expression, enabling actionable insight in settings where tissue sampling is infeasible or incomplete (Aonpong et al., 2021, Navarrete et al., 2022).

Tables below summarize selected validated pipelines and their clinical-molecular endpoints:

Disease/Setting	Imaging Modality	Molecular Endpoint	Framework	AUC/Accuracy	Reference
NSCLC recurrence	CT	RNA-seq genes	GGR (hybrid NN)	0.7667/83%	(Aonpong et al., 2021)
GBM driver mutations	MRI	EGFR, MGMT	SpACe LASSO+PPMM	~0.90	(Ismail et al., 2020)
ccRCC mutations	CT	VHL, PBRM1, BAP1	MCMO+ER+SMO	≥0.86	(Chen et al., 2018)
HCC	multiphase CT	APOB, COL11A1, ATRX	4-phase CNN+biomrk	~0.75	(Gu et al., 2020)
AD/MCI	MRI	APOE, PSEN1/2 genes	Bipartite GNN	Noted high	(Raj et al., 14 May 2025)

6. Methodological Limitations, Reproducibility, and Future Directions

Current limitations in radiogenomic research include small sample sizes relative to feature dimension, variable imaging protocols, segmentation inconsistencies, and missing modalities—necessitating harmonization and external multi-site validation (Ghimire et al., 13 May 2024, Nadeem et al., 5 Jun 2024, Wijethilake et al., 2020). Bias assessments (e.g., QUADAS-2, CLAIM) identify high patient-selection and model-overfitting risk. Model interpretability, code transparency, and data-sharing remain insufficient (Ghimire et al., 13 May 2024).

Future directions are centered on:

• Robust, harmonized pipelines: Explicit stability filtering, federated learning of feature robustness, and containerized workflows.
• Integration of multi-omics (proteomics, metabolomics, spatial transcriptomics): Multipartite graph and transformer architectures to handle further heterogeneity and missingness (Oghenekaro, 30 Nov 2025, Raj et al., 14 May 2025).
• Prospective and multi-cohort validation: Large-scale, registered studies embedding radiogenomic arms into randomized clinical trials.
• Explainability and biological validation: SHAP, clustering/transition modeling, and explicit mapping of radiomic transitions to molecular events (Feng et al., 15 Oct 2025, Oghenekaro, 30 Nov 2025).
• Radial and spatial biology-aware feature engineering: Spherical radiomics, concentric layer models, and geometry-aware density descriptors yield features better aligned with biological processes (Feng et al., 15 Oct 2025, Mohammed et al., 2021).

Radiogenomics will increasingly function as an interpretive bridge between quantitative imaging, spatial biology, and therapeutic response, supporting a precision-medicine paradigm in effective, individualized clinical decision-making.