The finding matters because KRAS, once considered undruggable, is now one of the most consequential genes in lung oncology. Mutations in this gene define a distinct biological subtype of non-small cell lung cancer with specific treatment implications, and knowing a tumour’s KRAS status before tissue sampling or molecular testing could, in principle, sharpen clinical decision-making. Molecular profiling currently requires biopsy material or liquid biopsy assays, which take time and are not always available. If a non-invasive imaging signature could reliably hint at mutation status, it could help prioritise patients for targeted testing or clinical trial enrolment. The new study tests that idea with unusual statistical rigour for an exploratory radiomics analysis, and its results are a masterclass in why that rigour matters.

The team assembled a retrospective, single-centre cohort of 271 candidate records that had been intentionally enriched for KRAS-positive tumours. After a final eligibility review excluded 71 patients, 200 patients remained: 101 with KRAS-mutant lung adenocarcinomas and 99 with KRAS-wildtype tumours. The cohort was split 140 for training and 60 for internal testing, and the researchers deliberately preserved that original partition without re-partitioning, a choice that guards against the subtle optimism that creeps in when datasets are repeatedly reshuffled until a model looks good. The near-even balance of mutant and wildtype tumours was by design, since a KRAS-enriched cohort gives far more statistical power to detect mutation-associated imaging features than the roughly one-in-four prevalence seen in unselected populations.

The core of the analysis is a quantity the authors call the ITHscore, computed in both two-dimensional and three-dimensional versions from CT images. Radiomics approaches like this one convert the grayscale texture of a tumour on a scan into quantitative descriptors: regions of viable tissue, necrosis, mucin, haemorrhage and varying cell density all attenuate X-rays differently, and spatial variation in those attenuation patterns can be summarised mathematically. The intuition is straightforward. A tumour whose cells are genetically diverse tends to be physically diverse too, and that physical diversity — uneven density, irregular internal structure, heterogeneous enhancement — is exactly what CT is built to detect. The researchers asked whether the spatial arrangement of these patterns, not just their average intensity, differed between KRAS-mutant and KRAS-wildtype tumours.

The headline association was striking. Tumour volume correlated strongly with both ITH scores, with Spearman correlation coefficients of 0.658 for the 2D score and 0.679 for the 3D score — meaning bigger tumours were substantially more heterogeneous on imaging, which is biologically plausible since larger lesions have had more time to evolve subclones. After adjusting for log2-transformed tumour volume, T category, age, sex and smoking history in logistic regression models, each standard deviation increase in the 2D ITHscore was associated with 3.84-fold higher odds of harbouring a KRAS mutation (95% confidence interval, 2.23 to 6.59), while the 3D score yielded an odds ratio of 3.45 (95% CI, 1.98 to 6.00). Those are large effects for an imaging biomarker, and they held in both 2D and 3D formulations.

But the authors then did something that many radiomics studies skip: they stress-tested the association against the possibility that it was simply a proxy for tumour burden. Because ITH scores rise with volume, and volume itself differs systematically between KRAS-mutant and wildtype tumours, any apparent independence could be an artefact of comparing big heterogeneous tumours with small homogeneous ones. When the analysis was restricted to the overlapping tumour-volume range — 102 patients whose tumours fell within sizes shared by both mutation groups — the odds ratios attenuated to 2.39 (95% CI, 1.14 to 5.00) for the 2D score and 2.28 (95% CI, 0.999 to 5.22) for the 3D score. The association weakened but did not vanish, leaving the question genuinely open.

A second sensitivity analysis used overlap weighting, a statistical technique that reweights patients so that the modelled covariates are balanced between mutation groups, achieving absolute standardised mean differences below 0.001 — essentially perfect balance on the measured variables. Under this weighting, the odds ratios were 3.94 (95% CI, 2.53 to 10.91) for the 2D score and 3.07 (95% CI, 1.96 to 6.57) for the 3D score, stronger than in the restricted analysis. The authors interpret this carefully: attenuation occurred in one analysis but not the other, so independence from measured tumour burden was neither confirmed nor excluded. They also note that the restricted cohorts were not matched, and residual imbalance in sex and smoking history persisted in that subset — a reminder that even careful subgroup analyses carry their own confounders.

The predictive question got an equally sober answer. The team fitted six logistic-regression reference models in the training cohort and compared them in the internal-test partition using paired bootstrap confidence intervals, which account for the fact that the same patients are being scored by competing models. A clinical-plus-volume model achieved an area under the receiver operating characteristic curve of 0.844 (95% CI, 0.726 to 0.941) — respectable discrimination. Adding both ITH scores actually nudged performance down slightly, to 0.833 (95% CI, 0.725 to 0.925), with a paired difference in AUC of −0.011 (95% CI, −0.099 to 0.077), a change indistinguishable from zero. In plain terms, the heterogeneity scores carried no incremental information about KRAS status once tumour size and basic clinical variables were known.

This pattern — a robust association that fails to translate into added predictive value — is increasingly recognised as a signature of confounding by tumour burden in radiomics research. Heterogeneous imaging features often simply flag larger, more advanced tumours, and any molecular association they show can dissolve once size and stage are properly accounted for. What makes this study notable is that the authors report that dissolution themselves rather than burying it. Their conclusions state plainly that the findings do not establish molecular specificity or clinical utility, and that validation in representative external cohorts is required. In a field where single-centre radiomics models are frequently oversold, that restraint is itself a kind of scientific signal.

The study, approved by the Ethics Committee of the First Affiliated Hospital of Nanchang University and funded in part by the National Natural Science Foundation of China, arrives at a moment when the radiomics community is reckoning with reproducibility. Its lessons are twofold. First, CT-derived spatial heterogeneity is genuinely and strongly linked to KRAS mutation biology in lung adenocarcinoma, an observation that may illuminate tumour evolution even if it does not yet serve as a diagnostic shortcut. Second, the path from association to clinical tool runs through exactly the kind of adversarial statistics this team deployed — volume restriction, overlap weighting, paired bootstrap comparison — and through external validation in cohorts that reflect real-world KRAS prevalence rather than deliberately enriched samples. Until that work is done, the chaotic texture visible on a lung CT remains a compelling biological clue, but not yet a substitute for the molecular test.

Subject of Research: Association between CT-derived spatial intratumoural heterogeneity and KRAS mutation status in lung adenocarcinoma

Article Title: CT-derived spatial intratumoural heterogeneity and KRAS mutation status in a predominantly lung adenocarcinoma cohort: an exploratory study of association and incremental discrimination

Article References: Wu, L., Liu, X., Zhou, Y., Rao, Z., Su, R., Xu, F., Zhang, L., Peng, C., Li, H., Li, L., & Peng, D. (2026). CT-derived spatial intratumoural heterogeneity and KRAS mutation status in a predominantly lung adenocarcinoma cohort: an exploratory study of association and incremental discrimination. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02962-4

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02962-4

Keywords: lung adenocarcinoma, KRAS mutation, intratumoural heterogeneity, radiomics, CT imaging, tumour burden, non-small cell lung cancer, biomarker, machine learning, cancer imaging, molecular testing, CT-derived

The conceptual foundation dates back to 1976, when Peter Nowell proposed that tumour cell populations evolve through acquired genetic lability, with stepwise selection of variant sublines driving progression. Half a century of bulk sequencing vindicated the model, revealing branched evolution, intratumoural heterogeneity and the selective sweeps that follow treatment. Yet bulk measurements average across millions of cells, obscuring which mutation resides in which cell and, crucially, what that mutation actually does to the cell carrying it. The review’s authors argue that closing this genotype-to-phenotype gap is now the central task of cancer evolutionary biology.

One striking motivation comes from healthy tissue. Landmark studies of normal skin, oesophagus, colon, endometrium, bladder and bronchial epithelium have shown that somatic mutations in canonical cancer driver genes are under pervasive positive selection in tissue that looks entirely normal under the microscope. In sun-exposed skin, roughly a quarter of cells carry cancer-associated mutations, and mutation burdens in some cells rival those of tumours. Clonal haematopoiesis, the age-related expansion of blood cells carrying mutations in genes such as DNMT3A, TET2 and JAK2, likewise demonstrates that genetic mosaicism is a feature of ordinary physiology and ageing, seeding the pre-malignant landscape from which frank cancers emerge.

Mapping this diversity, however, is only the first step. The review emphasises that understanding somatic clonal evolution requires defining the phenotypes that give mutated clones a fitness advantage, whether those phenotypes involve altered differentiation, survival, proliferation or interaction with the microenvironment. This is where genotype-aware single-cell multi-omics enters. Methods such as G&T-seq and its descendants physically split or barcode the genome and transcriptome of an individual cell, while genotyping-of-transcriptomes approaches recover expressed mutations directly from single-cell RNA-sequencing data. Targeted strategies enrich for known mutant loci, and chromatin-focused assays now co-capture mutations alongside single-cell accessibility profiles.

The biological payoffs have been substantial. In clonal haematopoiesis, single-cell multi-omics has shown that the effects of a mutation are often cell-state specific. DNMT3A R882 mutations, for example, were found to perturb early progenitor states through selective hypomethylation, a phenotype invisible to bulk assays. Splicing aberrations in haematopoietic clonal outgrowths display distinct cell-type-specific impacts, and maps linking genotypes to chromatin accessibility profiles reveal how individual mutations reshape regulatory landscapes in a lineage-dependent manner. In myeloproliferative neoplasms, clonally resolved analyses have traced how JAK2 and CALR mutations propagate through differentiation hierarchies, while work in acute myeloid leukaemia has connected RAS-mutant leukaemia stem cells to clinical resistance against the BCL-2 inhibitor venetoclax.

Beyond single time points, the review highlights the power of coupling phylogenetic reconstruction with phenotypic measurement. Endogenous marks such as somatic point mutations, copy-number alterations, mitochondrial DNA mutations, microsatellite shifts and stochastic epimutations each leave heritable traces that allow researchers to infer the ancestral relationships among single cells. Mitochondrial mutations in particular have enabled lineage tracing directly in human samples, and somatic epimutations have recently been used to chart the dynamics of blood ageing. Reconstructed single-cell phylogenies can then be time-calibrated, converting a branching diagram into a chronogram that estimates when a clone originated within a patient’s lifespan.

Such temporal mapping demands careful statistical treatment. The authors describe phylogenetic frameworks that quantify heritability and plasticity of cell states, decoupling genetic inheritance from non-genetic, environmentally driven transitions. Molecular clock models, whether strict or relaxed, permit inference of mutation rates and timing of clonal expansions, and phylodynamic approaches borrowed from pathogen genetics now illuminate how tumour population sizes fluctuate over the course of disease. Applied to colorectal cancer, these tools have revisited the Big Bang model of tumour growth, in which most subclonal diversity is generated in an early expansion rather than through later selective sweeps, and have documented phenotypic plasticity under genetic control during malignant progression.

Spatial context adds a further dimension. Multiclonal invasion patterns in breast tumours, spatially resolved copy-number maps in benign and malignant tissue, and spatial genomics of cancer clones all demonstrate that evolutionary dynamics are constrained by tumour architecture. Mechanical confinement has been shown to govern phenotypic plasticity in melanoma, and harsh microenvironments select for glycolytic phenotypes in early breast cancer. Integrating spatially resolved or lineage-resolved phenotypes with genotype maps is therefore revealing how selection operates not just on mutations but on the cell states and niches in which those mutations find themselves.

Translationally, the authors argue that genotype-to-phenotype mapping can expose therapeutic vulnerabilities for the precision elimination of disease-propagating mutant cells. Because mutant phenotypes are often confined to specific cell states or lineages, vulnerabilities may exist that spare wild-type tissue. Single-cell analyses have identified drug-tolerant persister states, non-genetic determinants of clonal fitness, and epigenetically inherited plasticity that drives drug resistance through one-to-many genotype-to-phenotype relationships. In glioblastoma, recurring cellular states whose abundance is modulated by genetic aberrations suggest combination strategies; in IDH-mutant oligodendroglioma, mutant IDH inhibitors induce lineage differentiation detectable at single-cell resolution. Evolutionary steering, in which treatment is designed to guide tumours toward collateral sensitivities, becomes more tractable when clonal phenotypes can be read directly in patients.

The review closes with a sober assessment of remaining challenges. Whole-genome amplification artefacts, allelic dropout and tissue dissociation biases still limit sensitivity and fidelity, and artifacts in mitochondrial DNA analyses can misinform phylogenetic inference if uncorrected. Computational methods continue to mature, from probabilistic single-cell phylogeny inference that accounts for sequencing error to models that relax the infinite-sites assumption when back mutations and parallel evolution occur. Yet the trajectory is clear: by pairing genotype with phenotype in the same cell and embedding those pairs within time-calibrated phylogenies, researchers can now define, directly in primary human samples, the mechanisms underlying clonal expansion in both healthy and malignant tissues. For a disease that evolves to evade every therapy thrown at it, that may prove the most consequential lens oncology has yet acquired.

Subject of Research: Single-cell multi-omics mapping of genotype and phenotype co-evolution in clonal evolution of healthy and cancerous tissues

Article Title: A single-cell lens into the co-evolution of genotypes and phenotypes in cancer

Article References: Izzo, F., Prieto, T., Potenski, C., & Landau, D. A. (2026). A single-cell lens into the co-evolution of genotypes and phenotypes in cancer. Nature Reviews Cancer. https://doi.org/10.1038/s41568-026-00970-8

Image Credits: AI Generated

DOI: 10.1038/s41568-026-00970-8

Keywords: single-cell multi-omics, clonal evolution, genotype-phenotype mapping, tumour heterogeneity, clonal haematopoiesis, single-cell phylogenetics, lineage tracing, somatic mosaicism, phenotypic plasticity, cancer genomics, epigenetics, therapeutic vulnerabilities