Concurrent driver mutations induce distinct tumour morphologies on radiological imaging

Most radiogenomic studies examine single driver mutations in isolation, despite evidence that co-occurring driver alterations are common and functionally interacting in cancer. Whether concurrent mutations produce distinct imaging phenotypes or simply average single-mutation effects remains unknown.

A multi-cancer cohort of 1235 patients with 8633 segmented lesions from contrast-enhanced CT scans and matched genomic profiling was retrospectively assembled. Patients with no driver mutations, TP53 + other mutations, TP53-only, EGFR-only, KRAS-only, and strictly TP53 + EGFR or TP53 + KRAS mutations were compared. Primary analyses were conducted in the original feature space at the patient level (via median aggregation). Separation between mutation groups was quantified using centroid distance and cross-group inter-patient distance with 1000 patient-aware label-stratified bootstraps. Lesion-level dimensionality reduction was used for descriptive visualisation and mixed-effects models adjusted for tumour site. A cross-validated parental-axis geometry analysis tested emergence beyond additivity.

Single-mutant cohorts formed distinct phenotypes: EGFR-only versus KRAS-only centroid distance 18.4 (95% CI = 12.0-23.8); TP53-only versus EGFR-only 16.9 (12.3-21.3). Co-mutated tumours were separate from parents in the patient-level and lesion-level space. Patient-level centroid distances were TP53-only versus TP53 + EGFR = 13.4 (10.1-16.9); EGFR-only versus TP53 + EGFR = 7.7 (5.3-11.6); TP53-only versus TP53 + KRAS = 6.4 (4.5-9.1). Co-mutated lesions were morphologically closer to TP53-only than to the EGFR-only or KRAS-only cohort (Mahalanobis p < 0.001). Emergence beyond additivity was supported for TP53 + EGFR (off-axis p = 0.009) and trended for TP53 + KRAS (p = 0.062).

Across different cancers, CT radiomics identified genotype-specific phenotypes and showed that co-mutations produce distinct morphologies. Co-mutation context should be modelled explicitly in radiogenomics and prospectively validated for non-invasive molecular stratification.