KEY POINTS
- The analysis included 3,269 TCGA patients recorded as having received radiotherapy and with available gene-expression and outcome data. Eight cancer types had more than 100 clinically complete cases: breast (n = 553), cervical (n = 183), head and neck (n = 320), low-grade glioma (n = 316), lung adenocarcinoma (n = 104), melanoma (n = 125), thyroid (n = 326), and endometrial cancer (n = 257).
- Investigators calculated 100 genomic parameters from scores and dispersions for 50 MSigDB hallmark gene sets. For each cancer, the prediction model was trained using all other cancer types, meaning that the evaluated disease was completely excluded from model development.
- After adjustment for age, stage, and sex, each standard-deviation increase in the pan-cancer score remained associated with progression in breast cancer (HR 1.37, 95% CI 1.09–1.73; p = 0.007), cervical cancer (HR 2.05, 1.53–2.75; p < 0.001), head and neck cancer (HR 1.29, 1.08–1.54; p = 0.005), low-grade glioma (HR 1.49, 1.27–1.76; p < 0.001), and lung adenocarcinoma (HR 1.47, 1.17–1.85; p = 0.001).
- Associations were not significant for melanoma (HR 1.08, 95% CI 0.88–1.31), thyroid cancer (HR 1.04, 0.76–1.42), or endometrial cancer (HR 1.08, 0.86–1.37), showing that information shared across cancers did not improve prediction uniformly.
- The largest performance gain occurred in cervical cancer: the optimism-corrected C-statistic increased from 0.53 with clinical variables alone to 0.69 with the genomic score added. Corresponding changes were smaller for breast cancer (0.66 to 0.68), low-grade glioma (0.60 to 0.64), lung adenocarcinoma (0.54 to 0.60), and head and neck cancer (0.52 to 0.56).
- Models developed separately within each cancer using all 100 genomic parameters performed poorly in most diseases, whereas the pan-cancer associations largely persisted after adjustment for reduced-dimension cancer-specific predictions. This supports cross-cancer modeling as a possible response to inadequate event numbers in genomic datasets.
- The endpoint was progression-free interval, which combined progression, recurrence, metastasis, new primary tumors, and death with disease. Radiotherapy dose, fractionation, indication, technique, and systemic treatment were unavailable or incomplete, so the score should be interpreted as prognostic after radiotherapy rather than a validated predictor of radiotherapy benefit.
CLINICAL TAKEAWAY
Shared gene-expression information from other cancers may strengthen prognostic modeling when disease-specific datasets are too small, particularly in cervical cancer, low-grade glioma, head and neck cancer, and lung adenocarcinoma. The model is not ready for clinical use: calibration was limited, radiotherapy information was sparse, and independent validation is required.