Tumor Prognosis Modeling: Survival, Immune Infiltration, and Signatures

TCGA survival analysis · CIBERSORT and ESTIMATE · Methylation prognostic markers · TMB · Signature construction

Published 2026 · Last updated 2026-09-27 · Hailu Zhou

Most published prognostic signatures do not generalize. The failure is rarely the statistics — it is the study design: training and testing on the same cohort, no external validation, and a cutoff chosen to maximize a p-value. A prognosis model is only as good as the independent cohort in which it survives. Everything below is written with that constraint in mind.

TCGA survival data: respect its structure

TCGA provides clinical follow-up with overall survival, disease-free survival, and progression-free interval. The standard model is Cox proportional hazards, but you must handle it properly. Check the proportional hazards assumption with Schoenfeld residuals; if it fails, stratify or switch to a time-varying or accelerated failure time model. Treat the progression-free interval from the Liu et al. pan-cancer reanalysis as the preferred endpoint, since it has cleaner curation than raw clinical fields. Censoring is informative in follow-up-limited cohorts, so always report median follow-up and the number of events, not just the number of patients.

Discrimination is measured with the concordance index and time-dependent AUC; calibration needs a curve, not a number. A model with c-index 0.65 that is well calibrated is more useful clinically than one with 0.75 that is not.

Immune infiltration: deconvolution at a blunt resolution

CIBERSORT infers cell-type fractions from bulk expression using nu-support vector regression on reference signature matrices1. Its successor CIBERSORTx adds batch correction and single-cell-derived references, which materially improves fraction estimates. ESTIMATE takes a different route, computing stromal and immune scores from expression ranks to yield a tumor purity estimate2. The two are not interchangeable: CIBERSORT gives composition, ESTIMATE gives an aggregate load and purity that is often the most robust single feature in a prognosis model.

The honest caveats are these. Bulk deconvolution is unidentifiable without strong reference assumptions; fractions are relative, not absolute; and results shift with reference matrix choice. Use at least two methods and require that a prognostic immune feature carries the same sign in both before believing it.

Methylation-based prognosis and CpG islands

DNA methylation is a stable, heritable-ish mark that outperforms expression in several tumor types precisely because it is less noisy. CpG-island promoter hypermethylation typically silences tumor suppressors, while gene-body and intergenic hypomethylation is associated with activation and genomic instability. Prognostic methylation signatures are built either from individual CpGs selected by penalized Cox regression or from aggregated regional scores across CpG islands and shores.

There is a methodological trap worth naming: methylation and expression are not one-to-one, and correlation depends on genomic context. A CpG that predicts survival does not imply the nearby gene is the mechanism. Treat methylation signatures as prognostic features first and mechanistic claims second, and validate on an independent platform because array chemistry differences across 27K, 450K, and EPIC arrays change the measured value.

TMB and mutational features

Tumor mutational burden — nonsynonymous mutations per megabase — has become both a biomarker and a moving target, since it depends on panel size, variant-calling pipeline, and germline filtering. Within a fixed pipeline it is a useful prognostic and immunotherapy-response variable; across pipelines, thresholds do not transfer. Complement TMB with mutational signatures derived from the six substitution classes and their trinucleotide context, which capture the mutational processes and are more mechanistically informative than the raw count.

Constructing a signature that survives validation

Integration and the multiplicity problem

Strong models combine modalities: a methylation score, an immune-load score, TMB, and clinical stage in one Cox model. But every added feature multiplies the number of hypotheses tested during discovery. Pre-register the feature families, correct within families, and never select features on the validation cohort even informally. The reproducibility crisis in prognostic signatures is largely a multiplicity problem dressed as a modeling problem.

A prognosis model is a claim about future patients. Build it so that a stranger with your supplementary table can reproduce the number, and so that the number holds in their hospital's cohort, not just yours.

References

  1. Newman AM, Liu CL, Green MR, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12(5):453-457.
  2. Yoshihara K, Shahmoradgoli M, Martínez E, et al. Inferring tumour purity and stromal and immune cell admixture from expression data. Nat Commun. 2013;4:2612.
  3. Newman AM, Steen CB, Liu CL, et al. Determining cell type abundance and expression from bulk tissues with digital cytometry. Nat Biotechnol. 2019;37(7):773-782.
  4. Liu J, Lichtenberg T, Hoadley KA, et al. An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell. 2018;173(2):400-416.
  5. Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med. 1997;16(4):385-395.
  6. Alexandrov LB, Nik-Zainal S, Wedge DC, et al. Signatures of mutational processes in human cancer. Nature. 2013;500(7463):415-421.
  7. Chan TA, Yarchoan M, Jaffee E, et al. Development of tumor mutation burden as an immunotherapy biomarker. Ann Oncol. 2019;30(1):44-56.
  8. TCGA Research Network. Comprehensive molecular characterization of human colon and rectal cancer. Nature. 2012;487(7407):330-337.
  9. Harrell FE Jr, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996;15(4):361-387.