Ancestry-Invariant Prognostic Gene Signatures

A prognostic gene signature is a short weighted sum of a handful of tumor-gene expression measurements — typically three to twenty genes — that turns a patient's molecular profile into a single risk score for predicting how aggressively their cancer will progress. Virtually every such signature in use or in the literature is fitted on bulk cohorts that pool patients without regard to genetic ancestry, and those cohorts are overwhelmingly European-descent. Whether the resulting risk equations transfer to patients of African, East Asian, or other ancestries has, to our knowledge, never been systematically tested. This matters because germline genetics has already learned an uncomfortable lesson: the biological effects of genetic variants are almost perfectly conserved across ancestries — a correlation of 0.98 ± 0.07 across 47 traits between African- and European-ancestry segments in the UK Biobank (Fine-scale population structure and widespread conservation of genetic effect sizes between human groups across traits) — and yet predictors built in one ancestry fail badly in another, because the observable quantities (tagging variants, allele frequencies, baseline expression levels) shift between populations (On cross-ancestry cancer polygenic risk scores).

We propose to bring that same lens to transcriptomic prognostic signatures. The core hypothesis is that the relationship between a tumor gene's expression and patient survival is largely conserved across ancestries for some biological processes (we predict proliferation, following the evidence that three non-overlapping breast cancer signatures all track the same proliferative program and agree on 135 of 198 patients in the TRANSBIG validation, Comparison of prognostic gene expression signatures for breast cancer), but that signatures accidentally include genes whose baseline regulation and survival association are ancestry-specific — immune and microenvironment genes being the prime suspects, since African American colon tumors show systematically altered immune-cell fractions and downregulated MHC and checkpoint genes relative to European American tumors (Differential immune signatures in the tumor microenvironment are associated with colon cancer racial disparities). The deliverable is a new selection principle: build signatures only from genes whose survival association is demonstrably invariant across ancestry-defined environments, using ideas from invariant causal discovery (Invariant Ancestry Search), and pair this with within-ancestry score recalibration, which is exactly what rescued cross-ancestry cancer polygenic risk scores (breast cancer top-decile enrichment odds ratios of 2.81 in Europeans and 2.88 in Africans after recalibration, On cross-ancestry cancer polygenic risk scores).

Core Claim

The claim has three parts, in increasing order of ambition. First, an audit: standard prognostic signatures trained on European-ancestry TCGA patients lose measurable accuracy when applied to African- and East Asian-ancestry patients, and the size of that loss varies systematically by the biological module the signature draws from. Second, a measurement: the per-gene correlation of Cox regression coefficients between ancestries — the transcriptomic analogue of the 0.98 germline figure — is high but imperfect, and the non-invariant genes are concentrated in immune and microenvironment pathways rather than cell-cycle pathways. Third, a construction: restricting the standard two-stage LASSO Cox pipeline to an ancestry-invariant gene pool yields signatures that (a) match within-European performance and (b) substantially close the cross-ancestry transfer gap, without needing ancestry-specific training cohorts that are too small to support model fitting in the first place.

The invariance criterion is the conceptual heart of the proposal. Following the environment-invariance idea from Invariant Ancestry Search, which identifies causal ancestors as genes whose behavior is stable across perturbation environments, we treat genetically inferred ancestry as the environment and ask for genes whose conditional relationship to survival does not depend on it. For a candidate gene set $S$ with expression values $X_S$, survival time $Y$, and ancestry environment $E$ (inferred as continuous or fine-scale components, not self-reported categories), invariance means

$$P(Y \mid X_S, E = e) = P(Y \mid X_S) \quad \text{for all environments } e,$$

and the risk score keeps the standard form $\text{Risk} = \sum_i \beta_i \times \text{Expression of gene } i$ with coefficients fitted on the pooled cohort. Invariant Ancestry Search validated this machinery on yeast deletions and synthetic graphs, finding 23 invariant pairs (7 confirmed true) versus 8 for its predecessor; it has never touched human cancer data, so applying it where the "perturbation" is human demographic history rather than a knockout is itself a genuine test of the method's reach.

Research Program

The program has four components. The first is ancestry inference on TCGA: from existing germline genotype data, infer fine-scale ancestry components following the methodology of Fine-scale population structure and widespread conservation of genetic effect sizes between human groups across traits, which showed that categorical labels and even standard principal components fail to correct stratification where continuous components succeed, optionally augmented with the continuous geographic-coordinate inference of XGMix (Addressing Ancestry Disparities in Genomic Medicine: A Geographic-aware Algorithm). This replaces self-reported race — the coarse label that weakened the colon-cancer disparities analysis, which had only 58 African American patients against 196 European Americans and could not resolve overall survival differences (Differential immune signatures in the tumor microenvironment are associated with colon cancer racial disparities) — with quantitative environments.

The second component is a gene-level invariance filter. For each gene, we fit univariate Cox models stratified by ancestry environment and test coefficient homogeneity (an ancestry-by-expression interaction test, with permutation inference and false-discovery-rate control). Because per-cancer non-European samples are small, we pool evidence meta-analytically across cancer types, treating cancer type as a stratification variable, and we report a simulation-based power analysis so reviewers can see what effect sizes are detectable at the available sample sizes. Genes passing the filter form the invariant pool; a matched non-invariant pool (genes that fail it) is retained for contrast.

The third component is signature construction: the standard two-stage pipeline (univariate screening, then LASSO Cox) run twice — once on all genes (the status quo) and once restricted to the invariant pool. We additionally test a set-level variant that adapts the minimally-invariant-set search of Invariant Ancestry Search to groups of co-regulated genes, in case invariance holds for modules but not single genes. Scores are then calibrated two ways: a single pooled cutoff, and within-ancestry decile recalibration modeled directly on the cross-ancestry PRS result that relative risk enrichment is conserved once absolute score scales are recentered per ancestry (On cross-ancestry cancer polygenic risk scores).

The fourth component is interpretation: we annotate invariant versus non-invariant genes against the African and European eQTL catalogs from The genetic architecture of gene expression in individuals of African and European ancestry, which found that most detected eQTLs are population-specific in practice (21,130 in the African blood cohort versus 5,661 in the European, with African fine-mapping credible sets roughly two times smaller), to test whether non-invariant prognostic genes are disproportionately governed by ancestry-specific regulatory variants. We flag up front that those catalogs cover blood and lymphoblastoid lines, not tumors, so this annotation is suggestive rather than decisive — a limitation the field should fix with multi-ancestry tumor eQTL studies.

Proposed Experiments

The portability audit (Experiment 1) uses TCGA bulk RNA-seq and survival data across the cancer types with at least roughly forty non-European patients (breast, lung adenocarcinoma, colon, prostate, endometrial among the candidates). Standard LASSO Cox signatures and a fine-tuned BulkRNAbert (BulkRNABert: Cancer prognosis from bulk RNA-seq based language models, using the released pre-trained weights in instadeepai/multiomics-open-research) are trained on European-ancestry patients only and evaluated with Uno's concordance index on held-out European, African, and East Asian patients separately; we report the transfer gap, the difference between within-European and cross-ancestry C-indices, per cancer type and per signature module (proliferation, immune, ferroptosis/anoikis-style death programs). The prior from Differential immune signatures in the tumor microenvironment are associated with colon cancer racial disparities is that immune-module signatures degrade most.

Experiment 2 estimates the transcriptomic conservation coefficient: per gene and per cancer type, Cox coefficients fitted within each ancestry, combined by precision-weighted meta-analysis, and summarized as a cross-ancestry correlation directly comparable to the germline 0.98 ± 0.07 benchmark of Fine-scale population structure and widespread conservation of genetic effect sizes between human groups across traits. A materially lower correlation for immune genes than proliferation genes would confirm the mechanism the proposal is built on; a uniformly high correlation would instead localize the portability problem in baseline expression shifts rather than effect divergence, which points the remedy at recalibration alone — either outcome is informative and we commit to reporting both.

Experiment 3 is the head-to-head evaluation, run under ancestry-stratified cross-validation with identical splits, identical LASSO penalty grids, and identical gene-count budgets for every method: clinical-only Cox, pooled LASSO Cox over all genes, LASSO Cox restricted to the invariant pool, and fine-tuned BulkRNAbert. Primary endpoint is the worst-ancestry (min across ancestry groups) Uno C-index, with integrated Brier score and per-ancestry Kaplan-Meier log-rank hazard ratios as secondaries, evaluated both within individual cancer types and on the six-cohort TCGA-SurvReport setting used by CACSurv: Concordance-Aligned Comparative Learning with Large Language Models for Cancer Survival Prediction, where CACSurv reported an average C-index of 0.722 against 0.657 for the strongest non-LLM baseline. Robustness to small-sample variance is checked by repeating the pipeline on the 174 few-shot tasks of the TCGA Meta-Dataset Clinical Benchmark (The TCGA Meta-Dataset Clinical Benchmark), whose baselines (a neural network at 71.19% ± 14.62% average accuracy) show how much variance small clinical tasks carry; we ask whether the invariant pool reduces that variance, since it shrinks the effective search space.

The decisive ablation (Experiment 4) replaces the invariant pool with a random gene pool matched for size and for pooled univariate survival significance. If invariance selection is doing real work, the matched-random signature should show the same transfer gap as the unrestricted pooled signature while the invariant one closes it; if matched-random does just as well, the invariance story is wrong and the gap was driven by univariate fragility, not ancestry — a result we would report and act on.

Baselines

The naive reference point is a Cox proportional hazards model using only standard clinical variables (age, sex, stage, where available). It is the floor every molecular method must clear, and it is a genuinely strong floor: the large TCGA benchmark across 18 cancer types and 11 methods found that unstructured multi-omics models, including flat Lasso and random forests, failed to outperform this simple clinical Cox model, with only the structured Block Forest marginally beating it on average (Large-scale benchmark study of survival prediction methods using multi-omics data). No reported result found on Lacuna for a numeric C-index of that clinical-only baseline, so we will establish its value ourselves on our splits and treat it as the do-nothing reference. Beating it means a strictly higher worst-ancestry Uno C-index with confidence intervals from the stratified cross-validation folds.

The strongest existing method depends on the axis of comparison. For raw pan-cancer survival accuracy, the strongest cited system is BulkRNAbert, a transformer encoder pre-trained with masked modeling on more than 30,000 bulk RNA-seq samples and fine-tuned with parameter-efficient adaptation, which reports a pan-cancer survival C-index of 0.765 (BulkRNABert: Cancer prognosis from bulk RNA-seq based language models). We will fine-tune the released weights on our European-only training splits with the same compute and tuning budget as our linear models receive, and compare worst-ancestry C-indices; because BulkRNAbert ignores population structure entirely, our claim is not that we exceed 0.765 overall but that our worst-ancestry number beats its worst-ancestry number at equal within-European performance. For the portable-linear-sigature question, the strongest competitor is the status-quo pipeline itself: the two-stage univariate-screen-then-LASSO-Cox architecture that dominates the prognostic-signature literature, run on the pooled cohort without any ancestry treatment. It shares our data, splits, and penalty grids; the only difference is the absence of the invariance constraint. Beating it requires a strictly smaller cross-ancestry transfer gap at matched gene count, with within-European performance within 0.01–0.02 C-index. The matched-random-pool ablation described above isolates the key ingredient — ancestry-invariance selection as such — from generic dimensionality reduction, and is the check that most directly tests the proposal's causal story.

Why not just…?

Why not just add self-reported race as a covariate in the Cox model? This is the first thing a clinician would suggest, and the file's evidence is against it twice over. Addressing Ancestry Disparities in Genomic Medicine: A Geographic-aware Algorithm shows discrete racial labels cannot represent continuous ancestry gradients or admixed individuals, and Fine-scale population structure and widespread conservation of genetic effect sizes between human groups across traits shows even standard principal components leave stratification artifacts that fine-scale components remove. More fundamentally, a covariate adjusts the intercept, not the gene coefficients: it says nothing about whether a gene's survival signal itself differs by ancestry, which is the quantity our invariance filter tests. Our Experiment 3 will include the race-covariate model explicitly so the comparison is empirical, not rhetorical.

Why not just train a separate signature for each ancestry? With 58 African American colon cancer patients in TCGA (Differential immune signatures in the tumor microenvironment are associated with colon cancer racial disparities) and similar ratios across tumors, per-ancestry training is statistically starved; the few-shot results in The TCGA Meta-Dataset Clinical Benchmark (models trained on 50 samples with double-digit accuracy standard deviations) indicate what happens when you shrink training sets further. Meanwhile the conservation result of Fine-scale population structure and widespread conservation of genetic effect sizes between human groups across traits says most of the underlying biology is shared, so siloing models throws away the pooled sample exactly where sharing is safe. Our design keeps the pooling and adds a constraint, which is the statistically efficient compromise. The honest caveat: if Experiment 2 reveals large genuinely ancestry-specific effects, per-ancestry modeling becomes more attractive, and we will say so — but at current sample sizes it cannot be executed reliably regardless.

Why not just use clinical variables and skip molecular signatures? Partly a real threat, and we treat it as such rather than argue around it: Large-scale benchmark study of survival prediction methods using multi-omics data is a strong negative result for unstructured molecular models. But the same study shows structured molecular methods do add value in specific cancers, and Comparison of prognostic gene expression signatures for breast cancer shows the 70-gene signature identifying patients 7.1 times more likely to develop distant metastasis, with a 10-year C-index of 0.90 on TRANSBIG. Our framing also changes what "sufficient" means: the clinical baseline is evaluated for equity across ancestries as well as average accuracy, and if it turns out to be both accurate and equitable while molecular signatures are not, that is a publishable and clinically important finding of the audit itself — the proposal does not depend on the molecular model winning everywhere, only on the portability question being answered honestly.

Decisive Checks

The proposal lives or dies on three numbers. The first is the measured transfer gap of Experiment 1: if European-trained signatures show negligible cross-ancestry degradation, the motivating problem evaporates and the paper becomes a short negative result, which we would write. The second is the module split of Experiment 2: conservation near the germline 0.98 benchmark for proliferation genes and materially lower for immune genes would confirm the mechanism; uniform conservation would redirect the remedy toward pure recalibration, in line with the PRS finding that within-ancestry relative ranking survives transfer (On cross-ancestry cancer polygenic risk scores). The third is the matched-random ablation of Experiment 4: only if invariant-pool selection beats matched-random selection at closing the transfer gap does the invariance principle — as opposed to any sparse-selection routine — earn its place in the signature-builder's toolkit. All three checks are pre-specified, run on public data (TCGA, with 1000 Genomes reference panels for ancestry inference), and can be reproduced end-to-end from the released BulkRNAbert codebase (instadeepai/multiomics-open-research) plus standard survival tooling.