Cancer outcome models stop splitting by ancestry: DA-ICL models the genetic continuum and tops mean AUROC in 10 of 18 TCGA tasks

Omics-based cancer outcome prediction moves from ancestry groups to modeling along the genetic continuum: each ancestry point gets its own context instead of samples being sorted into discrete African, European and other groups and predicted uniformly within a group. Ancestry-discrete modeling partitions samples into disjoint ancestry groups and fits a separate model within each group, giving the same form of prediction to all samples in a group; large biobank analyses have shown accuracy declining approximately linearly with genetic distance from the training set, even within a single ancestry label. In synthetic experiments, an ancestry-discrete logistic regression degrades as a sample's distance from its ancestry-group centroid grows, while an oracle continuum model drawing from the true distribution and an empirical continuum model trained on the 100 nearest training samples stay stable over that distance range; as the rotation and offset parameters controlling how ancestry shifts the decision boundary increase, both continuum models achieve substantially higher AUROC than the discrete model.

The method couples the ancestry continuum with in-context learning: for each target ancestry point, nearby samples in genetic space form a context, and a pretrained tabular foundation model predicts from it at inference time without updating parameters for the target task. Conventional continuum modeling requires fitting a local model for every query, whereas DA-ICL conditions TabPFN 2.6 on a query-specific context and shifts the cost to constructing that context at inference time. On TCGA BRCA, KIRC and UCEC, mRNA expression, microRNA expression and DNA methylation are each paired with overall survival and progression-free interval for 18 tasks; evaluated on African-ancestry samples with three folds across 30 seeds, DA-ICL achieves the highest mean AUROC on 10 tasks, ICL with discrete context without enhancement on 6, and ICL with continuum context without enhancement on 2.

Distribution-aware context enhancement (DACE) supplements the local context with distributionally similar samples from across the continuum, easing the sparse-context problem of relying only on genetic neighbors. When context is built only from genetic-space neighbors, nearby samples are insufficient to characterize the omics-endpoint relationship in some tasks, and the continuum context without enhancement achieves the highest mean AUROC on only 2 of 18 tasks. DACE builds a separate minimum spanning tree within each outcome class in an omics autoencoder embedding and adds candidates on edges crossing into the nearest-neighbor context; with it, DA-ICL achieves the highest mean AUROC on a majority but not all tasks, so the benefit is task-dependent.

The evaluation covers synthetic data and TCGA BRCA, KIRC and UCEC mRNA, microRNA and DNA methylation paired with overall survival and progression-free interval, tested on African-ancestry samples; the theorem compares Bayes-optimal classifiers under the stated data-generating process.

The theoretical result concerns optimal classifiers and does not guarantee that a finite-sample DA-ICL implementation will outperform an ancestry-stratified model; real-data gains are task-dependent, and DACE achieves the highest mean AUROC on a majority but not all tasks.