Summary
A bioRxiv preprint reports that TabICL and TabPFN prediction performance declines as genetic ancestry dispersion increases in evaluation data. The authors also report that instruction-tuning TabICL on synthetic tasks designed to model ancestry-dependent effects improves performance stability across ancestry-distance groups.
A bioRxiv preprint reports that two tabular foundation models, TabICL and TabPFN, perform less well on genotype-to-phenotype prediction as genetic ancestry becomes more dispersed in the data. The authors also report that instruction-tuning TabICL on synthetic tasks designed to represent ancestry-dependent genetic effects makes its performance more stable across ancestry groups, including groups less represented among the examples provided to the model.
The study combines controlled simulations with evaluations using data from the All of Us biobank. Its findings point to a specific challenge for models trained on generic synthetic table tasks: genetic effects associated with a phenotype can vary across the ancestry spectrum, and a model may struggle when its examples do not reflect that variation.
Why genetic ancestry can challenge tabular models
Tabular foundation models are designed to make predictions from data arranged in rows and columns. In this study, the models use in-context examples: a table of examples supplied to the model to guide predictions on other cases. The authors tested whether models built for varied tabular tasks can also handle genetic data, where the relationship between a genetic variant and a phenotype may differ across ancestry groups.
The paper describes this as ancestry-dependent non-stationarity in allele effect sizes. In this context, non-stationarity means that the predictive contribution associated with a genetic variant is not constant across the ancestry spectrum. The authors argue that generic synthetic tasks used to pretrain tabular foundation models do not capture this structure.
In controlled stress tests using hierarchical Gaussian-process simulations, TabICL and TabPFN were robust when ancestry dispersion was low. As dispersion increased, both models’ predictive performance deteriorated. The researchers then examined cancer phenotypes in All of Us, a biobank with whole-genome sequencing and electronic health-record data spanning the ancestry spectrum. When they held the in-context table size, phenotype and feature set fixed, ancestry-specific tables consistently performed better than size-matched tables combining multiple ancestries. The authors interpret this comparison as evidence that ancestry dispersion was driving the performance drop.
Synthetic tasks offer a targeted adjustment
The researchers developed two synthetic task families that encode ancestry-dependent changes in genetic effects. They used tasks sampled from both families to instruction-tune an off-the-shelf TabICL model, without changing its base architecture.
In held-out All of Us evaluations covering cancer and respiratory disease phenotypes, the tuned model showed more stable performance across ancestry-distance groups. The reported improvement was especially strong among subjects whose ancestry was farthest from the centre of the model’s in-context examples—the group least represented by those examples.
The work is a computational evaluation, not a clinical study of patient outcomes. Its practical implication is that the composition of examples supplied to a genetic prediction model may matter, and that training tasks can be designed to represent ancestry-dependent effects. The preprint abstract reports the direction of these comparisons but not cohort counts or numerical performance estimates, so the findings are described here qualitatively.