Tabular Foundation Models Lack Robustness to Microbiome Data Shifts

Tabular Foundation Models (TFMs) have shown strong performance in analyzing microbiome abundance data, but their resilience to realistic distribution shifts, which are common in biological datasets, remains largely unexplored. This study addresses this gap by presenting a new benchmark designed to assess TFM robustness under biologically inspired perturbations across six gut microbiome datasets, covering four different disease contexts. The evaluation focuses on an in-context learning setting, where models are given unperturbed support sets and then tested on perturbed query samples. To isolate robustness beyond simple "shortcut" features, the researchers preserved the most discriminative taxa while applying three controlled perturbation strategies: removing high-abundance uninformative taxa, increasing zero-inflation (sparsification), and introducing spurious non-zero injections (zero-imputation). The results indicate that merely protecting discriminative features is insufficient to guarantee stability under support-query shifts. All perturbations degraded model performance, with zero-imputation consistently proving the most harmful. This suggests that corrupting the global feature structure significantly impairs generalization, even when key taxa are retained. Furthermore, sparsification disproportionately affected TFMs compared to a classical random forest baseline, highlighting TFMs' greater sensitivity to zero-inflation-type shifts.