Are Tabular Foundation Models Robust to Realistic Query Distribution Shifts in Microbiome Data?

As tabular foundation models (TFMs) gain popularity for analyzing microbiome data, a critical question arises: how robust are these models when the query data distribution differs from the support set provided during in-context learning? Because microbiome data is characterized by compositionality and high zero-inflation, the authors investigate whether TFMs can maintain performance when faced with realistic shifts in these structural properties.

The authors introduce a benchmark using six gut microbiome datasets across four disease contexts. They employ an in-context learning evaluation protocol where the support set remains unperturbed, while the query samples undergo three types of biologically inspired perturbations: feature removal, increased zero-inflation (sparsification), and zero-imputation. To ensure the findings reflect robustness to distribution shift rather than the loss of predictive signal, the authors use a data-driven method to identify and protect the most discriminative taxa from these perturbations.

The study reveals that protecting discriminative features is insufficient to guarantee model stability. All three perturbation strategies consistently degrade performance across all tested TFMs. Zero-imputation—filling in zero values with observed abundances—proved to be the most harmful, indicating that TFMs are highly sensitive to changes in the global feature structure. Furthermore, TFMs showed greater sensitivity to sparsification (increased zero-inflation) compared to a classical random forest baseline, suggesting that the inductive biases of current foundation models may be poorly suited to the extreme sparsity inherent in microbiome data.