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

Researchers benchmarked tabular foundation models (TFMs) on microbiome data to test their robustness against realistic distribution shifts, finding that all models degrade significantly under perturbations even when key discriminative features are preserved. The study reveals that TFMs are particularly vulnerable to zero-inflation shifts and global feature structure corruption, suggesting current foundation model architectures may struggle with real-world data variability in biological applications.

This research addresses a critical gap in foundation model evaluation by moving beyond static benchmarks to test performance under realistic biological perturbations. The study acknowledges that while tabular foundation models show promise on microbiome datasets, their behavior under distribution shift—a common problem in real-world deployments—remains largely unexplored. The researchers developed a rigorous evaluation framework using six gut microbiome datasets across four disease contexts, employing an in-context learning setup that mirrors practical use cases where models must generalize from clean training data to noisy query samples.

The findings present sobering implications for deploying TFMs in biomedical applications. Despite preserving the most discriminative taxa, all perturbation strategies degraded model performance, with zero-imputation (spurious non-zero injections) proving most damaging. This indicates that foundation models rely on global feature relationships rather than isolated key features, making them brittle when data structure changes. Notably, TFMs showed disproportionate sensitivity to sparsification compared to classical random forests, suggesting that the architectural advantages of foundation models may not extend to handling missing-data patterns common in real microbiome studies.

These results carry substantial weight for the machine learning and biotech sectors. Organizations deploying foundation models for biological analysis cannot assume robustness based on static benchmark performance. The research suggests that simpler, more interpretable baseline methods may offer better reliability guarantees in high-stakes applications where distribution shifts are unavoidable. Future development should prioritize architectural innovations that handle zero-inflation patterns and feature-level corruption more gracefully.

• →Tabular foundation models show significant performance degradation under realistic microbiome data perturbations despite preserving discriminative features.
• →Zero-imputation and sparsification perturbations prove most harmful to TFM robustness, indicating vulnerability to global feature structure changes.
• →TFMs exhibit greater sensitivity to zero-inflation shifts than classical random forest baselines, questioning their reliability advantage in biological applications.
• →Current foundation model architectures rely on implicit feature relationships that break under distribution shift, limiting deployment in high-variability real-world settings.
• →The benchmark framework enables systematic evaluation of foundation model robustness and provides code for reproducible testing across microbiome datasets.