TaxoMFM: A Taxonomy-Aware Foundation Model for the Human Gut Microbiome

Haochen Xue; Anastasia Godneva; Yulong Li; Qi Cao; Zhaodong Wu; Kaiyu Liang; Jionglong Su; Eran Segal

cs.LGOct 29, 2026

TaxoMFM: A Taxonomy-Aware Foundation Model for the Human Gut Microbiome

Authors: Haochen Xue, Anastasia Godneva, Yulong Li, Qi Cao, Zhaodong Wu, Kaiyu Liang, Jionglong Su, Eran Segal,

Organizations: Xi’an Jiaotong-Liverpool University · Mohamed bin Zayed University of Artificial Intelligence · Weizmann Institute of Science

Abstract

Existing gut microbiome foundation models typically use flat taxon representations, limiting their ability to exploit known taxonomic hierarchies. We introduce TaxoMFM, a microbiome foundation model that integrates taxonomic structure into self-supervised learning. TaxoMFM jointly learns abundance patterns across taxonomic ranks and relative proportions within groups through taxonomy-aware encoding and hierarchical compositional reconstruction. We further construct TaxoGut-HPP, comprising 100,806 gut microbiome profiles from 231 source cohorts and covering 40 classification and 8 regression tasks. TaxoMFM achieves competitive performance across disease, gut health, host trait, and liver ultrasound phenotype prediction. In cross-cohort zero-shot transfer, where target cohorts are excluded from both pretraining and downstream adaptation, TaxoMFM achieves AUROCs of 0.838, 0.736, and 0.722 for colorectal cancer, type 2 diabetes, and inflammatory bowel disease, respectively. Further analyses show that TaxoMFM recovers community changes across taxonomic scales, while preserving community information and producing consistent representations under database updates and species-vocabulary shifts. Our code is available at https://anonymous.4open.science/r/TaxoMFM-F16D/.

Figures & tables

Figure 1: Conceptual comparison of flat and taxonomy-aware microbiome foundation models. Both approaches use the same species-level abundance profile. Existing models represent species as a flat set, whereas TaxoMFM aggregates their abundances along known taxonomic lineages and jointly models species, genus, family, and order nodes to learn a structure-aware representation.

Model	Size	Cohorts	Downstream task counts			Downstream task coverage
Model	Size	Cohorts	Total	Classification	Regression	Disease	Gut health	Host traits	Clinical phenotype
Pope et al. ( Pope et al., 2025 )	18,480	1	3	3	0	$\checkmark$	$\times$	$\checkmark$	$\times$
BiomeGPT ( Medearis et al., 2026 )	13,524	86	27	27	0	$\checkmark$	$\times$	$\times$	$\times$
GUT-FORMer ( Karwowska et al., 2026 )	21,024	93	27	25	2	$\checkmark$	$\times$	$\checkmark$	$\times$
MBEmbed ( Zahavi et al., 2025 )	85,364	3	7	4	3	$\checkmark$	$\times$	$\checkmark$	$\checkmark$
TaxoMFM (ours)	100,806	231	48	40	8	$\bm{\checkmark}$	$\bm{\checkmark}$	$\bm{\checkmark}$	$\bm{\checkmark}$

Table 1: Comparison of pretraining and downstream evaluation datasets used by human gut microbiome foundation models.

Figure 2: TaxoMFM architecture and self-supervised pretraining. (A) The encoder combines taxon identity, taxonomic rank, and abundance embeddings with relation-biased self-attention and rank-wise pooling to generate the sample representation

z

. (B) The decoder predicts child proportions, reconstructing abundance profiles top-down through the taxonomy. (C) Pretraining masks species or subtrees and optimizes full-tree reconstruction and within-parent child-proportion matching.

Standard baselines
Model	Disease					Gut health (HPP)
Model	Colorectal cancer	Type 2 diabetes	IBD	Liver cirrhosis	NAFLD	Current abdominal pain	Abdominal pain frequency	Current bloating	GI sensitivity	Irritable bowel syndrome	H. pylori infection/history	Gastro-oesophageal reflux disease
Elastic Net	0.826	0.791	0.868	0.811	0.807	0.593	0.572	0.637	0.632	0.582	0.628	0.651
Random Forest	0.839	0.917	0.853	0.836	0.819	0.607	0.583	0.691	0.646	0.547	0.667	0.663
DeepMicro	0.803	0.763	0.865	0.816	0.832	0.602	0.574	0.658	0.627	0.562	0.628	0.646
Taxonomy-aware models

Table 2: Representative classification performance. Binary tasks use AUROC and abdominal pain frequency uses macro-AUROC (

\uparrow

). The baselines use the locked TaxoGut-HPP tasks and splits.

Standard baselines
Model	Host traits			Liver ultrasound phenotypes
Model	Age	BMI	VAT mass	Attenuation coefficient	Speed of sound	Velocity
Elastic Net	0.132 / 0.038	0.148 / 0.022	0.214 / 0.048	0.136 / 0.038	0.113 / 0.036	0.111 / 0.026
Random Forest	0.117 / 0.031	0.191 / 0.045	0.246 / 0.075	0.128 / 0.040	0.193 / 0.042	0.136 / 0.024
DeepMicro	0.280 / 0.066	0.221 / 0.065	0.326 / 0.108	0.162 / 0.061	0.128 / 0.033	0.135 / 0.037
Taxonomy-aware models

Table 3: Representative regression results. In each paired result, the first value is Spearman

\rho

and the second is

R^{2}

(

\uparrow

). The baselines use the locked TaxoGut-HPP tasks and splits.

Model variant	Disease classification			Gut health classification			Liver phenotype regression
Model variant	Colorectal cancer	IBD	Type 2 diabetes	Current abdominal pain	Current bloating	Irritable bowel syndrome	Attenuation coefficient	Speed of sound	Velocity
Structural ablations
Species-only (flat)	0.914	0.943	0.909	0.736	0.758	0.723	0.313/0.113	0.320/0.125	0.226/0.060
w/o relation bias	0.934	0.964	0.926	0.771	0.782	0.715	0.327/0.123	0.331/0.130	0.273 / 0.088
Random taxonomy	0.906	0.933	0.880	0.686	0.694	0.635	0.223/0.084	0.206/0.071	0.158/0.035
Pretraining objective ablations

Table 4: Ablation study of TaxoMFM. Classification tasks report AUROC (

\uparrow

), and liver phenotype regression reports Spearman’s

\rho

R^{2}

(

\uparrow

Table 7

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Standard baselines
Model	Healthy vs. diseased	Crohn’s disease	Ulcerative colitis	Multiple sclerosis	Rheumatoid arthritis
Elastic Net	0.760	0.872	0.849	0.869	0.834
Random Forest	0.782	0.908	0.872	0.901	0.859
DeepMicro	0.809	0.873	0.868	0.908	0.861
ExtraTrees	0.773	0.902	0.882	0.895	0.864
Taxonomy-aware models

Appendix

Table A1: Additional AUROC results for general and immune-related diseases (part I).

Standard baselines
Model	Ankylosing spondylitis	Graves’ disease	Behçet’s disease	Myalgic encephalomyelitis / chronic fatigue syndrome
Elastic Net	0.892	0.883	0.904	0.862
Random Forest	0.923	0.902	0.947	0.889
DeepMicro	0.907	0.910	0.922	0.902
ExtraTrees	0.917	0.930	0.941	0.877
Taxonomy-aware models

Appendix

Table A2: Additional AUROC results for general and immune-related diseases (part II).

Standard baselines
Model	Colorectal adenoma	Pancreatic cancer	Coronary artery disease	Atherosclerotic cardiovascular disease	Impaired glucose tolerance	Melanoma
Elastic Net	0.833	0.896	0.874	0.903	0.788	0.874
Random Forest	0.868	0.932	0.903	0.934	0.817	0.906
DeepMicro	0.851	0.918	0.882	0.919	0.825	0.895
ExtraTrees	0.873	0.925	0.896	0.928	0.806	0.899
Taxonomy-aware models

Appendix

Table A3: Additional AUROC results for neoplastic and cardiometabolic diseases.

Standard baselines
Model	Parkinson’s disease	Schizophrenia	Soil-transmitted helminths	Clostridioides difficile infection	Acute diarrhea	Premature birth
Elastic Net	0.898	0.883	0.916	0.869	0.922	0.904
Random Forest	0.924	0.918	0.944	0.899	0.957	0.932
DeepMicro	0.916	0.925	0.930	0.920	0.949	0.941
ExtraTrees	0.919	0.911	0.956	0.883	0.950	0.928
Taxonomy-aware models

Appendix

Table A4: Additional AUROC results for neurological, infectious, and host-related tasks.

Standard baselines
Model	Soft-stool frequency	Hard-stool frequency	Nausea in the past three months	Anorectal disease	Gallstone / cholecystectomy
Elastic Net	0.596	0.587	0.664	0.624	0.669
Random Forest	0.638	0.625	0.698	0.657	0.706
DeepMicro	0.632	0.619	0.728	0.668	0.719
ExtraTrees	0.613	0.621	0.687	0.646	0.698
Taxonomy-aware models

Appendix

Table A5: Additional HPP classification results for gastrointestinal outcomes. Soft- and hard-stool frequency use macro-AUROC; the remaining tasks use AUROC.

Standard baselines
Model	Fatty liver history	Peptic ulcer disease	Bowel-movement satisfaction	Daily bowel-movement count
Elastic Net	0.657	0.629	0.118 / -0.012	0.183 / 0.041
Random Forest	0.741	0.612	0.156 / 0.019	0.258 / 0.096
ExtraTrees	0.726	0.624	0.147 / 0.023	0.246 / 0.101
DeepMicro	0.693	0.655	0.193 / 0.031	0.229 / 0.071
Taxonomy-aware models

Appendix

Table A6: Additional HPP classification and regression results. The first two columns report AUROC; the final two report Spearman

\rho

R^{2}

(all

\uparrow

Task	Source samples	Target study	Target samples	Positive/ negative
CRC	1,517	Yang et al. (2021)	194	97/97
T2D	465	MetaCardis ( Forslund et al., 2021 )	783	549/234
IBD	318	HMP/IBDMDB ( Lloyd-Price et al., 2019 )	130	103/27

Appendix

Table A7: Source and target cohort splits for cross-cohort transfer. Positive and negative counts refer to the target test sets.

Task	Role	Resource	Resource study ID	PMID
CRC	Source	cMD	GuptaA_2019	31719139
CRC	Source	cMD	HanniganGD_2017	30459201
CRC	Source	cMD	ThomasAM_2018b	30936548
CRC	Source	cMD	VogtmannE_2016	27171425
CRC	Source	cMD	WirbelJ_2018	30936547
CRC	Source	cMD	YachidaS_2019	31171880

Appendix

Table A8: Resource study IDs and PubMed identifiers for cross-cohort transfer.

Model	Parameters (M)	GFLOPs / sample	Latency (ms) $\downarrow$	Throughput (samples/s) $\uparrow$	Peak memory (GiB) $\downarrow$
MGM	6.83	0.692	3.399	2,113	0.063
Waypoint-45M	45.15	4.844	5.732	1,341	0.237
BiomeGPT	14.73	1.991	4.262	1,809	0.114
GUT-FORMer	1.22	0.483	0.951	8,392	0.029
TaxoMFM	18.25	5.626	0.759	3,852	0.412

Appendix

Table A9: Computational efficiency of representation extraction.

Bioinformatics

See all 43 papers

Are Tabular Foundation Models Robust to Realistic Query Distribution Shifts in Microbiome Data?
Jun 23, 2026Giulia Perciballi, Ahmad Fall, Federica Granese +2Distribution Shift Robustness Tabular Foundation Models
IRD, Sorbonne Université, Unité de Modélisation Mathématique et Informatique des Systèmes Complexes (UMMISCO), Paris, France · Inria, CNRS, I3S, Université Côte d’Azur, Valbonne, France · INSERM, Nutrition et Obésités; Approches Systémiques, NutriOmique, AP-HP, Hôpital Pitié-Salpêtrière, Paris, France
HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference
Oct 19, 2026Akash Rao, Sushim Gupta, Vishalini Ramnath +2Bioinformatics Representation Learning
CSSE Department, Auburn University, Auburn, AL, USA · Oklahoma Animal Disease Diagnostic Laboratory, Oklahoma State University, Stillwater, OK, USA
Set-Aggregated Genome Embeddings for Microbiome Abundance Prediction
May 12, 2026Younhun Kim, Georg K. Gerber, Travis E. GibsonRepresentation Learning Genomics
1Brigham and Women’s Hospital, Boston, MA, USA · 2Harvard Medical School, Boston, MA, USA

Jun 23, 2026J/K move · Enter open · S save

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

Giulia Perciballi, Ahmad Fall, Federica Granese, Edi Prifti+1

IRD, Sorbonne Université, Unité de Modélisation Mathématique et Informatique des Systèmes Complexes (UMMISCO), Paris, France · Inria, CNRS, I3S, Université Côte d’Azur, Valbonne, France · INSERM, Nutrition et Obésités; Approches Systémiques, NutriOmique, AP-HP, Hôpital Pitié-Salpêtrière, Paris, France

Distribution Shift Robustness Tabular Foundation Models Tabular ICL

Open paper AskPDF arXiv

Tabular foundation models (TFMs) achieve strong performance on microbiome abundance data, yet their robustness under realistic distribution shift remains poorly characterized. We introduce a benchmark that evaluates the robustness of TFMs to biologically inspired perturbations across six gut microbiome datasets spanning four disease contexts. In this in-context learning setting, models receive unperturbed support sets as context and are evaluated on perturbed query samples. To isolate robustness beyond "shortcut" features, we preserve the most discriminative taxa and apply three controlled perturbation strategies: (i) removal of high-abundance (uninformative) taxa, (ii) sparsification via increased zero-inflation, and (iii) zero-imputation via spurious non-zero injections. Our results show that protecting discriminative features is insufficient to guarantee stability under support-query shift: across datasets, all perturbations degrade model performance, with zero-imputation consistently the most harmful, indicating that corrupting global feature structure can break generalization even when key taxa are retained. Sparsification disproportionately affects TFMs relative to a classical random forest baseline, suggesting greater sensitivity to zero-inflation-type shifts. The code is publicly available at: https://github.com/UMMISCO/metagenomics-fm/.

Abstract

Bioinformatics

Explore similar work

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

HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference

Set-Aggregated Genome Embeddings for Microbiome Abundance Prediction