Microbiome Self-Supervised Learning

The gut microbiome shapes many aspects of human health, but machine learning on metagenomic data is hard: labeled samples are scarce, feature spaces are high-dimensional, and batch effects differ sharply between cohorts. This work from the Segal lab at the Weizmann Institute of Science, with collaborators including Katherine Pollard, was released as a preprint in October 2025. It adapts the foundation-model recipe—self-supervised pretraining followed by transfer—to bacterial abundance profiles so that downstream predictors can work in data-limited and cross-cohort settings.

The authors pretrain several self-supervised models on a large collection of 85,364 metagenomic samples. These include masked autoencoders with varying masking rates and single-cell RNA-seq architectures—scVI and scGPT—adapted from their original transcriptomic setting to microbiome abundance tables. The learned embeddings act as drop-in replacements for raw bacterial abundances when training predictive models, and because they can be computed once and frozen, they transfer across datasets without retraining the representation.

Rather than a single named model, the study presents a family of representation-learning approaches and shows consistent advantages of learned embeddings over raw abundances across multiple phenotype-prediction tasks.

#Key Features

• Large pretraining corpus: Representations are learned from 85,364 gut metagenomic samples, an unusually large collection for microbiome machine learning.
• Multiple self-supervised methods: Includes masked autoencoders at several masking rates alongside scVI and scGPT architectures adapted from single-cell RNA-seq.
• Data-efficient prediction: Learned embeddings outperform raw abundances when only very limited labeled data are available for training downstream models.
• Cross-cohort transfer: Frozen embeddings generalize across datasets, consistently beating raw-abundance baselines in cross-dataset evaluation and mitigating batch effects.

#Technical Details

The models operate on bacterial abundance profiles rather than raw sequencing reads, learning lower-dimensional embeddings through self-supervised objectives (masked reconstruction for the autoencoders, and the generative/variational objectives inherited from the adapted scVI and scGPT architectures). With very limited labeled data, embeddings improve continuous-trait prediction over raw abundances: age (Pearson r = 0.14 vs. 0.06), BMI (r = 0.16 vs. 0.11), and visceral fat mass (r = 0.25 vs. 0.18), as well as drug-usage classification (PR-AUC = 0.81 vs. 0.73). The same embeddings deliver consistent gains when models are transferred across cohorts.

#Applications

The approach is aimed at microbiome researchers building predictive models of host phenotypes—age, adiposity, medication use, and related traits—from metagenomic profiles, especially when labeled samples are scarce or when a model trained on one cohort must generalize to another. By supplying pretrained embeddings that absorb structure from tens of thousands of unlabeled samples, it offers a practical way to improve accuracy and robustness in the small-data, multi-cohort regime that characterizes much of human microbiome science.

#Impact

This study provides a framework for bringing self-supervised and transfer-learning ideas—already central to protein, genomic, and single-cell modeling—to gut microbiome data, and shows that learned representations reliably outperform raw abundances in the data-limited and cross-cohort settings that most constrain the field. It also demonstrates that single-cell architectures like scVI and scGPT can be repurposed for metagenomic abundance tables. As a preprint awaiting peer review, its results are established through in-silico benchmarks, and no code or model weights accompany the current release.