Use Waypoint Online
Commercially Available Waypoint No-Code Web Server
Waypoint: Pretrained Transformer Foundation Models for Microbiome Taxonomic Profiles
Waypoint is a family of GPT-2-style causal language models for microbiome data, pretrained on taxonomic abundance profiles treated as token sequences. It was developed by Neythen J. Treloar, Saif Ur-Rehman and Jenny Yang at Outpost Bio, and is introduced in the paper "Learning the Language of the Microbiome with Transformers". The released artifacts are licensed under Apache 2.0, and the authors state they are releasing Waypoint-6M and Waypoint-45M.
Self-supervised pretraining has transformed protein and genomic modeling, but the authors note that it has received limited attention for microbiome data. Most microbiome prediction still relies on classical models, such as random forests and linear models, trained from scratch on bag-of-taxa abundance vectors for each new study.
Waypoint addresses this gap with three connected resources: Atlas, a large pretraining dataset; Waypoint, the models; and Compass, a benchmark of eight downstream tasks for evaluating microbiome foundation models.
Item | Detail from the paper |
|---|---|
Model type | GPT-2-style causal language model, 6M to 170M parameters |
Pretraining data (Atlas) | 539,308 datapoints from MGnify, drawn from over 4,100 studies |
Data modalities | 16S amplicon, shotgun metagenomic, metagenomic assembly, metatranscriptomic |
Vocabulary | 14,394 tokens (14,389 taxonomic plus 5 special) |
Context length | 512 tokens |
Benchmark | Compass, 8 predictive tasks |
How Waypoint Turns Microbiome Profiles into Model Inputs
Genus-level tokenization: each taxon becomes a token at genus level, falling back to the most specific higher rank when genus is unassigned.
Abundance ordering: taxa in each sample are ordered by z-scored relative abundance, and sequences are truncated to 510 tokens plus start and end tokens.
Autoregressive pretraining: the model learns by predicting the next taxon token.
Fine-tuning: a linear head on the end-of-sequence representation is trained for classification or regression, updating all weights.
Waypoint Results on the Compass Microbiome Benchmark
Compass covers biome classification, drug-microbiome interactions, drug degradation and infant gut development.
Finding | Reported result |
|---|---|
Effect of pretraining | Pretrained transformers beat non-pretrained versions and classical baselines |
Data threshold | Transformers pass the random forest baseline from about 10,000 training examples and underperform below about 1,000 |
Model size | Pretrained models improve modestly with size; the 45M model had the highest mean benchmark score |
Versus prior model | 6M-MGM, with larger data and fallback tokenization, outperforms the original MGM |
Where Baselines Still Win
On tasks 4 (drug versus control) and 7 (infant age), baselines that drop unknown taxa outperformed the transformers, and the random forest beat them on some other tasks.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Waypoint on Tamarind Bio
Biome classification: label samples across several levels of the biome hierarchy, including digestive system samples.
Drug-microbiome studies: predict the source community of a drug-perturbed sample, drug versus control, or drug class applied.
Drug degradation: estimate degradation rates of drugs by gut microbial communities.
Infant gut development: predict infant age and delivery mode from microbiome profiles.
How to Use Waypoint on Tamarind Bio
Open the tool: log in to tamarind.bio and open Waypoint.
Provide your profiles: supply per-sample taxonomic names with relative abundances, the input the model is built on.
Choose a model size: select Waypoint-6M or Waypoint-45M if exposed.
Select the task: choose embedding extraction or a fine-tuned prediction head, if exposed.
Run the job: submit it and wait for processing.
Download outputs: retrieve the embeddings or predicted labels and values.
Use downstream: feed the outputs into your own classifiers or statistical analyses.
Things to Keep in Mind
Compass covers gut and environmental tasks only, not oral, skin or respiratory microbiomes.
The fixed 512-token context may truncate samples with very high richness.
Genus-level tokenization may limit taxonomic resolution, and taxa outside the vocabulary are unknown to the model.
Models underperform classical methods when fewer than about 1,000 labeled examples are available.
Only GPT-2-style architectures were tested, and baselines were limited to random forest and linear models.
Source: Treloar NJ, Ur-Rehman S, Yang J. Learning the Language of the Microbiome with Transformers. bioRxiv 2026. DOI: 10.64898/2026.05.02.722381. Code: https://github.com/Outpost-Bio/waypoint