Waypoint Bio

Waypoint: Outpost Bio's Open Microbiome Foundation Models

Overview

Outpost Bio open-sourced three artefacts under Apache 2.0, described in Treloar et al., bioRxiv 2026.05.02.722381:

The unifying idea: a microbiome sample is a sentence. Each taxon is one token, tokens are ordered by descending abundance z-score, and the model is trained with next-token prediction. A pretrained checkpoint then supplies sample-level embeddings or a fine-tuning backbone for prediction tasks.

All of it is driven by one CLI, waypoint, with five subcommands: prepare-dataset, embed, finetune, benchmark, pretrain.

When to use

• Embedding 16S/shotgun taxonomic profiles into fixed-size vectors for clustering, visualisation, or a downstream classifier.
• Fine-tuning a Waypoint checkpoint to predict a phenotype, treatment, or continuous readout from community composition.
• Scoring your own microbiome model on Compass with an explicitly recorded protocol.
• Pretraining a taxonomic language model on Atlas or on your own corpus.
• Converting profiler output (MetaPhlAn, Kraken2/Bracken, QIIME 2, MGnify TSVs) into the input format these tools expect.

For small labelled datasets, start with a random-forest baseline and grouped validation. The paper's sample-size crossover is an empirical result from its experiments, not a universal cutoff for using embeddings or a guarantee of performance on a new study.

Setup

pip install "waypoint-bio==1.0.2" "transformers==4.57.6" "peft==0.18.1" "huggingface-hub<1"

The 2026-10-01 review checked the published wheel, GitHub main f45eee6d07a480bfc90f84ab8082bbc969dec15d (1.0.4), and current Hub metadata. Small native CPU/tokenizer tests used Python 3.12, Torch 2.14.1, Transformers 4.57.6, PEFT 0.18.1 and pandas 3.0.6; no pretrained weights or gated rows were downloaded. The package leaves dependencies unbounded: Transformers 5 removes its pretraining logging_dir argument. Keep a separate compatible environment.

Upstream limitations: local TaxonomicTokenizer.save_pretrained() raises NotImplementedError in both reviewed versions, blocking pretrain before training and finetune export when using that local tokenizer. Released 1.0.2 does not merge LoRA adapters or write training logs; GitHub 1.0.4 does. See references/upstream-review.md before training. Training examples below are source-checked templates, not completed scientific runs.

Atlas, Compass, and every Waypoint checkpoint are gated. Access is auto-approved, but you must click through once per repo and then authenticate:

1. Request access on each repo page you need: Waypoint-6m, Waypoint-45m, Waypoint-170m, Atlas, Compass.

2. Authenticate locally:

   hf auth login          # or: export HF_TOKEN=hf_...

For 401/403 errors, check token validity, read scope, and approval for the specific repo. The tokenizer loader executes repository code with trust_remote_code=True. Review that code and pin an immutable Hub commit. The upstream CLI has no --revision; download a reviewed snapshot and pass its local path so model, tokenizer and ordering statistics share one revision (see references/python-api.md).

The waypoint data format

Everything except prepare-dataset consumes waypoint format: a .parquet / .csv / .tsv whose rows are samples, with two aligned list-columns plus any label columns you need.

Use parquet to preserve list values and sample IDs. Upstream CSV/TSV loading loses the sample-ID index and, with pandas 3, does not parse string lists. The bundled coverage reader handles the converter's CSV/TSV, but that does not repair the upstream loader.

Give full lineages, not bare names. The tokenizer extracts the genus segment (g__) from each lineage and falls back to the most specific higher rank when genus is missing. Bare names disable that fallback entirely.

Workflow

1. Get your data into waypoint format

If you already have a sample × taxa (or taxa × sample) abundance matrix with lineage labels:

waypoint prepare-dataset \
    --input abundance_matrix.tsv \
    --metadata sample_labels.csv \
    --output dataset.parquet

Orientation is auto-detected from the first column header (taxonomy, lineage, taxon, otu, #otu id ⇒ taxa-as-rows); override with --orientation. Rows are normalised to sum to 1 unless you pass --no_normalize, and zeros are dropped unless you pass --keep_zeros.

prepare-dataset cannot read profiler output directly — MetaPhlAn uses | separators, Kraken2 reports encode the hierarchy as indentation, and QIIME 2/SILVA prefixes the domain d__ instead of k__ (which the tokenizer silently ignores). Use the bundled converter for those:

python scripts/profiler_to_waypoint.py \
    --input merged_metaphlan.tsv --format metaphlan \
    --output dataset.parquet

python scripts/profiler_to_waypoint.py \
    --input reports/*.kreport --format kraken \
    --output dataset.parquet

python scripts/profiler_to_waypoint.py \
    --input feature-table.tsv --format qiime2 --taxonomy-column taxonomy \
    --output dataset.parquet

See references/data-preparation.md for every input layout, rank handling, and the d__/| gotchas.

2. Check vocabulary coverage before anything else

Waypoint's vocabulary is fixed at pretraining time from Atlas. Taxa absent from it become <unk> and are silently dropped by waypoint embed; the paper names this as the models' main limitation. A sample whose taxa are all out-of-vocabulary yields a degenerate [BOS][EOS] embedding.

python scripts/vocab_coverage.py --model outpost-bio/Waypoint-6m --data dataset.parquet

It reports per-sample and abundance-weighted coverage and flags samples below a threshold. Treat the script's default 0.8 threshold as a heuristic, not a validated biological quality cutoff. Low abundance-weighted coverage is a reason to re-examine your taxonomy labels before trusting any downstream number.

3. Embed samples

waypoint embed \
    --model outpost-bio/Waypoint-6m \
    --data dataset.parquet \
    --output embeddings.parquet

Output is indexed by sample ID with columns dim_0 … dim_{H-1} (H = 256 for 6m, 512 for 45m, 768 for 170m). Defaults: --pooling last_token, --batch_size 32, --max_length 512, device auto-detected (cuda → mps → cpu).

Record the fraction of samples truncated at the chosen max_length, separately from vocabulary coverage. Samples can have excellent vocabulary coverage and still lose lower-ranked taxa after abundance/z-score sorting. Keep this limit consistent across embedding comparisons and report any sensitivity analysis.

Use --pooling last_token to match the supervised benchmark and finetune default. Pretraining itself uses a next-token loss, without a sample-level pooling objective. mean is a reasonable alternative for unsupervised use; first_token/cls_token return the BOS position and carry little signal in a causal LM.

4. Fine-tune on your labels

# classification
waypoint finetune \
    --model outpost-bio/Waypoint-45m \
    --data dataset.parquet \
    --output_dir outputs/ft_disease \
    --task_type classification \
    --target "Disease Status" \
    --config configs/finetune_classification.yaml

# regression, with a categorical covariate one-hot appended to the pooled embedding
waypoint finetune \
    --model outpost-bio/Waypoint-45m \
    --data dataset.parquet \
    --output_dir outputs/ft_degradation \
    --task_type regression \
    --target "Degradation Rate" \
    --covariate_column Drug \
    --config configs/finetune_regression.yaml

Config paths resolve against the bundled waypoint_bio/configs/ tree, so configs/... works from any directory without cloning.

For small datasets, choose warmup and evaluation intervals that actually fit the number of optimizer steps, and enough epochs for validation and early stopping. The shipped one-epoch benchmark config differs from the paper's up-to-300-epoch protocol. On PyPI 1.0.2, keep use_lora: false for export compatible with plain AutoModel; automatic adapter merging is only in GitHub 1.0.4. LoRA reduces trainable parameters but still needs the base model.

Splits default to a random 80/10/10. Set split_column to a Split column whenever samples are correlated — repeated measures, one donor sampled over time, technical replicates — or a random split leaks and the test score is meaningless.

Outputs land in --output_dir: best_model/ (loadable by embed/benchmark), test_metrics.json and finetune_results.json after successful serialization. Training logs (training_log.csv + .html) and prediction CSVs are GitHub 1.0.4 additions.