manylatents-omics 0.1.3

  A T G . . C A T        .  . .
  . G C A T . G .   -->  . .. .  -->  λ(·)
   T . A G C . A T        .  . .
     A T G . C A

     m a n y l a t e n t s - o m i c s

        from sequence to manifold

Population genetics, single-cell, and foundation model encoders for manylatents. Extends the core DR framework with biological data types and domain-specific metrics.

Install

uv add manylatents-omics

Optional extras:

uv add "manylatents-omics[popgen]"      # population genetics
uv add "manylatents-omics[singlecell]"  # single-cell (scanpy, anndata)
uv add "manylatents-omics[dogma]"       # protein + RNA encoders (ESM3, Orthrus)

DNA encoder (Evo2) requires a separate venv due to torch version conflicts. See scripts/setup-dna-venv.sh.

Or from the core manylatents repo:

uv sync --extra omics   # installs manylatents-omics as a namespace extension

git clone https://github.com/latent-reasoning-works/manylatents-omics.git
cd manylatents-omics && uv sync

Architecture

manylatents-omics is a namespace extension of manylatents. It lives alongside the core repo and adds domain-specific modules under the manylatents.* namespace via pkgutil.extend_path().

lrw/
├── manylatents/    # core DR engine
├── omics/          # this repo — popgen, singlecell, dogma encoders
└── shop/           # cluster infrastructure

Design decision: The core engine stays domain-agnostic. Each "flavor pack" (omics, vision, etc.) is a separate repo/package that extends the manylatents namespace without polluting the core with domain-specific dependencies. Experiment configs (ClinVar pipelines, fusion sweeps, cluster resource presets) belong in downstream experiment repos, not here — this package ships only instantiation configs that define what encoders, datasets, and algorithms are.

Quick start

Omics configs are auto-discovered when the package is installed. The datasets themselves are not shipped with the package (they're gitignored and not in the wheel) — fetch them first:

# Download the PBMC datasets into the location ${omics_data:} resolves to.
# (Writes to <repo>/data from a checkout, else ~/.cache/manylatents/data.)
python scripts/download_pbmc.py --dataset 3k

python -m manylatents.main --config-name=config \
  experiment=single_algorithm data=pbmc_3k

# Sweep on cluster
python -m manylatents.main -m \
  cluster=tamia resources=gpu \
  data=hgdp,pbmc_10k algorithms/latent=umap,phate

The ${omics_data:} config resolver and download_pbmc.py agree on where data lives. To pin it explicitly, set MANYLATENTS_DATA (both honor it):

export MANYLATENTS_DATA=/path/to/data
python scripts/download_pbmc.py --dataset 3k   # writes $MANYLATENTS_DATA/single_cell/

Modules

popgen — Population genetics via the manifold-genetics CSV pipeline. HGDP+1KGP, UK Biobank, All of Us. Admixture proportions, geographic metadata, QC/relatedness filtering. Requires preprocessing via manifold-genetics (a separate tool, not a Python dependency). Configs: popgen/configs/

• GeographicPreservation — Spearman correlation between haversine and embedding distances
• AdmixturePreservation — Geodesic distance fidelity in admixture simplex vs. latent space

singlecell — AnnData .h5ad loader for scRNA-seq, scATAC-seq, CITE-seq. Ships with PBMC 3k/10k/68k and Embryoid Body. Any .h5ad works via AnnDataset. Configs: singlecell/configs/

dogma — Foundation model encoders for DNA, RNA, and protein sequences. Supports single-modality encoding, multi-layer extraction, and cross-modal fusion. All encoders inherit from FoundationEncoder — lazy model loading, batched encoding with OOM retry, standard fit()/transform() interface. Configs: dogma/configs/

• ESM3 — Protein, 1536-dim, masked mean-pool, true batched forward
• Evo2 — DNA, 1920/4096/8192-dim (1B/7B/40B), multi-layer extraction, 1M bp context
• Orthrus — RNA, 256/512-dim (4-track/6-track), Mamba SSM re-implementation for mamba-ssm 2.x
• AlphaGenome — DNA, 1536/3072-dim (1bp/128bp), JAX-based, regulatory track predictions, chunked encoding

ClinVar pipeline

Reference pipeline for variant-effect analysis via geometric methods. Encodes DNA and protein sequences flanking ClinVar variants, then applies dimensionality reduction to study how pathogenic vs. benign variants separate in embedding space. Three stages: DNA encoding, protein encoding, and geometric analysis (fusion + DR). See docs/clinvar_pipeline.md for full details. Experiment configs live in downstream repos (e.g. merging_dogma), not in this package.

Development

uv sync
pytest tests/ -v

Citing

If manylatents-omics was useful in your research, a citation goes a long way:

@software{manylatents_omics2026,
  title     = {manyLatents-Omics: Biological Extensions for Unified Dimensionality Reduction},
  author    = {{Latent Reasoning Works}},
  year      = {2026},
  url       = {https://github.com/latent-reasoning-works/manylatents-omics},
  license   = {MIT}
}