alrobles/biocellai-foundations

BioCellAI

Foundations — a research toolkit for controlled cell–text representation learning.

BioCellAI investigates whether aligning single-cell expression profiles with biological text improves generalization to unseen donors beyond equivalent non-text supervision — and how much of any improvement survives controls for cell composition, region, and cohort.

The core comparison is deliberately strict: same cells, same labels, same encoder, same training budget — biological text versus non-text supervision of equivalent strength.

This is an independent implementation of the research questions described for CellOLMo (Ai2 × Allen Institute): open language models × single-cell brain data on SEA-AD, evaluated on held-out donors across cell, region, and donor levels. See docs/cellolmo-alignment.md for the artifact-level mapping to that agenda. Not affiliated with Ai2 or the Allen Institute.

Status

Research alpha. The data-preparation and evaluation infrastructure is implemented and tested; the corrected v2 benchmark is still being executed. No release tag exists yet. The repository remains bioAI-devel and the Python package remains bioai; Foundations is the codename for this infrastructure stage, not a publication claim.

Capability	State
Donor-level splits, train-only HVG selection, counts validation	Implemented; tested on synthetic data
Immutable fold builder with manifests, hashes, donor/gene lists	Implemented
Train-only PCA, feature scaling, ridge readouts	Implemented
Paired bootstrap and exact gate utilities	Implemented
One-hot / random-prototype supervision controls	Implemented as components
Integrated v2 runner (all control arms, equivalent budgets)	Pending
Corrected real-cohort benchmarks (blood, SEA-AD, ROSMAP)	Pending
Strict source-fitted cross-cohort transfer	Pending
Blinded biological caption review	Pending

Full task registry and dependencies: spec/revalidation_tasks.json — python scripts/revalidation_status.py prints the live state.

Installation

Requires Python ≥ 3.11. From source:

git clone https://github.com/alrobles/bioAI-devel
cd bioAI-devel
python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev,data,text]"
pytest -q   # 63 tests

Extras: data (scanpy/anndata/cellxgene-census), text (sentence-transformers; downloads encoder weights on first use), dev.

Quickstart: v2 fold preparation

scripts/v2_prepare.py builds immutable donor-held-out folds from a verified count matrix. With any AnnData .h5ad whose X (or a named layer) contains raw counts and obs["donor_id"] identifies donors:

python scripts/v2_prepare.py \
    --input cohort.h5ad --counts-layer X \
    --cohort seaad --source-release "SEA-AD MTG 2024" \
    --seeds 0 1 2 --out experiments/v2_revalidation/seaad_run1

Each run writes fold_s*.h5ad plus per-fold JSON manifests (source hash, environment versions, code hashes, donor/gene lists, matrix hash) and refuses to overwrite an existing --out directory. Prepared ≠ evaluated: the PREPARED.json marker records scientific_results_complete: false.

Synthetic smoke test without real data:

python - <<'PY'
import numpy as np, pandas as pd
from anndata import AnnData
rng = np.random.default_rng(0)
AnnData(
    X=rng.poisson(3.0, (80, 400)).astype(np.float32),
    obs=pd.DataFrame({"donor_id": np.repeat([f"d{i}" for i in range(8)], 10)}),
    var=pd.DataFrame(index=[f"g{i}" for i in range(400)]),
).write_h5ad("/tmp/synth.h5ad")
PY
python scripts/v2_prepare.py --input /tmp/synth.h5ad --counts-layer X \
    --cohort synthetic --source-release local --seeds 0 --out /tmp/v2_folds

A historical end-to-end pipeline check exists (python -m bioai.experiment m1 --dataset pbmc3k), but its PBMC donor IDs are randomly assigned pseudo-donors: it exercises code paths only and demonstrates nothing about generalization.

Input requirements

• Counts: X or a named layer must contain nonnegative, integer-valued counts. Numerical checks do not establish provenance — you must identify which layer holds verified raw counts. Preprocessed inputs (existing log1p state) are rejected.
• Donors: obs["donor_id"] is required; splits are donor-level and unified across seeds (split_donor_ids).
• Genes: unique identifiers; symbol mapping via data/gene_aliases.json.
• HVGs are selected on train donors only; normalization happens before the gene subset; the fold records bioai_data_state and protocol metadata in uns.

Workflow

verified counts → v2_prepare (frozen folds + manifests)
               → v2 runner (text arms + equivalent-supervision controls)  [pending]
               → inductive readouts (train-fit PCA/scaler/ridge)
               → paired comparisons, permutations, exact gates

All scientific outputs go under experiments/v2_revalidation/<run-id>/. Historical experiments/m* directories are preserved read-only in spirit: the tooling refuses to overwrite prepared outputs, and historical runners now reject renormalizing processed files.

Scientific scope and limitations

• Grounding where a caption is assigned to a cell by its label is text-mediated supervision, not label-free learning — even when the text is external knowledge or omits the class name. The interesting question is what biological text adds beyond equivalent supervision; that is what the pending control arms measure.
• Historical M0–M12 results are exploratory. The audit found repeated preprocessing, transductive components, and non-equivalent comparisons; see docs/milestones.md for the full log and caveats.
• Nothing here is a claim that text helps, that composition explains the pathology axis, or that encoder scale is irrelevant — those are open questions the v2 matrix is designed to answer.
• Cross-sectional postmortem pseudoprogression scores are not longitudinal trajectories.

Data and artifacts

• Included: dataset manifests, gene aliases, retrieval corpora (data/retrieval/), captions/provenance (data/text/), specs, ADRs, code, tests, and historical milestone reports.
• Not included: raw single-cell datasets (Tabula Sapiens, SEA-AD, ROSMAP), model checkpoints, or large embeddings — these live on KU HPC and are subject to their own access terms. No large data or weights are stored in git.
• The m1 pipeline check downloads pbmc3k and a text encoder on first run.

Documentation

• spec/00-spec.md — project specification
• spec/adr/ — architecture decision records
• spec/adr/ADR-011-scientific-revalidation.md — v2 protocol
• spec/revalidation_tasks.json — task registry (scripts/revalidation_status.py)
• docs/milestones.md — historical M0–M13 log
• docs/cellolmo-alignment.md — mapping to the CellOLMo research agenda
• docs/historical-results.md — honest review of exploratory results
• docs/release-plan.md — weights/data release plan and license stack
• paper/ — manuscript (provisional; full rewrite pending v2 evidence)
• scripts/slurm/ — version-controlled KU HPC jobs

License and citation

Code and released model weights are Apache-2.0 (see LICENSE); project-generated captions, retrieval corpora, and manifests are CC-BY-4.0 (see NOTICE). Third-party datasets keep their own terms — controlled-access data are not redistributed. Per-artifact release permissions are tracked in R7-LICENSE and docs/release-plan.md. If you use this repository in work leading to a publication, cite the GitHub repo and note the Foundations codename.

Credits / context

Developed within the alrobles ecosystem: KU HPC, Apptainer, Ollama, PubMed indexing (genominer), and spec-driven development under spec/.