README.md

---
license: cc-by-4.0
tags:
  - single-cell
  - flow-cytometry
  - spectral-flow-cytometry
  - haematopoiesis
  - experimental-design
size_categories:
  - 10M<n<100M
---

# LabCompass — Spectral Flow Cytometry haematopoiesis dataset

Measurements underlying **LabCompass**, a method for generative modeling of experimental design in
single-cell data. This dataset contains Spectral Flow Cytometry (SFC) profiles of *in vitro*
haematopoietic differentiation cultures, collected over successive rounds of a closed-loop
experimental design cycle.

Each round — a **loop** — proposes new culture protocols, runs them at the bench, and measures the
resulting cells. The measurements from each loop are published here as a separate file.

- **Code and full reproduction pipeline:** <https://github.com/theislab/LabCompass>
- **Contents:** per-loop measurements in `loops/`, trained models in `checkpoints/`, designed protocols in `solutions/`
- **Wet-lab experiments and measurements:** Göttgens Lab
- **License:** CC-BY-4.0

## ⚠️ These files are per-loop, not cumulative

`loops/loop3.h5ad` contains **only the cells measured in loop 3** — not loops 0–3 together. Models in
the paper are trained on the *accumulated* data, so a loop's training set is the concatenation of
every loop up to and including it:

```
dataset(N) = concat(dataset(N-1), loopN)
```

Concatenating them yourself is a few lines of `anndata`, but the exact chain matters (one loop
introduces new protocol axes that must be zero-filled on the earlier data — see below). The
reproduction repository ships a script that does it correctly:

```bash
git clone https://github.com/theislab/LabCompass.git
python scripts/data/build_loop_datasets.py              # downloads from this repo and builds the chain
python scripts/data/build_loop_datasets.py --variants 500k   # subsampled only: far smaller and faster
```

## Files

Every loop is published in two variants: the full measurement set, and a subsampled version
(`_500k` suffix) intended for fast iteration. The suffix is a naming convention carried over from
the source data, not a guaranteed cell count — the subsampled files vary in size.

| Loop | Full | Subsampled | Approx. size (full) |
| --- | --- | --- | --- |
| 0 (baseline) | `loops/loop0.h5ad` | `loops/loop0_500k.h5ad` | 36 GB |
| 1 | `loops/loop1.h5ad` | `loops/loop1_500k.h5ad` | 2.5 GB |
| 2 | `loops/loop2.h5ad` | `loops/loop2_500k.h5ad` | 3.9 GB |
| 2.5 | `loops/loop2p5.h5ad` | `loops/loop2p5_500k.h5ad` | 2.7 GB |
| 3 | `loops/loop3.h5ad` | `loops/loop3_500k.h5ad` | 6.3 GB |
| 4 | `loops/loop4.h5ad` | `loops/loop4_500k.h5ad` | 0.9 GB |
| 4.5 | `loops/loop4p5.h5ad` | `loops/loop4p5_500k.h5ad` | 0.5 GB |
| 5 | `loops/loop5.h5ad` | `loops/loop5_500k.h5ad` | 6.3 GB |

Loop 0 is the baseline screen and is by far the largest. The half-steps (2.5, 4.5) are follow-up
rounds within a design cycle and accumulate like any other loop, giving the chain

```
loop0 → loop1 → loop2 → loop2p5 → loop3 → loop4 → loop4p5 → loop5
```

The full set is roughly 60 GB; the subsampled set is a few GB.

## Format

Each file is an [AnnData](https://anndata.readthedocs.io/) `.h5ad` object:

- **`X`** — logicle-transformed SFC intensities: fluorescence channels and morphological scatter
  features, one row per cell.
- **`obs`** — per-cell metadata, in three groups:
  - *Acquisition:* `experiment_number`, `experiment_id`, `replicate`, `date`, `well_id`,
    `cytometer`, `cytometer_serial_no`, `count_beads`, `cell_counts`, `source_id`.
  - *Protocol axes* — the culture recipe, and the space LabCompass searches over. Cytokines and small
    molecules carry their units in the column name, e.g. `scf_[ng_ml]`, `tpo_[ng_ml]`,
    `il3_[ng_ml]`, `gm-csf_[ng_ml]`, `rhflt3l_[ng_ml]`, `ldl_[ng_ml]`, `sr1_[nm]`, `um171_[nm]`,
    `um729_[µm]`, `butyzamide_[nm]`, `retinoic_acid_[µm]`, `mtg_[µm]`, `740-yp_[µm]`, alongside
    culture conditions such as `o2_[%]` and `hydrogel_type`.
  - *Annotation:* cell-type labels, where available.

`experiment_number` identifies the physical experiment a cell came from (loop 1, for instance, spans
experiments 206–210), which makes it a convenient way to check which loops are present in a
concatenated object.

### The protocol schema grows across loops

Later loops vary axes that earlier loops never did. Loop 3 introduces `il7_[ng_ml]`,
`mcsf_[ng_ml]` and `ly_cocktail_[ul/well]`, which are absent from loops 0–2.5. When concatenating,
these must be **zero-filled on the earlier data** (they were held at zero, not missing) so both sides
share an `obs` schema. `build_loop_datasets.py` does this; a naive `anndata.concat` will silently
drop the columns instead.

## Loading

```python
import anndata as ad
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="theislab/LabCompass",
    filename="loops/loop3_500k.h5ad",
    repo_type="dataset",
)
adata = ad.read_h5ad(path)
```

## Model checkpoints

`checkpoints/` holds the trained models behind the paper's designs, one folder per loop:

| Folder | Size | Forward model | Config group |
| --- | --- | --- | --- |
| `checkpoints/loop0/` | 1.7 GB | `likely-donkey-20` | `paths=loop0` |
| `checkpoints/loop1/` | 1.9 GB | `eager-feather-1` | `paths=loop1` |
| `checkpoints/loop2/` | 2.1 GB | `fresh-bee-21` | `paths=loop2_replicate` |
| `checkpoints/loop2p5/` | 2.3 GB | `celestial-fire-40` | `paths=loop2p5_replicate` |
| `checkpoints/loop3/` | 2.6 GB | `rich-sunset-44` | `paths=loop3_replicate` |
| `checkpoints/loop4/` | 2.6 GB | `fast-gorge-13` | `paths=loop4_replicate` |

**Loop *N*'s models are the ones that generated the designs executed in loop *N+1*.** So to reproduce
the candidates that were run at the bench in loop 1, use `checkpoints/loop0/`.

Each folder contains the four models the pipeline needs:

- **`*_FlowMatching.pkl`** — the forward model: predicts the cell-state distribution a protocol induces.
- **`*_TargetPredictionModel.pkl`** — the cell-type classifier, i.e. the phenotypic readout that the
  inverse objective is defined against.
- **`*_FlowMatchingWithScore.pkl`** — the generative prior over protocols.
- **`*_FlowMap.pkl`** — a distilled few-step version of that prior.

File names are the Weights & Biases run names, unchanged from training. `checkpoints/manifest.json`
records, for every file, which `paths` config group refers to it, its original size, and a sha256.

### Three things to know before using them

**They are inference-only.** The training and validation data that the original checkpoints carried
inside them has been removed — that is why a 113 GB file is 1 GB here. Everything inference touches
is intact (network weights bitwise unchanged, normalisation parameters, cell-type labels, the
condition key), and every checkpoint was verified tensor-by-tensor against its original and exercised
end-to-end through the pipeline. But you **cannot retrain a prior from these**: the scripts that do
so read the forward model's embedded training set, which is gone. Retrain the forward model from the
`loops/` data instead.

**They are CPU-resident.** The tensors load on any machine, with or without a GPU; move the model to
your device as you would any PyTorch module. The original checkpoints held CUDA tensors and could
only be loaded on a GPU node.

**Loops 3 and 4 use the expanded design space.** They were trained after M-CSF and the lymphoid
cocktail were added, so they expect the wider protocol vector and will fail with a shape mismatch if
you load them with the earlier annotation. Use `annotation=bloodplus_loop3` for those two; the
earlier loops use the default.

### Loading

```python
from huggingface_hub import hf_hub_download
from labcompass.models import FlowMatching

path = hf_hub_download(
    repo_id="theislab/LabCompass",
    filename="checkpoints/loop3/rich-sunset-44_FlowMatching.pkl",
    repo_type="dataset",
)
model = FlowMatching.load(path)
```

In the reproduction repository these are wired up through the `paths` config group, so pointing a run
at a downloaded loop is a matter of overriding the four checkpoint paths.

## Designed protocols (`solutions/`)

`solutions/` contains **every candidate protocol LabCompass generated**, across all sweeps behind the
paper — about 1.07 million designs. The raw output is a tree of ~16,000 run directories; each loop is
flattened here into a single gzipped CSV, with the directory structure turned into columns.

| File | Designs | Runs | Size |
| --- | --- | --- | --- |
| `solutions/loop0.csv.gz` | 74,400 | 744 | 35 MB |
| `solutions/loop1.csv.gz` | 373,900 | 3,111 | 169 MB |
| `solutions/loop2.csv.gz` | 69,012 | 1,386 | 32 MB |
| `solutions/loop2p5.csv.gz` | 152,050 | 3,041 | 53 MB |
| `solutions/loop3.csv.gz` | 195,000 | 3,900 | 86 MB |
| `solutions/loop4.csv.gz` | 208,700 | 4,174 | 91 MB |

Each row is one designed protocol. Column counts differ between loops (170–187) because the design
space and the cell-type panel both grew over the campaign, which is why these are six files rather
than one.

### Columns

- **Provenance** — `loop`, `experiment_type` (the optimisation variant and guidance schedule, e.g.
  `penalized_all_axes-pure_populations-constant`), `cell_type` (the target the run optimised for),
  `run_id`, and `uncertainty_scoring`.
- **The design** — one column per protocol axis (`tpo_[ng_ml]`, `um171_[nm]`, `o2_[%]`,
  `days_of_culture`, …), plus `:rescaled` variants.
- **Predicted outcome** — `<cell_type>_prop` for every cell type in the panel, and `loss`.
- **Uncertainty** — `<cell_type>_prop_std`, `target_ct_loss_mean`, `target_ct_loss_std`,
  `ct_prop_total_variance`.
- **Configuration** — ~96 `cfg:*` columns recording the resolved hydra config for that run, so every
  design can be traced back to exactly how it was produced.

### These are unfiltered

Nothing here has been filtered or ranked. The paper's analysis applies thresholds *downstream* —
minimum predicted enrichment, oxygenation and culture-duration bounds, and a margin on the measured
design range — and those thresholds differ per loop and per target cell type. Publishing the full
search record keeps that choice in the reader's hands, and preserves the configurations that did not
work alongside those that did.

The `uncertainty_scoring` column says how each row was scored:

- `same_loop` — scored under that loop's own forward model, the usual case.
- `next_loop` — the same candidates re-scored under a *later* loop's model. This is what shows
  predictive uncertainty falling as data accumulates; present for loops 0 and 1.
- `none` — uncertainty estimation never ran for that run, so the `_std` columns are empty. This
  affects **roughly half of loop 2.5** (71,450 of 152,050 rows); the designs and their predicted
  proportions are still there.

### Not included

The per-run `.npz` files holding the guidance trajectories and per-candidate forward samples are not
published — roughly 157 GB, around thirty times everything else, and needed only for trajectory and
sensitivity plots. They can be regenerated from the published `checkpoints/`.

### Loading

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="theislab/LabCompass",
    filename="solutions/loop3.csv.gz",
    repo_type="dataset",
)
designs = pd.read_csv(path)

# e.g. the most promising MgkPro designs that carry a scored uncertainty
mgk = designs[
    (designs["cell_type"] == "late_MgkPro")
    & (designs["uncertainty_scoring"] == "same_loop")
].nlargest(20, "late_MgkPro_prop")
```

## Citation

<!-- TODO: replace with the published reference before release. -->

```bibtex
@article{labcompass,
  title   = {TODO},
  author  = {Consoli, Lorenzo and Palma, Alessandro and others},
  journal = {TODO},
  year    = {TODO},
}
```