btraven/clust-libreville-graphs

Hao 2021 PBMC kNN graphs

The graphs behind the Validation section of clust-libreville-rs (docs/design.md), a deterministic parallel Leiden. Four symmetric, weighted kNN graphs (scanpy-style UMAP connectivities) of the Hao et al. 2021 multimodal PBMC data, at 10k, 50k, 100k and 154k cells. bench/validate/validate.py in that repository recomputes every number of the section from these files.

file	cells	stored entries	sha256
hao2021_10k.npz	9,999	213,358	2db665667c38b66b20bc6abc0a5d3faf35c55fd61aecee6deda4ce937de9ca10
hao2021_50k.npz	50,001	1,109,160	20c262b30546380689d978b53d17fe44fe92cfaac9e260aacf2b9f807d7e1ce0
hao2021_100k.npz	100,000	2,250,304	bfa7c3de7e122999d861ae4efe84284a241bcf05045f5189ca7c33192398d826
hao2021_154k.npz	154,283	3,502,318	d0e16995aebeb58d21aea6999c737b281a10ccc1a3de53c9ba127d6ac2968a8d

sha256sum -c SHA256SUMS checks them.

Format

NumPy .npz, one CSR matrix per file, both directions of every edge stored, no diagonal:

• indptr (int64), indices (int32), data (float64; the values are exactly float32), shape (int64, [n, n]);
• cell_ids (str): the cell barcodes, in row order.

import numpy as np, scipy.sparse as sp
z = np.load("hao2021_10k.npz")
A = sp.csr_matrix((z["data"], z["indices"], z["indptr"]), shape=tuple(z["shape"]))

How they were made

All four start from CELLxGENE dataset 4078abd1 (Hao et al. 2021), filtered, normalised (log1p CP10k) and reduced to 2,000 genes by the pinned Omnibenchmark plan of sc-brrr (prep/benchmark.yaml).

• 10k, 50k, 100k: the cells of btraven/sc-brrr-hao2021 (nested, stratified subsets). sc-brrr's exact-kNN reference (baselines/reference.py): scanpy PCA (50 components, arpack, float64, random_state 0), exact kNN (sc.pp.neighbors, 15 neighbours, sklearn brute force), UMAP connectivities.
• 154k: every cell of the filtered dataset (the superset of the three above). PCA (50 components, float32 covariance) and exact kNN (faiss flat, 15 neighbours) as in brrr-gve's method.py, UMAP connectivities in float64 (clust-libreville-rs bench/py/build_graph.py).

Licence and attribution

CC BY 4.0, as the source data. Please cite the original study:

Hao Y, Hao S, Andersen-Nissen E, et al. Integrated analysis of multimodal single-cell data. Cell. 2021;184(13):3573–3587.e29. doi:10.1016/j.cell.2021.04.048.