kozo2/metabolomics-edges-expected-ge5

Cross-study metabolomics co-response edges (expected frequency ≥ 5)

620,265 edges over 34,378 nodes, drawn from pairwise metabolite co-response statistics across MetaboLights and Metabolomics Workbench studies, together with the node properties, two PyTorch Geometric graphs, and the full pipeline that produces them.

A node is one differential comparison within one study assay — MTBLS1285_0001_00000028 is study MTBLS1285, assay 0001, feature 00000028; ST002832_AN004625_00002191 is the Metabolomics Workbench equivalent. An edge carries the association between two comparisons, summarised as a 2×2 contingency table of InChIKey counts across the up/down directions of each side, plus the log2 odds ratio and chi-square p-value.

Companion model repository: kozo2/metabolomics-edge-node2vec.

Two filters define this release

1. Cross-study only. Rows whose two endpoint IDs share their first underscore-delimited token (the study accession) are dropped, so every edge joins comparisons from different studies. This reduces 223,871,380 candidate rows to the five *_filtered tables and is the step the source TSVs (not released here) would be needed to re-run.

2. Chi-square validity. With the 2×2 table written [[a,b],[c,d]] = [[uu_n, ud_n],[du_n, dd_n]] and n = a+b+c+d, the four expected frequencies are (a+b)(a+c)/n, (a+b)(b+d)/n, (c+d)(a+c)/n, (c+d)(b+d)/n. The smallest is always min(row total) × min(col total) / n, so the test for all four ≥ 5 collapses to one comparison, evaluated in integer arithmetic so there is no floating-point error at exactly 5:

n > 0 AND least(a+b, c+d) * least(a+c, b+d) >= 5 * n

source_table	rows before	n ≥ 20	released (all E ≥ 5)
edge_ML	6,144,267	75,300	3,697
edge_MLvsMW	71,625,459	2,614,319	89,277
edge_MW_1	62,246,217	3,711,703	294,714
edge_MW_2	4,379,904	0	0
edge_MW_3	79,475,533	4,949,342	232,577
total	223,871,380	11,350,664	620,265

edge_MW_2 contributes nothing: not one of its rows reaches even n = 20, so it cannot produce a valid chi-square table. Its Parquet file is present and empty, and build_graph.py exits with a message if asked to build a graph from it.

What is here

Path	Contents	Size
data/edges_filtered_expected_ge5.parquet	All 620,265 edges, 12 columns + source_table	23 MB
data/edge_*_filtered_expected_ge5.parquet	The same edges split by source table, 12 columns	23 MB
data/nodes_expected_ge5.parquet	34,378 node property rows (one per released node), 12 columns	425 KB
graph/edge_ML_filtered_expected_ge5_pyg.pt	PyG Data: 848 nodes, 3,697 undirected edges	272 KB
graph/edge_MLvsMW_filtered_expected_ge5_pyg.pt	PyG Data: 15,758 nodes, 89,277 undirected edges	6.3 MB
figures/umap_species_*_{light,dark}.png	UMAP of the embeddings, coloured by species	1.6 MB
figures/umap_database_edge_MLvsMW_filtered_{light,dark}.png	The same UMAP coloured by source database	956 KB
figures/umap_coords_*.npy	Cached 2D UMAP coordinates	140 KB
bootstrap.py	Rebuilds edges_filtered.duckdb from the two Parquet files	—
sql/	The DuckDB import and Parquet export statements, as run	—
scripts/	The pipeline: graph build, training, evaluation, figures	—
scripts/umap_app.py	Streamlit explorer for the UMAP, coloured by node property	—
scripts/config.py	Per-source-table paths and node metadata — one place to repoint	—

Edge columns: start_id, end_id, OddsRatio_log2, ChiTestsPValue, {upup,updown,downup,downdown}InChIKeyCount, {upup,updown,downup,downdown}InChIKey (pipe-delimited lists, NULL when empty). Node columns: id, factor, Study, Analytics, factor1, factor2, factor1_org, factor2_org, species, color_group, color_group_upper1, color_group_upper2.

The graphs

Built with to_undirected(..., reduce="mean"), which adds the reverse of every edge and coalesces duplicates. Both edge attributes are invariant under transposing the 2×2 table, so a reverse edge carries identical values. Node indices are assigned lexicographically over the union of both endpoint columns, so the mapping is reproducible; the original string IDs ride along as data.node_id. Edge attributes are float64 because ChiTestsPValue reaches 1.0e-20, below the float32 normal range.

	edge_ML	edge_MLvsMW
nodes	848	15,758
undirected edges	3,697	89,277
edge_index columns	7,394	178,554
self-loops / isolated nodes	0 / 0	0 / 0
structure	MetaboLights only	bipartite: 4,489 MTBLS × 11,269 ST

edge_MLvsMW is bipartite — every start_id is a MetaboLights ID and every end_id a Metabolomics Workbench ID, with no within-database edges. Any method that assumes a homogeneous graph should account for that. edge_MW_1 (294,714 edges) and edge_MW_3 (232,577 edges) are released as Parquet but their graphs are not built here; the same command builds them.

Reproducing it

uv sync                          # Python 3.12, torch 2.14.0+cu130, PyG 2.8.0
uv run python bootstrap.py       # -> edges_filtered.duckdb

cd scripts
SRC=edge_MLvsMW                  # or edge_ML | edge_MW_1 | edge_MW_3
uv run python build_graph.py    --source-table $SRC
uv run python verify_load.py    --source-table $SRC
uv run python node2vec_model.py --source-table $SRC --epochs 200
uv run python heldout_check.py  --source-table $SRC --epochs 200
uv run python species_purity.py --source-table $SRC [--by database]
uv run python umap_species.py   --source-table $SRC [--color-by database] light

Explore the embeddings interactively instead of re-rendering the static figure:

cd scripts
uv run streamlit run umap_app.py            # SOURCE_TABLE=edge_MLvsMW to switch graphs
uv run python verify_app.py                 # the same figure, checked without a browser

The app needs an embedding file from the companion model repository next to the data (<source_table>_filtered_expected_ge5_n2v.pt at the repository root) plus the matching figures/umap_coords_*.npy; both are listed above. Click a point for its 10 nearest embeddings, each flagged for whether it is also a true graph neighbour.

Every script resolves its paths from the repository root, so no DATA_DIR is needed in this layout; set it only to point elsewhere. build_graph.py rebuilds the released .pt files deterministically from the Parquet edges. pyg-lib is required — PyG's Node2Vec needs it for random walks — and is pinned to the data.pyg.org wheel index in pyproject.toml.

Provenance and limits

• The source edge TSVs (~120 GB) and the full node property tables are not included; data/nodes_expected_ge5.parquet covers exactly the released nodes, which is what every script here needs.
• Node properties come from the study metadata as-is. A species value may name a co-culture pair (Eubacterium rectale/Streptococcus salivarius); 128 distinct values appear across the released nodes.
• ChiTestsPValue is the p-value as computed upstream; no multiple-testing correction is applied in this release.