StripeSankey 0.4.0

---

StripeSankey fits a mixed membership model at several k (a sweep) and draws all of it as one figure. Every ribbon entering or leaving a node sums to exactly that node, and a ribbon's thickness is a number of samples you can read off.

The pipeline supports LDA and NMF, and takes three tables: a count matrix, a per-feature annotation (like taxa in microbiome) and per-sample metadata.

table	index	columns
counts	sample id	feature id
feature_annotation	feature id	annotation levels
metadata	sample id	sample variables

The pipeline also lets users bring their own sweep, computed elsewhere (see Fit your customized models).

pip install stripesankey

Fit models with StripeSankey

1 — prepare

prepare checks that the three tables line up (matching sample and feature ids, usable counts), collapses the counts to one annotation level if you give one, and writes the result to base/intermediate/ for sweep to read.

import stripesankey as ss
base = 'LOCAL FOLDER PATH TO SAVE CACHE'
# fit LDA/NMF models via StripeSankey, example of microbiome count table
ss.prepare(base, counts, feature_annotation, metadata, level="genus")

level is any column of feature_annotation. For instance, in microbiome or ecology, you may want to group your count table at genus level.

2 — sweep

One fit per k, each saved to disk as it finishes.

ss.sweep(base, k_range=range(2, 11), implementation="nmf", random_state=42)

Implementations include:

Algorithm	implementation	source lib	API
LDA	Variational Bayes	gensim	lda_gensim
LDA	Markov Chain Monte Carlo	MALLET	lda_mallet
NMF	generalized KL loss, initialised with nndsvdar	scikit-learn	nmf

lda_mallet takes one more argument, mallet: the path to MALLET's bin/mallet launcher. MALLET is a Java program and is not bundled with StripeSankey, so download it yourself and make sure Java is availble.

ss.sweep(base, k_range=range(2, 11), implementation="lda_mallet",
         mallet="/path/to/Mallet", random_state=42)

Expect lda_mallet to be much slower than lda_gensim and nmf, for two reasons:

1. MALLET fits LDA by Markov chain Monte Carlo (collapsed Gibbs sampling)
2. Convergence has to be watched.

After fitting, results are saved in the base folder with the structure below:

base/
├── intermediate/                       # written by prepare
│   ├── features.csv                    
│   ├── feature_annotation.csv
│   ├── manifest.json
│   └── mallet/corpus.txt               # the count table as MALLET's input
└── results/                            # written by sweep
    ├── W/W_{k}.csv                     # one file per k
    ├── H/H_{k}.csv                     # one file per k
    ├── diagnostics/topic_diff_{k}.csv  # LDA (gensim) only
    ├── diagnostics/ll_{k}.csv          # LDA (MALLET) only
    ├── consensus.json                  # consensus cluster of each sample
    ├── run_info.json                   # k list, implementation, convergence per k
    └── metrics.csv

	shape	on disk
W	sample × LC	results/W/W_{k}.csv
H	LC × feature	results/H/H_{k}.csv

Fit your customized models

Bring your own model via python

Define a function that takes k and a seed and returns an unfitted model, then pass it as implementation. sweep fits it at every k and saves W, H and the consensus, as it does for the built-in implementations:

from sklearn.decomposition import NMF

def my_nmf(k, seed):
    return NMF(n_components=k, beta_loss="kullback-leibler", solver="mu",
               init="nndsvdar", max_iter=3000, random_state=seed)

ss.sweep(base, k_range=range(2, 11), implementation=my_nmf, random_state=42)

The model needs scikit-learn's interface: fit_transform(X) returns W (sample × LC) and components_ holds H (LC × feature).

Bring your own sweep via other programming languages

If your model runs in another language, such as R, or in a well-established tool like ADMIXTURE, run the sweep there and save each W as base/results/W/W_{k}.csv: the sample id in the first column, then one column per LC, rows summing to 1, and the same samples in the same order at every k. Then run sweep without k_range:

ss.sweep(base)

example/population does this with pong's precomputed ADMIXTURE results for the 1000 Genomes (K = 2…8).

Plotting

Both figures read the sweep from base, wherever it was fitted.

stripesankey — coloured by consensus cluster

fig = ss.stripesankey(base, metadata_df=metadata, hover_field=["group"])

An interactive version is available in example/population/viz/stripesankey_K2-8.html.

Click a band, a ribbon or a cluster chip to trace those samples across every k. Click a K = label to focus that column — it restacks by the first hover_field. Read the selection back from fig.selected_samples.

metasankey — the same figure, cut by metadata

fig = ss.metasankey(base, metadata_df=metadata, hover_field=["group"])

An interactive version is available in example/population/viz/metasankey_K2-8.html.

Identical layout, one thing changed: a node is sliced into a metadata field's categories instead of into clusters.

affinity heatmap — the matrix the clusters came from

Q = ss.consensus_affinity(ss.load_W(base))                                         # the affinity matrix
fig, order, labels = ss.affinity_heatmap(Q, above=metadata, annotations=["group"])  # the plot

Q is the signed, chance-corrected co-membership: positive means two samples share components more than chance, negative less. Samples are ordered by consensus cluster by default.

Hard labels

For an ensemble of hard labels instead of soft assignments — k-means at several k, or different algorithms at one k:

fig = ss.hardstripe(labels_df, colorby="consensus", metadata_df=metadata)

Citation

License

MIT