cpss

ElasticNet stability selection with the Shah & Samworth (2013) r-concave PFER bound, for protein-level biomarker discovery from DIA-NN output.

The selection-probability threshold π is derived per-run from the r-concave tail bound (S&S 2013, equation 8) using the data-dependent quantity θ = q/p. This is tighter than the Meinshausen & Buhlmann (2010) worst-case bound and so retains more features for the same PFER budget. The package operates on protein-level DIA-NN matrices (pg_matrix); peptide-level input is rejected explicitly.

---

Installation

Requires Python ≥ 3.10.

pip install cpss

To also install test dependencies:

pip install "cpss[test]"

To install the latest development version directly from GitHub:

pip install git+https://github.com/byrnedaniel5-eng/cpss.git

Or for an editable install when developing the package itself:

git clone https://github.com/byrnedaniel5-eng/cpss.git
pip install -e ./cpss

Verify the install:

python -c "import cpss; print(cpss.__version__)"
# 1.0.0

Dependencies

Pulled in automatically by pip install: numpy, pandas, scipy, scikit-learn, joblib, matplotlib, seaborn, pyyaml, openpyxl.

---

Input data

Two files are required:

1. Protein-level abundance matrix (pg_matrix)

A CSV produced by DIA-NN. The first 6 columns are protein-level metadata (Protein.Group, Protein.Ids, Protein.Names, Genes, First.Protein.Description, N.Proteotypic.Sequences); every column from index 6 onwards is a sample, named by the run/biobank ID.

Peptide-level matrices (pr_matrix, detectable by a Stripped.Sequence column) are rejected at load time — see the docstring on io.py for why.

2. Sample-label spreadsheet

An .xlsx file with at least:

• a column matching the sample IDs used as column headers in the pg_matrix (e.g. Biobank Number)
• a group column with two values matching case_name and control_name from the config (e.g. Group containing PDAC / Non-cancer)

Samples present in the pg_matrix but missing from the label file are dropped with a warning.

---

Configuration

Runs are driven by a YAML config that you supply per-analysis. The fastest way to start is to drop a fully-commented template into your analysis directory:

python -m cpss init

This writes config_cpss.yaml next to where you ran the command. Edit at minimum data_file, label_file, case_name, control_name, and output_dir. The template is also browsable at examples/config_template.yaml.

The most important keys:

Key	Meaning
data_file, label_file	Paths to the two input files
case_name, control_name	Labels in the group column to map to y=1 / y=0
group_column, sample_id_column	Column names in the label file
min_proteotypic_peptides	Drops proteins with N.Proteotypic.Sequences below this (default 2)
max_missing	Per-protein NaN fraction allowed after sample alignment
imputation_strategy	knn, minimum, mean, or median
log_transform	Apply log2(x+1) before scaling
max_candidates (q)	Number of top-
stability_iterations	2 × B; S&S §3.4.1 recommends B ≤ 50 so leave at 100 unless you understand the r-concavity assumption
stability_pfer	Absolute PFER budget E[V]
l1_ratio	ElasticNet mixing parameter (1 = lasso, 0 = ridge)
output_dir	Results destination

The threshold π is not specified directly; it is derived per-run from (q, p, B, PFER) via the r-concave bound after preprocessing fixes p.

---

Running on real data

From the directory containing your config (and data files referenced by it):

python -m cpss --config config_cpss.yaml

This prints class balance, q/p, B, the r-concave vs M&B threshold comparison, calibrated C_ref, the top-20 stable features by selection probability π̂, and writes results to output_dir. Relative paths in the config are resolved from the current working directory.

---

Toy example with synthetic data

Use this to sanity-check the install on a tiny fabricated dataset (no DIA-NN run required). It produces a pg_matrix-shaped CSV, an Excel label file, and a config, then runs the pipeline.

Save as toy_run.py next to the package and run with python toy_run.py:

"""Minimal end-to-end smoke test for cpss on synthetic data."""
import numpy as np
import pandas as pd
import yaml
from pathlib import Path

rng = np.random.default_rng(0)
n_samples_per_group = 20
n_proteins = 200
n_true_signal = 10  # proteins with a real case/control difference

work = Path("toy_workspace")
work.mkdir(exist_ok=True)

# 1. Build a synthetic pg_matrix-shaped CSV.
sample_ids = [f"S{i:03d}" for i in range(2 * n_samples_per_group)]
y = np.array([0] * n_samples_per_group + [1] * n_samples_per_group)

X = rng.normal(loc=20.0, scale=1.0, size=(n_proteins, len(sample_ids)))
# Inject a real signal into the first n_true_signal proteins (up in cases).
X[:n_true_signal, y == 1] += rng.normal(loc=2.0, scale=0.3,
                                        size=(n_true_signal, n_samples_per_group))

metadata = pd.DataFrame({
    "Protein.Group":              [f"P{i:05d}" for i in range(n_proteins)],
    "Protein.Ids":                [f"P{i:05d}" for i in range(n_proteins)],
    "Protein.Names":              [f"PROT{i}_HUMAN" for i in range(n_proteins)],
    "Genes":                      [f"GENE{i}" for i in range(n_proteins)],
    "First.Protein.Description":  [f"Synthetic protein {i}" for i in range(n_proteins)],
    "N.Proteotypic.Sequences":    rng.integers(2, 8, size=n_proteins),
})
pg_matrix = pd.concat([metadata, pd.DataFrame(X, columns=sample_ids)], axis=1)
pg_matrix.to_csv(work / "toy.pg_matrix.csv", index=False)

# 2. Build a label spreadsheet matching the sample IDs.
labels = pd.DataFrame({
    "Biobank Number": sample_ids,
    "Group":          ["Non-cancer" if yi == 0 else "PDAC" for yi in y],
})
labels.to_excel(work / "toy_labels.xlsx", index=False)

# 3. Write a config.
cfg = {
    "data_file":               str(work / "toy.pg_matrix.csv"),
    "label_file":              str(work / "toy_labels.xlsx"),
    "control_name":            "Non-cancer",
    "case_name":               "PDAC",
    "group_column":            "Group",
    "sample_id_column":        "Biobank Number",
    "random_state":            42,
    "max_missing":             0.0,
    "log_transform":           False,   # data is already on a log-like scale
    "imputation_strategy":     "knn",
    "knn_neighbors":           5,
    "min_proteotypic_peptides": 2,
    "max_iterations":          5000,
    "max_candidates":          20,
    "stability_iterations":    100,     # 2 × B = 2 × 50
    "stability_pfer":          5,
    "l1_ratio":                0.3,
    "n_jobs":                  -1,
    "output_dir":              str(work / "toy_results"),
}
cfg_path = work / "toy_config.yaml"
with open(cfg_path, "w") as f:
    yaml.safe_dump(cfg, f)

# 4. Run the pipeline programmatically.
from cpss import (
    Config, load_data, save_results, create_visualizations,
    full_dataset_stability_elasticnet,
)

config = Config.from_yaml(cfg_path)
np.random.seed(config.RANDOM_STATE)

X_raw, y, sample_ids, protein_metadata = load_data(config)
results = full_dataset_stability_elasticnet(X_raw, y, config)

if results is None:
    print("No stable features found.")
else:
    save_results(results, protein_metadata, config)
    create_visualizations(results, protein_metadata, config)

    sel_probs = results["feature_stability"]["selection_probabilities"]
    stable    = results["feature_stability"]["stable_features"]
    top = sorted(stable, key=lambda f: sel_probs[f], reverse=True)[:10]
    print("\nTop stable features:")
    for f in top:
        gene = protein_metadata.loc[f, "Genes"]
        print(f"  {gene:>10s}  π̂ = {sel_probs[f]:.3f}")

If the install is healthy you should see most of the first 10 injected proteins (GENE0–GENE9) show up in the stable list. The toy dataset is small enough to finish in a few seconds on a laptop.

---

Outputs

output_dir will contain four timestamped artefacts per run:

File	Contents
stable_features_<ts>.csv	One row per measured protein with selection_probability, cohens_d, direction, n_case, n_control, imputation_rate. Sorted by π̂. The "stable" subset is everything with π̂ ≥ threshold (recorded in the manifest).
stable_features_<ts>.schema.json	Per-column descriptions for the CSV.
full_results_<ts>.pkl	Pickled results dict (feature_stability, stability_selection, preprocessing, sample_counts).
run_manifest_<ts>.json	Reproducibility manifest: package + dependency versions, input SHA-256 hashes, upstream DIA-NN log provenance, resolved config, derived parameters (C_ref, threshold, M&B threshold, θ, B, PFER), preprocessing summary, convergence stats.

create_visualizations additionally writes selection-probability and threshold-comparison plots to the same directory.

---

Programmatic API

The CLI is just a thin wrapper. The same pipeline from Python:

from cpss import (
    Config, load_data, save_results, create_visualizations,
    full_dataset_stability_elasticnet,
)

config = Config.from_yaml("config_cpss.yaml")
X_raw, y, sample_ids, protein_metadata = load_data(config)
results = full_dataset_stability_elasticnet(X_raw, y, config)
save_results(results, protein_metadata, config)
create_visualizations(results, protein_metadata, config)

Lower-level entry points exposed at the package root:

• Preprocessor(config).fit_transform(X, y) — leakage-safe missing-filter + log + impute + z-score; reusable on held-out data via .transform.
• stability_selection_elasticnet(X, y, config) — returns (stable_features, threshold, selection_probs, C_ref, threshold_info).
• compute_rconcave_threshold(q, p, B, pfer) and compute_D(...) — the threshold maths in isolation.
• calculate_cohens_d(X, y, features) — signed effect size on the preprocessed matrix.

---

Tests

pytest cpss/tests          # fast suite
pytest cpss/tests --runslow  # includes the empirical PFER calibration test

The suite covers the r-concave bound (test_rconcave.py), an empirical PFER check (test_pfer_empirical.py), Cohen's d edge cases, peptide-input rejection, preprocessing leakage guarantees, and the end-to-end CSV/manifest schema.

---

Theory — one paragraph

For each subsample the procedure fits an ElasticNet logistic regression at a fixed regularisation C_ref (calibrated once on the full data so each subsample yields ≤ q non-zero coefficients), then records the top-q features by |β|. With B complementary pairs (2B subsamples) the selection probability π̂ for each feature is the empirical proportion of subsamples in which it appeared. A feature is called stable when π̂ ≥ π, where π is the smallest threshold for which the r-concave bound (Shah & Samworth 2013, eq. 8)

E[V] ≤ min{ D(θ², 2τ-1, B, -½), D(θ, τ, 2B, -¼) } × p

is below the user-specified PFER budget. θ = q/p is data-dependent and so π must be computed after preprocessing has fixed p. The bound reduces to M&B's 1/(2τ-1) bound when r-concavity is dropped, hence the manifest reports both for diagnostic purposes.

---

Relationship to Stabl

Stabl (Hédou et al., Nature Biotechnology 2024) is a widely used stability-selection variant for omics biomarker discovery, and both packages build on stability selection. Both refit a sparse model on repeated subsamples and keep features whose selection frequency exceeds a threshold; they differ in how the threshold is obtained.

	cpss	Stabl
Threshold	Analytical: Shah & Samworth (2013) r-concave bound for a user-specified PFER	Empirical: permuted or knockoff copies of the features are appended and the threshold minimising an estimated false-discovery proportion (FDP⁺) is chosen
Error quantity	Expected number of false positives, E[V]	Estimated false-discovery proportion
Subsampling	Complementary pairs, B ≤ 50 (required for the bound)	Bootstrap, typically 1000 draws
Regularisation	Single C_ref, calibrated once so ≤ q coefficients survive	Sweep over a λ grid; selection frequency is the maximum across the grid
Preprocessing	Imputation and scaling refit inside every subsample	Expects a preprocessed matrix
Scope	Binary classification on DIA-NN pg_matrix output	Multi-omics, regression, multi-class, late fusion

Taking the maximum selection frequency over a regularisation path places Stabl outside the assumptions of the Shah & Samworth bound, which is why it calibrates the threshold against synthetic null features instead. cpss keeps a fixed regularisation strength and complementary-pairs subsampling so that the analytical bound applies, at the cost of generality. Where a permuted-feature null is used, note that independently shuffled columns lose their correlation with every measured protein and so are more easily separated from signal than genuine uninformative proteins in a correlated proteome; the threshold derived from them is not guaranteed to control the false-discovery proportion among real features.

In practice the two methods rank features very similarly on the same data and differ mainly in where the inclusion cut-off falls. Choose Stabl for flexibility across data types and tasks; choose cpss when you want the retained set to carry a stated, interpretable bound on the expected number of false positives.

Citation

Shah, R. D. & Samworth, R. J. (2013). Variable selection with error control: another look at stability selection. JRSS B, 75(1):55–80.

Hédou, J., Marić, I., Bellan, G., et al. (2024). Discovery of sparse, reliable omic biomarkers with Stabl. Nature Biotechnology, 42:1581–1593. https://doi.org/10.1038/s41587-023-02033-x