loveslide

A Python interface to the SLIDE framework for latent factor discovery and statistical inference.

---

📘 Overview

loveslide wraps key components of the original SLIDE R package into a user-friendly Python interface, making it easier to incorporate into machine learning pipelines and bioinformatics workflows.

SLIDE (Statistical Latent Inference for Discovery and Explanation) combines:

• LOVE: A latent factor discovery algorithm using model-based overlapping clustering.
• Knockoffs: For statistically rigorous identification of significant standalone and interacting latent factors.

This Python implementation retains R underpinnings via rpy2 and is structured to be modular, extensible, and accessible from both the command line and within Python scripts or notebooks.

---

🔗 Related Repositories

• 📦 Original R package: https://github.com/jishnu-lab/SLIDE
• 🐍 Python wrapper: https://github.com/alw399/SLIDE_py

---

🚀 Installation

Set up a compatible Python environment:

module load anaconda3/2022.10
conda create -n loveslide_env python=3.9 r-base
conda activate loveslide_env
pip install loveslide

If needed, clone the environment used during development:

# On the cluster:
source activate /ix3/djishnu/alw399/envs/rhino

---

⚡ Quick Start

📿 Command Line

python slide.py \
  --x_path /path/to/your/features.csv \
  --y_path /path/to/your/labels.csv \
  --out_path /path/to/output/

Use full paths if not running from the src/loveslide directory.

---

🧪 In a Notebook

import loveslide

from loveslide import OptimizeSLIDE

input_params = {
    'x_path': '/path/to/features.csv',
    'y_path': '/path/to/labels.csv',
    'fdr': 0.1,
    'thresh_fdr': 0.1,
    'spec': 0.2,
    'y_factor': True,
    'niter': 500,
    'SLIDE_top_feats': 20,
    'rep_CV': 50,
    'pure_homo': True,
    'delta': [0.01],
    'lambda': [0.5, 0.1],
    'out_path': '/path/to/output/'
}

slider = OptimizeSLIDE(input_params)
slider.run_pipeline(verbose=True, n_workers=1)

---

🔬 Pipeline Overview

The run_pipeline() method follows three key stages:

🧩 Stage 1: Latent Factor Discovery

• LOVE Algorithm: Identifies overlapping latent factors in the data.
• Output: Latent factor matrix (z_matrix) and factor loadings.

📊 Stage 2: Statistical Inference with Knockoffs

• Identifies significant standalone and interacting latent factors.
• Controls False Discovery Rate (FDR) to maintain statistical rigor.

📈 Stage 3: Visualization

• Diagnostic plots
• Top genes/features for each latent factor (loadings > |0.05|)

---

⚙️ Parameters

Name	Type	Description	Default/Example
x_path	str	Path to feature matrix CSV	Required
y_path	str	Path to response/labels CSV	Required
fdr	float	Knockoff FDR threshold	0.1
thresh_fdr	float	FDR threshold in LOVE	0.1
spec	float	Minimum reproducibility for a factor	0.2
y_factor	bool	Treat y as categorical	True
niter	int	Iterations for LOVE	500
SLIDE_top_feats	int	Number of top features to plot	20
rep_CV	int	Repeats for cross-validation	50
pure_homo	bool	Use pure variables with loadings = 1	True
delta	list	Regularization parameters	[0.01]
lambda	list	Penalty parameters	[0.5, 0.1]
out_path	str	Output directory	Required

---

🏗️ Project Structure

SLIDE_py/
├── src/
│   ├── loveslide/             # Main Python & R wrappers
│   │   ├── slide.py           # Main entry point
│   │   ├── love.py
│   │   ├── knockoffs.py
│   │   ├── ...
│   │   ├── love_r/            # R reference implementation of LOVE
│   │   └── slide_r/           # R utilities for SLIDE (sourced via rpy2)
├── dist/
├── example/
├── ...

---

🧠 Design Notes

• Core statistical inference is done using R scripts via rpy2.
• Python acts as an orchestration layer to allow integration into ML workflows.
• Most plotting is done in R (e.g., pheatmap, ggplot2).

---

📌 Known Limitations and TODOs

• YAML → dictionary conversion for easier parameter management
• Extend y_factor handling to non-binary variables
• Parallelization of knockoff inference (e.g., in select_short_freq)
• Correlation networks visualization using networkx

---

📢 Citation & Contact

If you use loveslide in your work, please cite the original R implementation and this repository. For bugs or feature requests, please open an issue on GitHub.

• Homepage: SLIDE_py on GitHub

• Issues: Report an Issue

• Authors:

  • Ally Wang (alw399@pitt.edu)
  • Swapnil Keshari (swk25@pitt.edu)