loveslide 0.0.3

Python version of R package SLIDE

SLIDE_py

A bunch of python wrappers for R code

Overview

SLIDE combines the LOVE (Latent Model-Based Clustering for Biological Discovery) clustering algorithm with knockoff-based statistical inference to identify significant standalone and interacting latent factors. Link to R package: [!https://github.com/jishnu-lab/SLIDE]

Quick Start

Basic Usage

If running into trouble, feel free to use or clone the environment here: /ix3/djishnu/alw399/envs/rhino

From command line

Use the full path if you are not calling slide.py from the same directory

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

In a notebook

import sys
sys.path.append('src/SLIDE')

from slide import OptimizeSLIDE

# Configure input parameters
input_params = {
    'x_path': '/path/to/your/features.csv',
    'y_path': '/path/to/your/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/directory'
}

# Initialize and run SLIDE
slider = OptimizeSLIDE(input_params)
slider.run_pipeline(verbose=True, n_workers=1)

Pipeline Overview

The run_pipeline() has three main parts:

Stage 1: Latent Factor Discovery

• LOVE Algorithm: Runs the overlapping clustering algorithm to identify latent factors
• Output: Generates the latent factors (z_matrix) representing underlying data structure

Stage 2: Statistical Inference with SLIDE

• 2a) Standalone Factor Analysis: Uses knockoffs to identify statistically significant standalone latent factors
• 2b) Interaction Analysis: Applies knockoffs to discover significant interacting latent factor pairs
• Feature Selection: Controls false discovery rate (FDR) while maintaining statistical power

Stage 3: Visualization

• Control Plots: Generates diagnostic plots to assess model performance and statistical validity
• Latent Factor Genes: For each latent factor, plots the top features with loadings > abs(0.05)

Parameter Configuration

Parameter	Type	Description	Default/Example
x_path	str	Path to feature matrix CSV file	Required
y_path	str	Path to response labels CSV file	Required
fdr	float	False discovery rate threshold (Knockoffs)	0.1
thresh_fdr	float	FDR threshold for feature selection (LOVE)	0.1
spec	float	minimum % times an LF found to be significant in order to be included	0.2
y_factor	bool	Treat response as factor variable	True
niter	int	Number of iterations	500
SLIDE_top_feats	int	Number of top features to display	20
pure_homo	bool	Use homogeneous loadings for pure variables	True
delta	list	Regularization parameter(s)	[0.5, 0.1]
lambda	list	Penalty parameter(s)	[0.1]
out_path	str	Output directory path	Required

Advanced Configuration

• pure_homo=True: Forces pure variable loadings to be 1 (recommended)
• pure_homo=False: Relaxes the pure variable loading constraint being 1 without losing any guarantees. However, it is difficult to find the right delta parameter
• n_workers: Controls parallelization (1 for sequential processing), but CURRENTLY NOTHING IS PARALLELIZED
• verbose: Enables detailed progress reporting (just a bunch of print statements)

Project Structure

SLIDE_py/
├── src/
│   ├── SLIDE/              # Core SLIDE implementation
│   │   ├── slide.py        # Main Python interface
│   │   └── ...            # Supporting R functions
│   └── LOVE-master/        # Original LOVE algorithm
│       ├── ...            # Original LOVE code (do not use)
│       ├── ...            # pure_homo LOVE code (use carefully)
|   └── LOVE-SLIDE/        # SLIDE implementation of LOVE

Implementation Details

LOVE Algorithm Integration

• Primary Implementation: Located in src/SLIDE/get_Latent_Factors.R
• Alternative Version: Available in LOVE-master when pure_homo=False
• Note: The original LOVE code in LOVE-master may yield different results than the SLIDE implementation and is provided for reference

To-do list

These files

• Yaml conversion: Since people already have pipelines set up, it would be convenient to have a function to read yamls into dictionaries
• Other y_factor: Currently only binary y is accomodated.
• Parallelization: Knockoffs can be made much faster. Please see select_short_freq in src/SLIDE/knockoffs.py. I was trying to use concurrent futures/ pqdm but I couldn't figure out the errors and gave up.
• Correlation networks: I think networkx can make similar graph-like figures, but I'm not familiar with making them