jruffle/classical_samples_4d

Classical Dimensionality Reduction (Sample-centric, 4D)

Pre-trained PCA and UMAP models for transcriptomics data compression, part of the TRACERx Datathon 2025 project.

Model Details

• Methods: PCA and UMAP
• Compression Mode: Sample-centric
• Output Dimensions: 4
• Training Data: TRACERx open dataset (VST-normalized counts)

Contents

The model file contains:

• PCA: Principal Component Analysis model
• UMAP: Uniform Manifold Approximation and Projection model (2-4D only)
• Scaler: StandardScaler fitted on TRACERx data
• Feature Order: Gene/sample order for alignment

Usage

These models are designed to be used with the TRACERx Datathon 2025 analysis pipeline. They will be automatically downloaded and cached when needed.

import joblib

# Load the model bundle
model_data = joblib.load("model.joblib")

# Access components
pca = model_data['pca']
scaler = model_data['scaler']
gene_order = model_data.get('gene_order')  # For sample-centric

# Transform new data
scaled_data = scaler.transform(aligned_data)
embeddings = pca.transform(scaled_data)

Training Details

• Input Features: 20,136 genes
• Training Samples: 1,051 samples
• Preprocessing: StandardScaler normalization

Files

• model.joblib: Model bundle containing PCA, UMAP, scaler, and feature order