cpiVAE: Cross-platform Proteomics Imputation using Variational Autoencoders

Welcome to the comprehensive documentation for cpiVAE, a highly customizable and user friendly framework for cross-platform proteomics data imputation using Variational Autoencoders.

Overview

cpiVAE addresses the critical challenge of data harmonization in proteomics research by enabling accurate imputation between different measurement platforms (e.g., Olink and SomaScan). Our approach leverages deep generative models to learn shared latent representations that capture the underlying biological signals across platforms.

Key Features

• Multi-Architecture Support: Joint VAE, VampPrior, IWAE, VQ-VAE, and easy addition of custom models
• Robust Baseline Methods: KNN and WNN implementations for comparison
• Comprehensive Evaluation: Feature importance analysis and quality control metrics
• Production Ready: PyTorch Lightning framework with full experiment tracking
• Flexible Configuration: YAML-based configuration setup for easy customization

Quick Start

Get started with cpiVAE in three simple steps:

1. Install Dependencies

   pip install -r requirements.txt
   

2. Train a Model

   python scripts/train.py --config configs/default.yaml \
       --platform_a data/olink_train.csv \
       --platform_b data/somascan_train.csv
   

3. Perform Cross-Platform Imputation

   python scripts/impute.py --checkpoint outputs/model.ckpt \
       --platform_impute data/test.csv --impute_target b
   

Documentation Sections

Section	Description
Training	Model training with PyTorch Lightning
Baselines	KNN and WNN baseline implementations
Imputation	Cross-platform imputation pipeline
Model Tuning	Hyperparameter optimization and model selection
Quality Control	Data validation and preprocessing
Feature Importance	Analyzing learned representations
Latent Space	Visualization and interpretation
Confidence & Analysis	Confidence estimation and downstream analysis
Phenotype Association	Discovery and targeted phenotype associations
Comparison	Benchmarking against baselines

Architecture

cpiVAE employs a dual-encoder, dual-decoder architecture that learns platform-specific representations while maintaining a shared latent space. This design enables:

• Bidirectional Imputation: A→B and B→A cross-platform prediction
• Latent Alignment: Shared biological signal representation
• Platform Adaptation: Handling platform-specific measurement characteristics

Research Applications

• Multi-platform Studies: Harmonize data across different proteomics platforms
• Data Integration: Combine measurements from multiple cohorts or studies
• Method Validation: Compare measurement consistency across platforms

Citation

If you use cpiVAE in your research, please cite:

TBD

---