Omics Survival Embeddings

Benchmarking different embedding methods for survival prediction on omics data.

Overview

This project explores how various dimensionality reduction and embedding techniques affect survival analysis on omics datasets.
We compared classical linear methods with nonlinear and deep-learning approaches and combined them with modern predictive models.

Methods

• Embeddings: PCA, ICA, Factor Analysis, UMAP, VAE, AutoEncoder, MOFA, NMF, CP/Tucker Tensor, Scikit-Fusion
• Models: TabNet, CatBoost, Ridge Regression
• Evaluation metrics:
  • Concordance index (C-index)
  • Mean Absolute Percentage Error (MAPE)

Key Results

• Best combination: UMAP + TabNet (C-index = 0.568, MAPE = 79.26) without clinical features.
• VAE + TabNet (with clinical features) also performed well (C-index = 0.541, MAPE = 75.71).
• TabNet consistently outperformed other models across embeddings.
• Clinical features sometimes improved performance, but not universally.
• PLS + Ridge minimized MAPE but had poor C-index → not suitable for survival tasks.

Installation

git clone https://github.com/KonNik88/omics-survival-embeddings.git
cd omics-survival-embeddings
pip install -r requirements.txt

## Usage

Open the Jupyter notebooks in notebooks/ to reproduce experiments and visualizations.
All embeddings and models are implemented with scikit-learn, PyTorch, and PyTabNet.

## Visualizations

We provide comparisons of embedding spaces (UMAP, AE, VAE) and model performance metrics in results/.