Predicting Isoform Expression from Gene-Level Profiles using Representation Learning

A comparison of deep learning approaches for analyzing single-cell gene expression data.

Overview

This project explores different neural network architectures for single-cell RNA sequencing (scRNA-seq) data, comparing their performance across various embedding strategies.

Models

• MLP — with PCA, scGPT, and Gene-Aware embeddings
• VAE — Variational Autoencoder for dimensionality reduction
• LSTM — Sequential modeling of gene expression
• Transformer — Attention-based architecture

Project Structure

├── scripts/          # Core modules (data loading, models, training, utils)
├── train_scripts/    # Training scripts for each model variant
└── Results/          # Saved model outputs and metrics

Results

Trained models and evaluation metrics are saved in the Results/ folder, organized by architecture.

Notebook

The notebook reproduces the main results, importing functions from our main scripts.