EmbryoidNeuralODE

Neural ODE Implementation for Invertible Distribution Transport between an Embryoid Cell Distribution and the Gaussian Distribution

Overview

This project implements a Neural ODE to transport a Gaussian distribution to an embryoid cell distribution while retaining invertibility, allowing for the transformation back to a Gaussian distribution. The implementation utilizes TorchDyn and PyTorch Lightning to achieve these transformations with additional features like magnitude regularization and PCA-to-gene space inversion.

Technical Implementation

Neural ODE

• Implements a Neural ODE using TorchDyn.

PyTorch Lightning Module

• Implements the Learner(pl.LightningModule) class to handle the training and validation processes.
• Transports a Gaussian distribution to a cell distribution while retaining invertibility, allowing transformations from cells to a Gaussian and vice versa.

PCA to Gene Space Inversion

• Inverts generated cells from PCA space back to gene space.

Magnitude Regularization

• Utilizes TorchDyn for magnitude regularization.
• References Tong et al. 2020 for the regularization approach.

Feature Plots

PHATE vs PCA

• A figure with two subplots, one for each embedding.
• Cells are colored by time.

Generated Cells vs Real Cells in PCA Space

• A figure comparing generated cells and real cells in PCA space.
• Real cells use a different marker than generated cells.
• Cells are colored by time.

Generated Cells vs Real Cells Expression

• A figure with 10 subplots (2 rows of 5).
  • One row for generated data.
  • One row for ground truth data.
• Generated data is colored differently than ground truth data.