XTransferCDR: A Revolutionary Generative Framework for Feature Decoupling and Cross-Domain Transferable Representation Learning.

Introduction

XTransferCDR is a cross-domain transfer learning framework to predict cellular response at single-cell level.

Model architecture

The gene expression profiles $X^{(a)}$ and $X^{(b)}$ of a sample cell $X$ under perturbations $a$ and $b$, represented as vector-type data, serve as inputs. These inputs are decoupled by the cell state feature extractor $E_S$ and the perturbation feature extractor $E_P$, resulting in cell state feature embedding vectors $S^a$ and $S^b$ for cell $X$, as well as perturbation feature embedding vectors $P^a$ and $P^b$ for perturbations $a$ and $b$. These embedding vectors undergo cross-transfer and combination, yielding $P^a + S^a$, $P^b + S^a$, $P^a + S^b$, and $P^b + S^b$, which are subsequently fed into the decoder $D$ to produce the corresponding reconstructed gene expression profiles.

Requirements

The deep learning models were trained on 2*NVIDIA GeForce RTX 4090 on linux.

• Python 3.11
• CUDA 12.2
• PyTorch 2.0
• Pandas 2.1.4
• Numpy 1.26.3
• Scikit-learn 1.3.0

Usage

To setup the environment, install conda and run (Must run on servers with multiple GPUs):

conda create --name <your_env_name> --file requirements.txt

To train this model and obtain results, you need to download the dataset (Example: Sci-plex3), place it in the datasets folder, and then run:

torchrun --nproc_per_node=2 --master-port=29501  XTransferCDR/train.py

If you want to train other datasets, you need to modify the XTransferCDR/train.py section

Directory structure

• XTransferCDR: contains the code for the model, the evaluation, and the training loop.

• configs: Configuration file for hyperparameters.

• datasets: This directory needs to be created by yourself to store experimental data.

• results: This directory needs to be created by oneself. It is used to store the results of the experiment.

  • modules: Model storage directory.

  • plot: Training loss curve.

  • plot_data: Data of the training process.