scVI and SCANVI Training Pipeline

This repository contains a Snakemake workflow and helper scripts for training scVI and SCANVI models on single-cell RNA sequencing data. The pipeline performs data preprocessing, model training, benchmarking and differential expression analyses.

Features

• Data preprocessing – filters cells and genes, manages annotation labels and removes mitochondrial genes.
• Model training – trains scVI and SCANVI models with customizable parameters and checkpointing.
• Query mapping – maps query data sets into a trained reference latent space.
• Benchmarking – evaluates embeddings using PCA, Scanorama, Harmony, scVI and SCANVI.
• Result extraction – exports training metrics and differential expression results.

Installation

Create the conda environment used by the workflow:

mamba env create -f environment.yml

Activate the environment before running the pipeline.

Usage

Edit config/config.yaml to point to your input AnnData file and desired output directory. The workflow can then be executed locally with

snakemake --use-conda --cores 4 -s Snakefile

For HPC execution a Slurm profile is provided under config/gpu (see the README in that directory).

Individual steps can also be run directly via the scripts in scripts/.

Repository layout

• scripts/ – Python scripts used in the workflow.
• config/ – configuration files and cluster profiles.
• notebooks/ – example notebooks exploring various parts of the pipeline.