Predicting immune responses on multi-modal single-cell data with variational inference

This code accompanies the master thesis Predicting immune responses on multi-modal single-cell data with variational inference (https://repository.tudelft.nl/islandora/object/uuid%3A1b24699a-3967-4b08-9316-dae8d9577222?collection=education).

Author: Francesca Drummer

Supervisors: Dr. Ahmed Mahfouz and Mikhael Manurung

Package Structure

The repository is centered around the scr_trainer module in the new_model folder:

• src\_trainer.main contains training and evaluation functions
• src\_trainer.preprocessing contains data preprocessing
• src\_trainer.plotting contains ModelEvaluation class and functions for plotting
• src\_trainer.SCVI\_model contains scVI model trained with RNA
• src\_trainer.TOTALVI\_model contains totalVI model
• src\_trainer.cellPMVI\_model contains variants of cellPMVI model:
  • cellPMVI with isotropic normal prior (uses cellPMVAE module)
  • cellPMVI\_lp with Laplace prior (uses cellPMVAE\_lp module)
• src\_trainer.cellPMVI\_CITESEQ contains adaption of cellPMVI model that is based on totalVI (uses cellPMVVAE\_CITESEQ module)
• src\_trainer.my\_base\_component contains cellPMVI encoder variant
• src\_trainer.my\_training\_plan contains own extension of training plan
• src\_trainer.my\_vae contains cellPMVI VAE variant

Additional files and folders:

• notebooks contains notebooks to reproduce plots from the paper and detailed analysis of each model
• scripts contains the bash file for automatic running of the model
• CPA necessary adjustments to CPA to run with czi data
• input contains trained models
• diff_exp contains each cell types csv file with p-value of the differential expression analysis
• data contains datasets in h5ap format
• results contains the csv and pickle files after model evaluation

Run

There are two options for executing the main file: 1) Training and 2) Evalution of a trained model. The first argument --func defines which of them gets executed:

1. --func train\_model

2. --func evaluate\_model

Training

Mandatory arguments

• --dataset\_path: Respective location of .h5ad data to load
• --model\_type: Type of model to train. There are four different available types of models:
  • SCVI\_RNA: scvi model with RNA data
  • SCVI\_protein: scvi model with protein data
  • MMVAE: MMVAE model with one encoder for each RNA and protein
  • TotalVI: default TotalVI model from scvi-tools

Evaluation

Mandatory arguments:

• --filename: model name (DATE combination)
• --model\_type: Type of model to evaluate
• --training\_scenario: Training scenario 1,2, or 3 for evaluation