D4L-Hackaton

Setup

We recommend creating venv or conda environment with python==3.12.

Conda and requirements.txt

conda create -n multimodal_hackaton python=3.12
source activate multimodal_hackaton

And then:

pip3 install -r requirements.txt

Overview

Papers: Multimodal single cell data integration challenge: Results and lessons learned https://proceedings.mlr.press/v176/lance22a/lance22a.pdf Documentation: https://openproblems.bio/events/2021-09_neurips/documentation/data/about_multimodal When you want to run any experiment, run:

cd src

and then

python3 train_and_validate.py [ARGUMENTS]

with possible options:

• -h, --help: show this help message and exit
• --method {"omivae", "babel", "advae"}
• --config (default="standard"): Name of a configuration in src/config/{method} directory.
• --cv-seed (default=42): Seed used to make k folds for cross validation.
• --n-folds (default=5): Number of folds in cross validation.
• --subsample-frac (default=None): Fraction of the data to use for cross validation.
• --retrain (default=True): Retrain a model using the whole dataset.

Dataset

We use dataset from NeurIPS competition: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE194122 named GSE194122_openproblems_neurips2021_cite_BMMC_processed.h5ad.gz. We assume data is in path_to_repo\repo_name\data.

Metrics

Predicting one modality from another: As metrics, we consider root mean squared error (RMSE) and Pearson correlation on log-scaled counts, as well as Spearman correlation.

Matching cells between modalities: As metrics, we consider area under the precision recall curve (AUPR) and the average probability assigned to the correct matching. The latter is a relative measure per dataset that accounts for non-identifiability among cells with the same identity.

Viewing logs

To see the logs from different runs from the src directory run

tensorboard --logdir=logs