single-cell-multimodal

This is a repository of model experiments for the Multimodal Single-Cell Integration Open Problem.

For an overview of the problem this model is solving please refer to: https://openproblems.bio/neurips_docs/data/about_multimodal/

The current developed pipeline is made of two abstract flexible models, one for each of the two problems, Multiome and Citeseq, that can be built up, trained and tuned with different configurations and network topologies.

CiteSeq:

• A Truncated SVD layer. This layer is mandatory to make the problem feasible (original input space is ...)
• An optional multitask supervised embedding layer. The task employable are
  • autoenconding of the reduced input space
  • original regression problem
  • prediction of cell labelling available in the metadata (not yet developed)
• A multioutput sklearn wrapper that can be filled with a regressor for each dimension of the output:
  • currently the model employs lgbm regressors
• Instead of the multioutput a linear multiregression layer can be employed, this is not effective in case there's no supervision in the extraction of the embeddings

Multiome:

• Multiome model share the same input encoding possibilites of the Citeseq one
• on the output one can apply
  • a ridge multioutput regression
  • a lgbm that make prediction on a shrinked subspaced by mean of truncated-svd. The prediction is then projected on the full space by the inverse SVD.

Setup

pip install -r requirements.txt

chmod +x data_setup.sh

./data_setup.sh

Full Pipeline

The python entry point for the two problems are:

• main_cite.py

• main_multiome.py

They should be provided with a configuration for the model and the eventual network paramters, configurations key can be find in:

• scmm.problems.cite.configurations.__init__.py

• scmm.problems.multiome.configurations.__init__.py

an example run can be:

python3 scmm/main_cite.py baseline

python3 scmm/main_multiome.py svd_in_autsupervised_lgbm_out_svd

under each configurations module you can find the defined configurations in specific python files

The output predictions will be written under /out directory