Generating GP Gene Cosine Similarity

06.1_generate_gp_gene_cosine_similarity.py

import os
import shutil

import pandas as pd

import tripso

# Directory paths for loading/saving
root_dir = '/lustre/scratch126/cellgen/lotfollahi/mm58/gplearner_reproducibility/07_tutorial_zeng'
output_dir = os.path.join(root_dir, 'output_global')
target_dir = os.path.join(output_dir, 'gene_embeddings_analysis')
gpdb_path = os.path.join(root_dir, 'gpdb_tf.csv')

gpdb = pd.read_csv(gpdb_path)


def write_gene_embeddings(gp):
    gp_genes = gpdb[gp].dropna().tolist()

    data_dir = os.path.join(root_dir, 'data/processed/input_dataset')

    gp_downstream = tripso.gpEval(
        dataset_path=data_dir,
        gpdb_path=gpdb_path,
        output_dir=output_dir,
        tissue='zeng',
        model_type='Global',
    )

    dc = ['pDC', 'cDC']

    for split in ['train', 'val', 'test']:
        gp_downstream.generate_gene_embeddings(
            gp_for_forward=gp,
            gp_for_downstream=gp,
            split=split,
            obs_key='AuthorCellType_Broad',
            obs_value=dc,
            genes_to_keep=gp_genes,
            precision='16-mixed',
            return_gene_cosim='gene_to_gp',
        )

    # move saved embeddings to target directory
    if not os.path.exists(target_dir):
        os.makedirs(target_dir)

    emb_file_name = f"{gp}_gene_embeddings_{'_'.join(dc)}"

    shutil.move(
        os.path.join(output_dir, emb_file_name),
        os.path.join(target_dir, f'{gp}_dendritic_cells'),
    )


if __name__ == '__main__':
    for gp in ['GP_IRF1', 'GP_USF1']:
        write_gene_embeddings(gp)