DeepDoublet: A tool for high-resolution doublet decomposition based on deep learning

A tool for decomposing a doublet (mixed two single cells when sequencing) into two single cells based on their trascriptome.

The matrices in "umi" folder only included expression values of 8,676 selected genes, while the matrices in "FullGenes" folder reserved information of all available genes.

Hardware requirement

• Disk space and memory
  To generate 400k artificial doublets wiht around 9k genes included, you will need around 30GB disk space and 60GB memory.
• GPU
  To train a model on 400k artificial doublets with information of 9k genes. At least a 10GB GTX 3080 is required.
  In our application, we finished the training in half an hour with a 24GB Titan RTX.

Environment

Anaconda is needed to set a proper Python environment for running DeepDoublet:

conda env create --file environment.yml
source activate DeepDoublet

Generate artificial doublets

python GenAD_v1.py -h

• Example Generate 400,000 hepatocyte-LEC doublets

python GenAD_v1.py --s1 umi/hep_umi.csv --s2 umi/LEC_umi.csv --sampleSize 400000 --outdir ADoub_new

Artificial Doublets Storage and Composition

The artificial doublets are organized and stored as follows:

Directory Structure

• Folder: ADoub_new

  • Files:
    • ADoub.npy
      Contains all artificial doublets as a NumPy array.
    • ADoub_meta.csv
      Stores the composition metadata for each artificial doublet.

• Folder: umi

  • Files:
    • LEC_umi.csv
      Contains UMI (Unique Molecular Identifier) counts for LEC cells.
    • hep_umi.csv
      Contains UMI counts for HEP cells.

Metadata File: ADoub_meta.csv

This CSV file contains the composition information for each artificial doublet. Each row corresponds to an artificial doublet stored in ADoub.npy.

Column	Description
s1	Index of the first single cell
MF_s1	Proportion of the first single cell (s1)
s2	Index of the second single cell
MF_s2	Proportion of the second single cell (s2)

Single Cell Data

Single cell transcriptome information is stored in the umi folder. The folder contains two CSV files:

• Files:

  • LEC_umi.csv

    • Row Names: Gene names
    • Column Names: Cell names
    • Description: UMI counts for LEC cells. Only selected genes are included.

  • hep_umi.csv

    • Row Names: Gene names
    • Column Names: Cell names
    • Description: UMI counts for HEP cells. Only selected genes are included.

Notes:

• Ensure that the gene names across LEC_umi.csv and hep_umi.csv are consistent if they are to be used together.
• The selected genes in these CSV files should correspond to the genes used in downstream analyses.

Transcriptome Calculation

The transcriptome of each artificial doublet is calculated using the following formula:

Transcriptome_ADoub = Transcriptome_s1 X MF_s1 + Transcriptome_s2 X MF_s2

Where:

• Transcriptome_s1: Transcriptome data of the first single cell (s1)
• MF_s1: Proportion of the first single cell in the doublet
• Transcriptome_s2: Transcriptome data of the second single cell (s2)
• MF_s2: Proportion of the second single cell in the doublet

For example, if s1, MF_s1, s2, MF_s2 are 596, 0.7, 250, 0.3 respectively. The corresponding artificial doublet is

Transcriptome_s1_596 X 0.7 + Transcriptome_s2_250 X 0.3

Example Implementation in Python

import numpy as np
import pandas as pd

# Paths to the metadata and doublets
meta_path = 'ADoub_new/ADoub_meta.csv'
doublets_path = 'ADoub_new/ADoub.npy'
lec_umi_path = 'umi/LEC_umi.csv'
hep_umi_path = 'umi/hep_umi.csv'

# Load metadata
metadata = pd.read_csv(meta_path)

# Load single cell transcriptomes
# Load LEC and HEP UMI counts
lec_umi = pd.read_csv(lec_umi_path, index_col=0)
hep_umi = pd.read_csv(hep_umi_path, index_col=0)

# Load artificial doublets
ADoub = np.load(doublets_path)

Artificial Doublets and Trained deep learning models

The two deep learning models for decomposing hepatocyte-LEC doublets into hepatocytes and LECs, along with 400,000 artificial doublets for additional testing, have been uploaded here: https://figshare.com/articles/dataset/DeepDoublet_identifies_neighboring_cell_dependent_gene_expression/27217101

Workflow

• Train Decomposition Model
  Decomposition_Training.ipynb
• Predict with Decomposition Model
  Decomposition_Predict.ipynb
• Differential Expression Analysis
  DEA.ipynb

Logistic Regression

• Program
  LR.py
• LOG file
  LR.log