An interpretable deep-learning architecture of capsule networks for identifying cellular-type gene expression programs from single-cell RNA-seq data

This repository contains the official Keras implementation of:

An interpretable deep-learning architecture of capsule networks for identifying cellular-type gene expression programs from single-cell RNA-seq data

Requirements

• Python 3.6
• conda 4.4.10
• keras 2.2.4
• tensorflow 1.11.0
• establish the environment through code "conda create -n sccaps python=3.6 keras=2.2.4 tensorflow=1.11.0"

1. Model training

To unzip the PBMC_data.rar into dictionary 'data' then run:

• About this article

  
  #- default:
  python Model_Training.py

Augments:

'--inputdata', type=str, default='data/PBMC_data.npy', help='address for input data')

'--inputcelltype', type=str, default='data/PBMC_celltype.npy', help='address for celltype label')

'--num_classes', type=int, default=8, help='number of cell type')

'--randoms', type=int, default=30, help='random number to split dataset')

- with augments

python Model_Training.py --randoms 20

- *Further Explore*

- default:

python FE_Model_Training.py

Augments:

'--inputdata', type=str, default='data/PBMC_data.npy', help='address for input data'

'--inputcelltype', type=str, default='data/PBMC_celltype.npy', help='address for celltype label'

'--num_classes', type=int, default=8, help='number of cell type'

'--randoms', type=int, default=30, help='random number to split dataset into training and testing set'

'--dim_capsule', type=int, default=16, help='dimension of the capsule'

'--num_capsule', type=int, default=16, help='number of the primary capsule'

'--batch_size', type=int, default=400, help='training parameters batch_size'

'--epochs', type=int, default=10, help='training parameters epochs'

'--Model_weights', type=str, default='Modelweight.weights', help='Model_weight'

- with augments

python FE_Model_Training --randoms 20 --dim_capsule 32

**2. Model analysis**

- *Demo -- About this article*

The following documents (Jupyter Notebook) contain the codes for reproducing Figure in the main text.

demo_reproducing_figures.ipynb

In case you don't comfortable with Jupyter Notebook, you could copy the following file from the dictionary 'old_files' into current dictionary then run:

python Fig1CD_Fig2DE_Fig3.py python Fig2BC.py python Fig4.py python Fig5D.py

- *Further Explore*

- 1. heatmap for coupling coefficients, relationship between primary capsule and type capsule

python FE_Model_analysis_1.py

Augments:

'--inputdata', type=str, default='data/PBMC_data.npy', help='address for input data'

'--inputcelltype', type=str, default='data/PBMC_celltype.npy', help='address for celltype label'

'--num_classes', type=int, default=8, help='number of cell type'

'--randoms', type=int, default=30, help='random number to split dataset'

'--dim_capsule', type=int, default=16, help='dimension of the capsule'

'--num_capsule', type=int, default=16, help='number of the primary capsule'

'--weights', type=str, default='Modelweight.weights', help='trained weights'

Output:

FE_Model_analysis_1_heatmap.png FE_Model_analysis_1_overall_heatmap.png

- 2. select genes along one principle component for a specific primary capsule according to the heatmap

python FE_Model_analysis_2.py

Augments:

'--inputdata', type=str, default='data/PBMC_data.npy', help='address for input data'

'--inputcelltype', type=str, default='data/PBMC_celltype.npy', help='address for celltype label'

'--num_classes', type=int, default=8, help='number of cell type'

'--randoms', type=int, default=30, help='random number to split dataset'

'--dim_capsule', type=int, default=16, help='dimension of the capsule'

'--num_capsule', type=int, default=16, help='number of the primary capsule'

'--weights', type=str, default='Modelweight.weights', help='trained weights'

'--PC', type=int, default=1, help='indicate which principle component will be analyzed '

'--Primary_capsule', type=int, default=4, help='indicate which primary capsule will be analyzed'

'--Cell_type', type=int, default=-1, help='indicate which cell type will be analyzed'

Output:

Max2Min_genes.npy Min2Max_genes.npy FE_Model_analysis_2_resutls.png

**3. Comparison Model**

- SVM,RF,LDA,KNN

cd comparison_model python Machine_Learning.py

- Neural Networks

cd comparison_model python Neural_Networks.py

**capsule networks implementation**

the capsule parts refer to https://github.com/bojone/Capsule and https://kexue.fm/archives/5112