scCapsNet-mask: an updated version of scCapsNet with extended applicability in functional analysis related to scRNA-seq data

This repository contains the official Keras implementation of:

scCapsNet-mask: an updated version of scCapsNet with extended applicability in functional analysis related to scRNA-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 and analysis

• About this article

#Augments:
#'--inputdata', type=str, default='data/retina_data.npz', help='address for input data'
#'--inputcelltype', type=str, default='data/retina_celltype.npy', help='address for celltype label'
#'--num_classes', type=int, default=15, help='number of cell type'
#'--randoms', type=int, default=30, help='random number to split dataset'
#'--test_size', type=float, default=0.1, help='test size'
#'--dim_capsule', type=int, default=32, help='dimension of the capsule'
#'--activation_F', type=str, default='relu', help='activation function'
#'--batch_size', type=int, default=400, help='training parameters_batch_size'
#'--epochs', type=int, default=15, help='training parameters_epochs'
#'--lr', type=float, default=0.00015, help='Learning rate for Adam optimizers'
#'--training', type=str, default='T', help='training model(T) or loading model(F) '
#'--weights', type=str, default='data/retina_demo.weight', help='trained weights'
#'--plot_direction', type=str, default='one_side', help='display option, both_side or one_side'
#'--pc_slice', type=int, default=20, help='fineness divided along PC direction '
#'--threshold', type=float, default=0.05, help='threshold for setting dotted line'
#'--test_new_sample', type=str, default='F', help='test_new_sample (T) or not (F)'



For PBMC_dataset
python scCapsNet_mask.py --inputdata=data/PBMC_data.npz --inputcelltype=data/PBMC_celltype.npy --num_classes=8 --dim_capsule=16 --pc_slice=30 --epochs=10

For RBC_dataset
python scCapsNet_mask.py --inputdata=data/retina_data.npz --inputcelltype=data/retina_celltype.npy --num_classes=15 --dim_capsule=32 --pc_slice=20

• Further Explore

#- Testing your own dataset
Tips: set the dim_capsule twice as the number of cell types
python scCapsNet_mask.py --inputdata=your_data --inputcelltype=your_inputcelltype --num_classes=your_num_classes --dim_capsule=your_dim_capsule

2. Model analysis

• Demo -- About this article

The following codes could reproduce Figures in the article.

For PBMC_dataset
python scCapsNet_mask.py --inputdata=data/PBMC_data.npz --inputcelltype=data/PBMC_celltype.npy --num_classes=8 --dim_capsule=16 --pc_slice=30 --weights=data/PBMC_demo.weight --training=F

For RBC_dataset
python scCapsNet_mask.py --inputdata=data/retina_data.npz --inputcelltype=data/retina_celltype.npy --num_classes=15 --dim_capsule=32 --pc_slice=20 --weights=data/retina_demo.weight --training=F

3. Test new samples after model training

#- Testing new samples
python scCapsNet_mask.py --inputdata=your_data --num_classes=your_num_classes --dim_capsule=your_dim_capsule --weights=data/your.weight --training=F --test_new_sample=T

Output The output is in the results folder, including

• two plot

Prediction_accuracy_curve.png (Line_plot)
Choosen_genes.png (Scatter plot)

• training weight

training_n8_r30_dim16_e10_b400_.weight

• Prediction probability

Prediction_probability.npy

• Cell type related genes

total_select_genes_one_side.npy

capsule networks implementation

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