Automated Cell Type Identification using Neural Networks

August 18, 2019 ยท View on GitHub

Overview

ACTINN (Automated Cell Type Identification using Neural Networks) is a bioinformatic tool to quickly and accurately identify cell types in scRNA-Seq. For details, please read the paper: https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btz592/5540320 All datasets used in the paper are available here: https://figshare.com/articles/ACTINN/8967116

Prerequisite

python 3.6

python packages: tensorflow 1.10+, numpy 1.14+, pandas 0.23+, argparse 1.1+, scipy 1.1+

Convert format

We use HDF5 format for the scRNA-Seq expressiong matrix, which stores the compressed matrix and is fast to load. To convert the format, we first read the expression matrix as a pandas dataframe, then we use the to_hdf function to save the file as HDF5 format. For the to_hdf function, we use "dge", which stands for digital gene expression, for the key parameter.

Usage

python actinn_format.py -i input_file -o output_prefix -f format

Paramters

  • -i Path to the input file or the 10X directory
  • -o Prefix of the output file
  • -f Format of the input file (10X_V2, 10X_V3, txt, csv)

Output

The output will be an HDF5 formated file named after the output prefix with ".h5" extension

Examples

Convert 10X_V2 format

python actinn_format.py -i ./test_data/train_set_10x -o train_set -f 10X_V2

Convert 10X_V3 format

python actinn_format.py -i ./test_data/train_set_10x -o train_set -f 10X_V3

Convert txt format

python actinn_format.py -i ./test_data/train_set.txt.gz -o train_set -f txt

Convert csv format

python actinn_format.py -i ./test_data/train_set.csv.gz -o train_set -f csv

Predict cell types

We train a 4 layer (3 hidden layers) neural network on scRNA-Seq datasets with predifined cell types, then we use the trained parameters to predict cell types for other datasets.

Usage

python actinn_predict.py -trs training_set -trl training_label -ts test_set -lr learning_rat -ne num_epoch -ms minibatch_size -pc print_cost -op output_probability

Parameters

  • -trs Path to the training set, must be HDF5 format with key "dge".
  • -trl Path to the training label (the cell types for the training set), must be tab separated text file with no column and row names.
  • -ts Path to test sets, must be HDF5 format with key "dge".
  • -lr Learning rate (default: 0.0001). We can increase the learning rate if the cost drops too slow, or decrease the learning rate if the cost drops super fast in the beginning and starts to fluctuate in later epochs.
  • -ne Number of epochs (default: 50). The number of epochs can be determined by looking at the cost after each epoch. If the cost starts to decrease very slowly after ceartain epoch, then the "ne" parameter should be set to that epoch number.
  • -ms Minibatch size (default: 128). This parameter can be set larger when training a large dataset.
  • -pc Print cost (default: True). Whether to print cost after each 5 epochs.
  • -op Output probabilities for each cell being the cell types in the training data (default: False).

Output

The output will be a file named "predicted_label.txt". In the file, the first column will be the cell name, the second column will be the predicted cell type. If the "op" parameter is set to True, there will be another output file named "predicted_probablities.txt", where columns are cells and rows are cell types. The number in row i and column j will be the probablity that cell j being cell type i.

Example

python actinn_predict.py -trs ./test_data/train_set.h5 -trl ./test_data/train_label.txt.gz -ts ./test_data/test_set.h5 -lr 0.0001 -ne 50 -ms 128 -pc True -op False

Plots

We show an example on how to create a tSNE plot with the predicted cell types. The R command can be found in the "tSNE_Example" folder. tSNE Plot