fastscode 0.0.7

FastSCODE

Indroduction

• FastSCODE: an accelerated implementation of SCODE based on manycore computing

Installation

• :snake: Anaconda is recommended to use and develop FastSCODE.
• :penguin: Linux distros are tested and recommended to use and develop FastSCODE.

Anaconda virtual environment

After installing anaconda, create a conda virtual environment for FastSCODE. In the following command, you can change the Python version (e.g. python=3.12).

conda create -n fastscode python=3.12

Now, we can activate our virtual environment for FastSCODE as follows.

conda activate fastscode

Install from PyPi

pip install fastscode

• Default backend framework of the FastSCODE is PyTorch.- You need to install other backend frameworks such as CuPy, Jax, and TensorFlow

Install from GitHub repository

First, clone the recent version of this repository.

git clone https://github.com/cxinsys/fastscode.git

Now, we need to install FastSCODE as a module.

cd fastscode
pip install -e .

FastSCODE tutorial

Create FastSCODE instance

The FastSCODE class requires loaded files such as expression data arrays and pseudo time arrays, as well as several parameters for linear ODE optimization.

parameters

• exp_data: expression data array (Gene (G) x Cell (C)), required
• pseudotime: pseudotime data vector (C), required
• node_name: vector for name of genes (G), required
• droot: root directory for storing score matrix and RSS arrays, optional, default value is None, which means that the results are not saved
• num_tf: number of genes to use, optional, default value is None, and all genes are used
• num_cell: number of cells to use, optional, default value is None, and all cells are used
• num_z: length of vector z for optimization, optional, default: 4
• max_iter: number of iterations for optimization, optional, default: 100
• max_b: maximum initialization value for parameter b, optional, default: 2.0
• min_b: minimum initialization value for parameter b, optional, default: -10.0
• dtype: data type, optional, default: float32
• user_binary: save result matrix as binary file, optional, default: True

import fastscode as fs

exp_data = np.loadtxt(dpath_exp_data, delimiter=",", dtype=str)
node_name = exp_data[0, 1:]
exp_data = exp_data[1:, 1:].astype(np.float64).T  # gene x cell

pseudotime = np.loadtxt(dpath_trj_data, delimiter="\t")

worker = fs.FastSCODE(exp_data=exp_data,
                      pseudotime=pseudotime,
                      node_name=node_name,
                      droot=spath_droot_r,
                      num_tf=None,
                      num_cell=None,
                      num_z=num_z,
                      max_iter=max_iter,
                      dtype=np.float32,
                      use_binary=True)

Run FastSCODE

parameters

• backend: optional, default: 'cpu'
• device_ids: list or number of devcies to use, optional, default: [0] (cpu), [list of whole gpu devices] (gpu)
• batch_size_b: batch size of optimization parameter B, optional, default: 1
• batch_size: gene batch size of expression data, optional, default: None (compute all gene data at once, recommended)
• chunk_size: gene chunk size of expression data in inner loop of algorithm, optional, default: None (auto calculated)

rss, score_matrix = worker.run(backend='gpu',
                               device_ids=8,
                               sampling_batch=100,
                               batch_size=1024)

Run FastSCODE with run_scode.py

• Before run run_scode.py, batch_size_b and batch_size parameter must be modified to fit your gpu memory size

Usage

python run_scode.py --droot [root directory]
                    --fp_exp [expression file path]
                    --fp_trj [trajectory (pseudotime) file path] 
                    --fp_branch [cell select file path] 
                    --num_z [number of vector z]
                    --max_iter [number of optimization step]
                    --backend [name of backend framework]
                    --num_devices [number of devices]
                    --batch_size_b [number of parameter b]
                    --sp_droot [droot directory for saving results]
                    --num_repeat [total number of computation iterations]

Example

python run_scode.py --droot .
                    --fp_exp expression_dataTuck_sub.csv
                    --fp_trj pseudotimeTuck.txt
                    --fp_branch cell_selectTuck.txt
                    --num_z 10
                    --max_iter 100
                    --backend gpu
                    --num_devices 8
                    --batch_size_b 10
                    --sp_droot out
                    --num_repeat 6

Output

When use_binary is True

RSS.txt
ex)
3367844277.01837


score_matrix.npy
ex)
0	0.05	0.02	...	0.004
0.01	0	0.04	...	0.12
0.003	0.003	0	...	0.001
0.34	0.012	0.032	...	0


node_name.txt
ex)
GENE_1
GENE_2
GENE_3
.
.
.
GENE_M

When use_binary is False

RSS.txt
ex)
3367844277.01837

  
score_matrix.txt                            
ex)
Score	GENE_1	GENE_2	GENE_3	...	GENE_M
GENE_1	0	0.05	0.02	...	0.004
GENE_2	0.01	0	0.04	...	0.12
GENE_3	0.003	0.003	0	...	0.001
.
.
.
GENE_M	0.34	0.012	0.032	...	0

Downstream analysis tutorial

Create NetWeaver instance

parameters

• result_matrix: result score matrix of fastscode, required
• gene_names: gene names from result matrix, required
• tfs: tf list, optional
• fdr: specifying fdr, optional, default: 0.01
• links: specifying number of outdegrees, optional, default: 0
• is_trimming: if set True, trimming operation is applied on grn, optional, default: True
• trim_threshold: trimming threshold, optional, default: 0

result_matrix = np.loadtxt(fpath_result_matrix, delimiter='\t', dtype=str)
gene_name = result_matrix[0][1:]
result_matrix = result_matrix[1:, 1:].astype(np.float32)

tf = np.loadtxt(fpath_tf, dtype=str)

weaver = fs.NetWeaver(result_matrix=result_matrix,
                       gene_names=gene_name,
                       tfs=tf,
                       fdr=fdr,
                       links=links,
                       is_trimming=True,
                       trim_threshold=trim_threshold,
                       dtype=np.float32
                       )

Run weaver

• backend: optional, default: 'cpu'
• device_ids: list or number of devices to use, optional, default: [0] (cpu), [list of whole gpu devices] (gpu)
• batch_size: if set to 0, batch size will automatically calculated, optional, default: 0

grn, trimmed_grn = weaver.run(backend=backend,
                              device_ids=device_ids,
                              batch_size=batch_size)

Count outdegree

• grn: required

outdegrees = weaver.count_outdegree(grn)
trimmed_ods = weaver.count_outdegree(trimmed_grn)

Downstream analysis with reconstruct_grn.py

reconstruct_grn.py is a tutorial script for the output of grn and outdegree files.
When using a binary file, you must pass the path to the node_name.txt file to the --fp_gn parameter.
If it is not a binary file, the --fp_gn parameter is optional.

Usage

When specifying fdr

python reconstruct_grn.py --fp_rm [result matrix path]  --fp_gn [gene name file path] --fp_tf [tf file path] --fdr [fdr] --backend [backend] --device_ids [number of device]

Example

python reconstruct_grn.py --fp_rm score_result_matrix.txt --fp_gn node_name.txt --fp_tf mouse_tf.txt --fdr 0.01 --backend gpu --device_ids 1

Output

score_matrix.fdr0.01.sif, score_matrix.fdr0.01.sif.outdegrees.txt
score_matrix.fdr0.01.trimIndirect0.sif, score_matrix.fdr0.01.trimIndirect0.sif.outdegrees.txt

Usage

When specifying the links

python reconstruct_grn.py --fp_rm [result matrix path] --fp_gn [gene name file path]  --fp_tf [tf file path] --links [links] --backend [backend] --device_ids [number of device]

Example

python reconstruct_grn.py --fp_rm score_result_matrix.txt --fp_gn node_name.txt --fp_tf mouse_tf.txt --links 1000 --backend gpu --device_ids 1

Output

score_matrix.links1000.sif, score_matrix.links1000.sif.outdegrees.txt
score_matrix.links1000.trimIndirect0.sif, score_matrix.links1000.trimIndirect0.sif.outdegrees.txt

TODO

• Upload to PyPi