ONTraC 0.0.4b4

A niche-centered, machine learning method for constructing spatially continuous trajectories

ONTraC

ONTraC (Ordered Niche Trajectory Construction) is a niche-centered, machine learning method for constructing spatially continuous trajectories. ONTraC differs from existing tools in that it treats a niche, rather than an individual cell, as the basic unit for spatial trajectory analysis. In this context, we define niche as a multicellular, spatially localized region where different cell types may coexist and interact with each other. ONTraC seamlessly integrates cell-type composition and spatial information by using the graph neural network modeling framework. Its output, which is called the niche trajectory, can be viewed as a one dimensional representation of the tissue microenvironment continuum. By disentangling cell-level and niche- level properties, niche trajectory analysis provides a coherent framework to study coordinated responses from all the cells in association with continuous tissue microenvironment variations.

Required packages

pyyaml=6.0.1 pandas=2.1.1 pytorch=2.2.1 torch_geometric=2.5.1

Installation

• Setup environment

  # Create new environment with Python 3.11
  conda create -y -n ONTraC python=3.11
  conda activate ONTraC
  
  # Install required packages
  pip install pyyaml==6.0.1 pandas==2.2.1 torch==2.2.1 torch_geometric==2.5.0
  

• Install ONTraC

  git clone https://github.com/wwang-chcn/ONTraC.git
  cd ONTraC && pip install .
  

Tutorial

Input File

A example input file is provided in examples/stereo_seq_brain/original_data.csv. This file contains all input formation with five columns: Cell_ID, Sample, Cell_Type, x, and y.

Cell_ID	Sample	Cell_Type	x	y
E12_E1S3_100034	E12_E1S3	Fibro	15940	18584
E12_E1S3_100035	E12_E1S3	Fibro	15942	18623
...	...	...	...	...
E16_E2S7_326412	E16_E2S7	Fibro	32990.5	14475

For detailed information about input and output file, please see IO files explanation.

Running ONTraC

The required options for running ONTraC are the paths to the input file and the three output directories:

• preprocessing-dir: This directory stores preprocessed data and other intermediary datasets for analysis.
• GNN-dir: This directory stores output from running the GP (Graph Pooling) algorithm.
• NTScore-dir: This directory stores NTScore output.

cd examples/stereo_seq_brain
ONTraC -d original_data.csv --preprocessing-dir stereo_seq_preprocessing_dir --GNN-dir stereo_seq_GNN --NTScore-dir stereo_seq_NTScore

We recommand running ONTraC on GPU, it may take much more time on your own laptop with CPU only.

All available parameter options are listed below.

Options:
  --version             show program's version number and exit
  -h, --help            show this help message and exit
  
  IO:
    -d DATASET, --dataset=DATASET
                        Original input dataset.
    --preprocessing-dir=PREPROCESSING_DIR
                        Directory for preprocessing outputs.
    --GNN-dir=GNN_DIR   Directory for the GNN output.
    --NTScore-dir=NTSCORE_DIR
                        Directory for the NTScore output
  
  Niche Network Construction:
    --n-cpu=N_CPU       Number of CPUs used for parallel computing. Default is
                        4.
    --n-neighbors=N_NEIGHBORS
                        Number of neighbors used for kNN graph construction.
                        Default is 50.
  
  Options for training:
    --device=DEVICE     Device for training. We support cpu and cuda now. Auto
                        select if not specified.
    --epochs=EPOCHS     Number of maximum epochs for training. Default is
                        1000.
    --patience=PATIENCE
                        Number of epochs wait for better result. Default is
                        100.
    --min-delta=MIN_DELTA
                        Minimum delta for better result. Default is 0.001
    --min-epochs=MIN_EPOCHS
                        Minimum number of epochs for training. Default is 50.
                        Set to 0 to disable.
    --batch-size=BATCH_SIZE
                        Batch size for training. Default is 0 for whole
                        dataset.
    -s SEED, --seed=SEED
                        Random seed for training. Default is random.
    --lr=LR             Learning rate for training. Default is 0.03.
    --hidden-feats=HIDDEN_FEATS
                        Number of hidden features. Default is 4.
    -k K, --k-clusters=K
                        Number of niche clusters. Default is 6.
    --spectral-loss-weight=SPECTRAL_LOSS_WEIGHT
                        Weight for spectral loss. Default is 0.3.
    --cluster-loss-weight=CLUSTER_LOSS_WEIGHT
                        Weight for cluster loss. Default is 0.1.
    --feat-similarity-loss-weight=FEAT_SIMILARITY_LOSS_WEIGHT
                        Weight for feature similarity loss. Default is 300.
    --assign-exponent=ASSIGN_EXPONENT
                        Exponent for assignment. Default is 0.3.

Output

The intermediate and final results are located in preprocessing-dir, GNN-dir, and NTScore-dir directories. Please see IO files explanation for detailed infromation.

Visulization

Please see post analysis tutorial.

Citation