cellshape 0.0.30

3D shape analysis using deep learning

3D single-cell shape analysis of cancer cells using geometric deep learning

This is a Python package for 3D cell shape features and classes using deep learning. Please refer to our preprint here.

cellshape is the main package which imports from sub-packages:

• cellshape-helper: Facilitates point cloud generation from 3D binary masks.
• cellshape-cloud: Implementations of graph-based autoencoders for shape representation learning on point cloud input data.
• cellshape-voxel: Implementations of 3D convolutional autoencoders for shape representation learning on voxel input data.
• cellshape-cluster: Implementation of deep embedded clustering to add to autoencoder models.

Installation and requirements

Dependencies

The software requires Python 3.7 or greater. The following are package dependencies that are installed automatically when cellshape is installed: PyTorch, pyntcloud, numpy, scikit-learn, tensorboard, tqdm (The full list is shown in the setup.py file). This repo makes extensive use of cellshape-cloud, cellshape-cluster, cellshape-helper, and cellshape-voxel. To reproduce our results in our paper, only cellshape-cloud, cellshape-cluster are needed.

To install

1. We recommend creating a new conda environment. In the terminal, run:

conda create --name cellshape-env python=3.8 -y
conda activate cellshape-env
pip install --upgrade pip

2. Install cellshape from pip. In the same terminal, run:

pip install cellshape

This should take ~5mins or less.

Hardware requirements

We have tested this software on an Ubuntu 20.04LTS and 18.04LTS with 128Gb RAM and NVIDIA Quadro RTX 6000 GPU.

Data availability and structure

Data availability

Datasets to reproduce our results in our paper are available here.

• SamplePointCloudData.zip contains a sample dataset of a point cloud of cells in order to test our code.
• FullData.zip contains 3 plates of point cloud representations of cells for several treatments. This data can be used to reproduce our results.
• Output.zip contains trained model weights and deep learning cell geometric features extracted using these trained models.
• BinaryCellMasks.zip contains a sample set of binary masks of cells which can be used as input to cellshape-helper to test our point cloud generation code.

Data structure

We suggest testing our code on the data contained in SamplePointCloudData.zip. This data is structured in the following way:

cellshapeSamplePointCloudDatset/
    small_data.csv
    Plate1/
        stacked_pointcloud/
            Binimetinib/
                0010_0120_accelerator_20210315_bakal01_erk_main_21-03-15_12-37-27.ply
                ...
            Blebbistatin/
            ...
    Plate2/
        stacked_pointcloud/
    Plate3/
        stacked_pointcloud/

This data structure is only necessary if wanting to use our data. If you would like to use your own dataset, you may structure it in any way as long as the extension of the point clouds are .ply. If using your own data structure, please define the parameter --dataset_type as "Other".

Usage

The following steps assume that one already has point cloud representations of cells or nuclei. If you need to generate point clouds from 3D binary masks, please go to cellshape-helper.

Downloading the dataset

We suggest testing our code on the data contained in SamplePointCloudData.zip. Please download the data and unzip the contents into a directory of your choice. We recommend doing this in your ~Documents/ folder. This is used as parameters in the steps below, so please remember where you download the data to. Downloading and unzipping the data can be done in the terminal. You might need to first install wget and unzip with apt-get (e.g. apt-get install wget).

1. Download the data into the ~/Documents/ folder with wget

cd ~/Documents
wget https://sandbox.zenodo.org/record/1080300/files/SamplePointCloudDataset.zip

2. Unzip the data with unzip:

unzip SamplePointCloudDataset.zip

This will create a directory called cellshapeSamplePointCloudDatset under your ~Documents/ folder, i.e. /home/USER/Documents/cellshapeSamplePointCloudDatset/ (USER will be different for you).

Training

The training procedure follows two steps:

1. Training the dynamic graph convolutional foldingnet (DFN) autoencoder to automatically learn shape features.
2. Adding the clustering layer to refine shape features and learn shape classes simultaneously.

Inference can be done after each step.

Our training functions are run through a command line interface with the command cellshape-train. For help on all command line options, run the following in the terminal:

cellshape-train -h

1. Train DFN autoencoder

The first step trains the autoencoder without the additional clustering layer. Run the following in the terminal. Remember to change the --cloud_dataset_path, --dataframe_path, and --output_dir parmaeters to be specific to your directories, if you have saved the data somewhere else. To test the code, we train for 5 epochs. First make sure you're in the directory where you downloaded the data to. If this is your `~/Documents/ folder, go into this:

cd ~/Documents

Then run the following:

cellshape-train \
--model_type "cloud" \
--pretrain "True" \
--train_type "pretrain" \
--cloud_dataset_path "./cellshapeSamplePointCloudDataset/" \
--dataset_type "SingleCell" \
--dataframe_path "./cellshapeSamplePointCloudDataset/small_data.csv" \
--output_dir "./cellshapeOutput/" \
--num_epochs_autoencoder 5 \
--encoder_type "dgcnn" \
--decoder_type "foldingnetbasic" \
--num_features 128 \

This step will create an output directory /home/USER/Documents/cellshapeOutput/ with the subfolders: nets, reports, and runs which contain the model weights, logged outputs, and tensorboard runs, respectively, for each experiment. Each experiment is named with the following convention {encoder_type}_{decoder_type}_{num_features}_{train_type}_{xxx}, where {xxx} is a counter. For example, if this was the first experiment you have run, the trained model weights will be saved to: /home/USER/Documents/cellshapeOutput/nets/dgcnn_foldingnetbasic_128_pretrained_001.pt. This path will be used in the next step for the --pretrained-path parameter.

2. Add clustering layer to refine shape features and learn shape classes simultaneously

The next step is to add the clustering layer to refine the model weights. As before, run the following in the terminal. Remember to change the --cloud_dataset_path, --dataframe_path, --output_dir, and --pretrained-path parmaeters to be specific to your directories. If you have followed the previous steps, then you will still be in the ~Documents/ path. In the same terminal, run:

cellshape-train \
--model_type "cloud" \
--train_type "DEC" \
--pretrain False \
--cloud_dataset_path "./cellshapeSamplePointCloudDataset/" \
--dataset_type "SingleCell" \
--dataframe_path "./cellshapeSamplePointCloudDataset/small_data.csv" \
--output_dir "./cellshapeOutput/" \
--num_features 128 \
--num_clusters 5 \
--pretrained_path "./cellshapeOutput/nets/dgcnn_foldingnetbasic_128_pretrained_001.pt" \

To monitor the training using Tensorboard, in a new terminal run:

pip install tensorboard
cd ~/Documents
tensorboard --logdir "./cellshapeOutput/runs/"

Alternatively, the training steps can be run sequentially through one command line

This would be to state that you would like to pretrain and that you want to train DEC.

cellshape-train \
--model_type "cloud" \
--train_type "DEC" \
--pretrain True \
--cloud_dataset_path "./cellshapeSamplePointCloudDataset/" \
--dataset_type "SingleCell" \
--dataframe_path "./cellshapeSamplePointCloudDataset/small_data.csv" \
--output_dir "./cellshapeOutput/" \
--num_features 128 \
--num_clusters 5 \

Inference

Example inference notebooks can be found in the docs/notebooks/ folder.

Issues

If you have any problems, please raise an issue here

Citation

@article{DeVries2022single,
	author = {Matt De Vries and Lucas Dent and Nathan Curry and Leo Rowe-Brown and Vicky Bousgouni and Adam Tyson and Christopher Dunsby and Chris Bakal},
	title = {3D single-cell shape analysis using geometric deep learning},
	elocation-id = {2022.06.17.496550},
	year = {2023},
	doi = {10.1101/2022.06.17.496550},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2023/03/27/2022.06.17.496550},
	eprint = {https://www.biorxiv.org/content/early/2023/03/27/2022.06.17.496550.full.pdf},
	journal = {bioRxiv}
}

References

[1] An Tao, 'Unsupervised Point Cloud Reconstruction for Classific Feature Learning', GitHub Repo, 2020