DanuserLab/openLCH

Live Cell Histology: Extracting latent features from label-free live cell images using Adversarial Autoencoders

Live Cell Histology

• Extracting latent features from label-free live cell images using Adversarial Autoencoders
  • Inspired by: https://github.com/AllenCellModeling/torch_integrated_cell/

Associated paper

• Interpretable deep learning uncovers cellular properties in label-free live cell images that are predictive of highly metastatic melanoma, Cell Systems, 2021, 12(7):733-747.e6, written by Assaf Zaritsky, Andrew R Jamieson, Erik S Welf, Andres Nevarez, Justin Cillay, Ugur Eskiocak, Brandi L Cantarel, Gaudenz Danuser.

Setup & Running Source Code

Developed and tested on Red Hat Linux 7.

Installation Steps

• also see more detailed information here

Containers

• Set-up compute environment with containers:
  • Install Singularity
  • Tested with Singularity 3.5.3
  • Need CUDA 8.0+ compatible GPU and drivers (e.g. P100)
  • Pull Singularity container .sif image file from Singularity Hub
    • alternatively, a copy can be found here
    • Definition file provided in this repo
  • singularity pull shub://andrewjUTSW/openLCH:latest
  • Test GPU singularity exec --nv --cleanenv ./openLCH_latest.sif nvidia-smi

Download and Prepare Example Data

• Download 2000 image sample data set
  • curl https://cloud.biohpc.swmed.edu/index.php/s/FqZSqoKfHii6ony/download --output sample2000.tar.gz
  • unzip data
    • tar xvzf ./sample2000.tar.gz
  • Included are a random sample set of 256x256 cell images
  • Create image file list ls `pwd`/data2/*.png > imagePathList.txt
  • Full data set provided here:
  • Download previously trained autoencoder .t7 file
    • curl https://cloud.biohpc.swmed.edu/index.php/s/YAQQtpwTX2NKS89/download --output autoencoder_eval_56zTRAINED.t7

Run Provided Example Scripts

Linear interpolation between two reconstructed cell images using previously trained AAE

• interp_LatentSpace_LCH_MD_single_2.lua

	singularity exec --nv --cleanenv openLCH_latest.sif /bin/bash -c 'cd ./code; \
	LCH_PATH=YOUR_CODE_PATH_HERE; \
	th -i ./interp_LatentSpace_LCH_MD_single_2.lua \
	-imPathFile $LCH_PATH/imagePathList.txt \
	-autoencoder $LCH_PATH/autoencoder_eval_56zTRAINED.t7 \
	-outDir $LCH_PATH/output/interpOut/ \
	-gpu 1 \
	-img1 51 \
	-img2 81'

Train new AAE

• run_mainLCH_AAE_Train_2.lua

	singularity exec --nv --cleanenv openLCH_latest.sif /bin/bash -c 'cd ./code; \
	LCH_PATH=YOUR_CODE_PATH_HERE; \
	th ./run_mainLCH_AAE_Train_2.lua \
	-modelname AAEconv_CLEAN \
	-nLatentDims 56 \
	-imsize 256 \
	-imPathFile $LCH_PATH/imagePathList.txt \
	-savedir $LCH_PATH/outputNew/ \
	-epochs 100 \
	-gpu 1' 

Extract latent embeddings

• call_DynComputeEmbeddingsRobust_2.lua

	singularity exec --nv --cleanenv openLCH_latest.sif /bin/bash -c 'cd ./code; \
	LCH_PATH=YOUR_CODE_PATH_HERE; \
	th ./call_DynComputeEmbeddingsRobust_2.lua \
	-autoencoder $LCH_PATH/outputNew/autoencoder_eval.t7 \
	-imsize 256 \
	-dataProvider DynDataProviderRobust_2 \
	-imPathFile $LCH_PATH/imagePathList.txt \
	-gpu 2 \
	-embeddingFile $LCH_PATH/outputNew/embeddings_sampleTest.csv'

Explore latent space by shifting embedding vector values (one dimension at a time) of an input cell image and reconstructing shifted synthetic cell images

• exploreZ_LatentSpace_LCH_single_2.lua

	singularity exec --nv openLCH_latest.sif /bin/bash -c 'cd ./code; \
	LCH_PATH=YOUR_CODE_PATH_HERE; \
	th -i ./exploreZ_LatentSpace_LCH_single_2.lua \
	-imPathFile $LCH_PATH/imagePathList.txt \
	-autoencoder $LCH_PATH/autoencoder_eval_56zTRAINED.t7 \
	-outDir $LCH_PATH/outputOrig/zExploreOut \
	-img1 10 \
	-uR 1 \
	-numSteps 6'

Reconstruct images from latent codes

• zLatent2ReconBatchLCH_2.lua

	singularity exec --nv --cleanenv openLCH_latest.sif /bin/bash -c 'cd ./code; \
	LCH_PATH=YOUR_CODE_PATH_HERE; \
	th -i ./zLatent2ReconBatchLCH_2.lua \
	-autoencoder $LCH_PATH/autoencoder_eval_56zTRAINED.t7 \
	-zLatentFile $LCH_PATH/outputNew/embeddings_sampleTest.csv \
	-reconPath $LCH_PATH/outputNew/zRecon/ \
	-nLatentDims 56'

Citation

@article {Zaritskyj.cels.2021.05.003,
  author = {Zaritsky, Assaf and Jamieson, Andrew R. and Welf, Erik S. and Nevarez, Andres and Cillay, Justin and Eskiocak, Ugur and Cantarel, Brandi L. and Danuser, Gaudenz},
  title = {Interpretable deep learning uncovers cellular properties in label-free live cell images that are predictive of highly metastatic melanoma},
  journal = {Cell Systems},
  year = {2021},
  volume = {12},
  number = {7},
  pages = {733-747.e6},
  month = {Jul},
  doi = {10.1016/j.cels.2021.05.003},
  url = {https://doi.org/10.1016/j.cels.2021.05.003},
  eprint = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8353662/}
}

Danuser Lab Links

Danuser Lab Website

Software Links