This README outlines the steps required to reproduce the approach presented in the manuscript Rappez et al. DeepCycle reconstructs a cyclic cell cycle trajectory from unsegmented cell images using convolutional neural networks. Molecular Systems Biology, 2020. These instructions show how to run the model on the data used in the paper and on new unseen data.

keras, UMAP, cv2, albumentations, classification_models
Install customized version of SOMPY

1. Download data from EBI BioStudies repository and deep learning models from our EMBL hosting
2. Unzip to data/Timelapse_2019 folder preserving directory structure. You will have:

   root|
       |-data|
       |     |-Timelapse_2019|
       |                     |-BF/
       |                     |-Cy3/
       |                     |-DAPI/
       |                     |-GFP/
       |                     |-curated_tracks.csv
       |                     |- ...
       |-src/
       |...
   

Alternatively, DeepCycle can be run on newly generated data. For that, prepare the data as follows:

1. Organize the live-imaging microsocpy. Each channel of every timepoint are stored independently as:

   root|
       |-data|
       |     |-New_timelapse_data|
       |                         |-BF/
       |                         |-Cy3/
       |                         |-DAPI/
       |                         |-GFP/
   

2. Run TrackMate FIJI plugin on the nuclear staining channel (DAPI in our case). The output should be called Spots in tracks statistics.csv.
3. Run src/TrackMate_filter.py with the correct paths (see comments in file) to manually curate the tracks with one or two divisions.

1. cd src
2. Prepare the data:
   python data_prepare.py
   • Cleans and removes unnecessary columns. Stores as statistics_clean.csv in data/Timelapse_2019 dir
   • Aligns the curated tracks based on division events, calculates mean intensities track/frame wise. Stores as intensities.csv
   • Calculates intensity statistics and adds virtual class 1-4 to each tracked cell. Resulting data to be stored in statistics_mean_std.csv
3. Train the model:
   python model_train.py
   Trains the model on curated tracks (less double division tracks) using double division tracks as validation set. Saves best models in checkpoints dir
4. Generate cell descriptors with checkpoint.r34.sz48.03-0.73.hdf5 as default model:
   • from validation set (double division tracks) only:
     python encode.py --mode encode_val
   • from all available tracks:
     python encode.py --mode encode_all
     Descriptors are saved in descriptors.r34.sz48.pkl and descriptors_all.r34.sz48.pkl in data/Timelapse_2019 dir.
5. Generate embeddings for all dataset. Compute intense, consider using supplied embeddings_preds_all_batch<i>.npz instead:
   python all_cells_prediction.py
6. cd ..
7. start jupyter notebook and open timelapse_projection2019.ipynb