This is the repository for "Computational analysis of peripheral blood smears detects disease-associated cytomorphologies" (available on Nature Communications). In this work, we use the whole blood slides of >300 individuals with myelodyplastic syndromes and anaemias and use them to develop a method that is capable of automatically detecting cells in whole blood slides, predicting a disease and retrieving examples of cells which are relevant for each classification.

• Python
• Snakemake
• R (analysis and plotting)
• conda (recommended for virtual environment management)

Each folder requires different packages and a standard environment manager is advised (be it conda or virtualenv). Within each folder, a specification for the required packages is provided.

In each folder a short description is provided delineating the required packages to run each analysis.

1. pipeline_tf2 - contains Haemorasis, the pipeline for WBC and RBC detection and characterisation from WBS. Also available as a Docker container in Docker hub. The content of this folder is identical to the code in a separate repository which was created for convenience: Haemorasis
2. mil-comori - contains the code to train and run Morphotype analysis on the output from Haemorasis
3. analysis-plotting - contains the code to analyse and plot the results from the previous processes

It should be noted that, due to the high data volume required for this work, we have chose to make the necessary inputs and outputs for mil-comori, as well as the necessary inputs for analysis-plotting, available through a Figshare project. We provide simplified instructions on how to reproduce the results from the Morhpotype analysis and the downstream statistical analysis below. In any case, README files explaining how to run each analysis are contained in each folder.

This was developed and tested using Python 3.6.8 and on 8GB RAM on CentOS Linux 8 (kernel: Linux 4.18.0-240.22.1.el8_3.x86_64), and for Haemorasis, the blood cell detection pipeline (pipeline_tf2), we used NVidia Quadro M6000 GPUs. Approximately 3GB of free disk space are required to run both the Morphotype analysis and the downstream statistical analysis of the results (if you want to run only the latter approximately only 200MB of disk space are required). We also recommend having at least 8GB of RAM and a CentOS-based machine, and using conda to manage packages as this greatly simplifies dependencies.

1. Clone this Github repository to your local machine
   • git clone https://github.com/josegcpa/wbs-prediction.git
2. Download and unzip the necessary data from Figshare using the download script
   • sh download-and-unzip-data.sh
   • sh download-and-unzip-analysis-plotting-only.sh if you wish to perform only step 4
3. (optional for those wishing to perform the Morphotype analysis:) enter the Morphotype analysis directory (cd mil-comori) and follow the instructions presented there in the README.md file.
   • Please note that one of the steps, that of aggregating cells from the Haemorasis output and calculating cell proportions (using mil-comori/Snakefile-cell-collections which, in essence, uses the features for all cells accross all slides and calculates their respective morphotype classification, retrieves their images for later inspection and calculates the proportions of morphotypes in each PBS) has been occluded from these instructions as it requires massive amounts of compute power and storage. To ensure that the rest of these analyses still functions correctly, we have made the necessary output available together with the remaining downloads
4. After having performed the analysis in step 3., enter the statistical analysis and figure generation folder (cd analysis-plotting) and follow the instructions presented there