machine-learning-IFC

This repository is a tutorial for analyzing imaging flow cytometry (IFC) data with machine learning.
The step by step guide also contains examples where you can try out our workflow to get started, before beginning to analyze your own data. The workflow is based on the following articles (please cite when using the workflow):

• Hennig et al. An open-source solution for advanced imaging flow cytometry data analysis using machine learning. Methods 112, 201 (2017)
• Blasi et al. Label-free cell cycle analysis for high-throughput imaging flow cytometry. Nature Communications 7, 10256 (2016)

For our deep learning workflow, please see https://github.com/broadinstitute/deepometry

Step 1: extracting cell subpopulations (IDEAS)

Gating in the software IDEAS enables to define different subpuplations of cells (for example cancer cells, normal cells, treated cells). In the menu in IDEAS, choose extract subpopulation and generate a .cif file for each subpopulation. The .cif files will be processed in the next step.

Step 2: Generate image montages (python)

A montage is a collection of (e.g.) 900 cells, arranged in a 30x30 grid. The montages are automatically generated from a .cif file by a stitching script.
Input: example.cif file
Output: For each of the max. 12 channels, montages are created.
ch1.tif
ch2.tif
...
ch12.tif

Step 3: Extract features (CellProfiler)

How to identify the cells in the image montages and extract hundreds of features with CellProfiler?
This repository contains a CellProfiler pipeline IFC_template.cppipe. Start CellProfiler, import the pipeline and drag & drop your image montages into CellProfiler. You probably need to adapt the pipeline to your image assay.
Download Cellprofiler: http://cellprofiler.org/releases/
For help on how to improve or tweak your pipeline, see the CellProfiler forum.

Step 4: Machine learning (python)

A machine learning script in python is provided in this repository, which may serve as a starting point for your high-content image data analysis and exploration.

Examples & Tutorial

CellProfiler for IFC

This is an example of step 3 of the workflow. The goal in our example is to predict the cell cycle phase of the Jurkat cells. The examples folder contains a quick example with a small data set, which we hope may serve to become familiar with our workflow before analyzing your own data. The example data is also suitable for workshops & tutorials. The CellProfiler pipeline finishes in ~4 Minutes on a typical 4 core laptop.

1. Download the files from this repository
2. Install CellProfiler
3. Open the file Step3_CellProfiler.cppipe in CellProfiler. Drag & drop the folder Step2_output_tiled_tifs into CellProfiler
4. Hit the "analyze images" button

While this is a small data set suitable for tutorials and training purposes, the full data set with all montage images from Hennig et al. Methods 112, 201 (2017) is available on http://cellprofiler.org/imagingflowcytometry/index.html

Machine Learning for IFC

This is an example for step 4 of the workflow. The machine learning python script Step4_machine_learning.py reads the output data exported from CellProfiler. In this example, we will load all measurements taken from ca. 33.000 Jurkat cells stored in the .txt files in the folder Step3_AllData into python. The script then uses and compares a RandomForests and NaiveBayes classifier to predict the cell cycle phase of the Jurkat cells. It also performs feature selection, i.e., ranks the most important/informative features. The results are written to the machine learning output directory. For this data set, machine learning takes (1) a few seconds for NaiveBayes (2) several minutes for RandomForests including feature selection on a typical laptop.

1. Open the file Step4_machine_learning.py in python (install required python packages first)
2. Adjust the working directory
3. Run the script

FAQ

Q: I have an IFC experiment with several fluorescent markers. However, the Blasi paper focuses on a label-free analysis method (i.e., without any markers). Can I use the machine learning workflow including the marker channels and analyze correlations/colocalization between the channels?
A: Yes, there is already one marker image defined in the CellProfiler pipeline, simply add further marker images in the measurement modules in CellProfiler. For colocalization assays, there is an example pipeline on the CellProfiler website, or post your question on the CellProfiler forum.

Q: The object segmentation ("finding the cell objects") does not work well with my pipeline.
A: If you have a nuclear stain, you can adapt the pipeline accordingly (identify primary objects on the nuclei, and identify secondary objects on the cells with nuclei as guiding objects). For help on how to improve or tweak your pipeline, see the CellProfiler forum.

Q: My image montages are empty (all black)
A: Try the following

• Many image viewers don't scale the image range. Open the montages in CellProfiler or ImageJ
• Look up in IDEAS in which channels images have been recorded, the other channels are empty and the respective montages can be discarded.