Phenotypic Profiling

Scientists can now routinely extract high-content, high-dimensional cell morphology representations from microscopy images. However, these cell morphology features currently represent a "hidden code" that must be further interpreted in order to understand and assign biological meaning.

Here, we hypothesize that nuclear morphology can provide a window into single cell phenotype that can be broadly applied across cell types, treatments, and experimental designs (staining, microscopy acquisition parameters, etc.). We test this hypothesis by training machine learning models to predict specific phenotypes from easily-accessible and reproducible single-cell morphology representations.

Specifically, we use publicly-available data from the MitoCheck consortium, which includes 2,916 single-cells labeled with one of 15 different phenotypes, to train a multiclass logistic regression model to predict phenotype. We extracted CellProfiler and DeepProfiler features from all MitoCheck nuclei. See https://github.com/WayScience/mitocheck_data for details on how we accessed and processed these data.

We focused on assessing the generalizability of this approach to predict phenotype in new datasets not seen during model training. We tested generalizability performance in two scenarios: (1) Leave one image out analysis and (2) Predicting single-cell phenotype in the JUMP-CP Pilot data.

Figure 1 describes the dataset and our approach for training and evaluating our model.

Figure 1. Dataset and analysis approach. (A) Single-cell counts per labeled phenotype stratified by phenotype category. The labeled MitoCheck dataset included a total of 2,916 single nuclei. The original dataset contained labels for 16 classes, but we have removed “folded” because of low counts. (B) Our analysis pipeline incorporated image analysis, image-based profiling, and machine learning. We also assess model generalizability through a leave-one-image-out analysis and applying our models to the Joint Undertaking in Morphological Profiling Cell Painting (JUMP-CP) pilot dataset.

Environment Setup

Perform the following steps to set up the phenotypic_profiling environment necessary for processing data in this repository.

Step 1: Create Phenotypic Profiling Environment

# Run this command to create the conda environment for phenotypic profiling

conda env create -f phenotypic_profiling_env.yml

Step 2: Activate Phenotypic Profiling Environment

# Run this command to activate the conda environment for phenotypic profiling

conda activate phenotypic_profiling

Repository Structure:

The repository structure is as follows:

Order	Module	Description
0.download_data	Download training data	Download labeled single-cell dataset from mitocheck_data
1.split_data	Create data subsets	Create training and testing data subsets
2.train_model	Train model	Train ML models on combinations of features, data subsets, balance types, model types
3.evaluate_model	Evaluate model	Evaluate ML models on all data subsets
4.interpret_model	Interpret model	Interpret ML model coefficients
5.validate_model	Validate model	Validate ML models on other datasets
6.single_cell_images	Single cell images	View single cell images and model interpretation
7.figures	Figures	Create paper-worthy figures

Data

Specific data download/preprocessing instructions are available at: https://github.com/WayScience/mitocheck_data. This repository downloads labeled single-cell data from a specific version of the mitocheck_data repository. For more information see 0.download_data/.

We use the following 2 datasets from the mitocheck_data repository:

• ic: single-cell nuclei features extracted after performing illumination correction on images
• no_ic: single-cell nuclei features extracted without performing illumination correction on images

Supplementary Table 1 - Full list of JUMP-CP phenotype enrichment

We report the top 100 most enriched treatments per phenotype in Supplementary Table 1 of our paper. See jump_compare_cell_types_and_time_across_phenotypes.tsv.gz for the full list.

Machine Learning Models

We use Scikit-learn (sklearn) for data manipulation, model training, and model evaluation. Pedregosa et al., JMLR 12, pp. 2825-2830, 2011 describe scikit-learn as a machine learning library for Python. Its ease of implementation in a pipeline makes it ideal for our use case.

We consistently use the following parameters with many sklearn functions:

• n_jobs=-1: Use all CPU cores in parallel when completing a task.
• random_state=0: Use seed 0 when shuffling data or generating random numbers. This allows "random" sklearn operations to have consist results. We also use np.random.seed(0) to make "random" numpy operations have consistent results.

We use seaborn for data visualization. Waskom, M.L., 2021 describe seaborn as a library for making statisical graphics in python.

All parts of the machine learning pipeline are completed with the following feature types:

• CP: Use only CellProfiler features from MitoCheck labeled cells
• CP_zernike_only: Use only CellProfiler Zernike shape features from MitoCheck labeled cells
• CP_areashape_only: Use only CellProfiler AreaShape features from MitoCheck labeled cells
• DP: Use only DeepProfiler features from MitoCheck labeled cells
• CP_and_DP: Use CellProfiler and DeepProfiler features from MitoCheck labeled cells

See MitoCheck_data for more information on feature types and how they are extracted from MitoCheck labeled cells.

We create two versions of the same machine learning model:

• final: Model fit using real feature values
• shuffled_baseline: Model fit using shuffled training data (to create a suitable baseline comparison for the final model during evaluation)

Intermediate Data

Throughout this repository, we store intermediate .tsv data in tidy long format, a standardized data structure (see Tidy Data by Hadley Wickham for more details). This data structure makes later analysis easier.

Some intermediate data used in this repository are too large to be stored on GitHub. These intermediate data are available on the Way Lab Zenodo page.

Reproducibility

Specific code and steps used are available within each module folder.

The Way Lab always strives for readable, reproducible computational biology analyses and workflows. If you struggle to understand or reproduce anything in this repository please file an issue!