We aim to characterize cell morphology signatures of neurofibromin in Schwann cells. We applied a modified Cell Painting assay on two isogenic Schwann cell lines, one of the wildtype genotype (NF1+/+) and one of the null genotype (NF1-/-), both from the same patient. The modified assay stains for four organelles: nuclei, endoplasmic reticulum, mitochondria, and actin. We applied CellProfiler pipelines to perform quality control, illumination correction, segmentation, and feature extraction.
Image montage of dataset and analysis workflow. (A) Example image montages of the Cell Painting channels and composite image (all channels overlayed) for each NF1 genotype. The scale bar represents 25 μM. (B) The workflow of our analysis pipeline demonstrates the steps taken from image analysis to machine learning
We segmented 22,585 wild-type (WT) and null cells across three plates and utilized 907 significant morphology features representing various organelle shapes and intensity patterns.
We trained a logistic regression binary classifier to predict the NF1 genotype of single cells. The model shows high performance with with accuracy of 0.85 and 0.80 for the training and testing data splits respectively.
Logistic regression model predicts genotype with high performance. (A) Precision-recall curves comparing the final model applied to shuffled (dashed line) and non-shuffled data (solid line). Applying the model to a shuffled dataset performed worse than the non-shuffled data, demonstrating a biological signal between genotypes. (B) Confusion matrices from the training and testing data splits show higher performance across genotypes in non-shuffled data compared to the shuffled data. (C) Accuracy scores show high performance classifying cells with both genotypes from the training and testing data splits compared to shuffled data. Both panels B and C visualize the results from the optimized model.
The machine learning model learned from a total of 907 morphology features and assigned weights, or importance scores, per feature. This combination of morphology features represents a high-dimensional signature of how the NF1 genotype influences cell morphology in otherwise isogenic Schwann cells.
Significant cell morphology features come from multiple kinds of measurement and across organelles. (A) Top absolute value coefficients demonstrate which measurements/organelles/compartments are important in determining the differences between NF1 genotypes. Not one feature is important in this prediction; many features in our models are important for making accurate genotype predictions. The red boxes indicate the two most important features based on absolute value. (B) Min/max image montage of the top feature from panel A, representing the highest weighted feature for predicting null NF1 genotype. This shows the visual difference between null cells with the top values of radial distribution in the F-actin cytoskeleton compared to the wildtype cells with the lowest values. The title is the exact CellProfiler feature name. (C) Min/max image montage of the second top features from panel A, representing the top feature for predicting wildtype NF1 genotype. This montage shows the visual difference between the wildtype cells with high correlations between the nucleus and ER compared to the null cells with the lowest values. Both the image montages from panels B and C show composite single-cell images where the colors represent each organelle: blue = nucleus, green = ER, magenta = mitochondria, and red = actin.
We look to improve upon this preliminary model in the future. We aim to generate further data which includes the heterozygous genotype (NF1+/-). AS well, we plan to apply an improved model to large-scale drug screens to capture candidate drugs that make NF1 patient cells look healthy.
NOTE: All image analysis and image-based profiling pipelines can be found in the nf1_cellpainting_data repository.
This analysis is categorized as follows:
| Module | Purpose | Description |
|---|---|---|
| 0.data_analysis | Preliminary analysis of NF1 data | We find interesting patterns in the NF1 cell morphology data. |
| 1.train_models | Train NF1 models | Optimize the NF1 model by training multiple models with a random search. |
| 2.evaluate_models | Evaluate final NF1 model | After training the final NF1 model, we extract the model performance and model feature importances. |
| 3.figures | Generate manuscript figures | We interpret the results from the model by generating figures to include in the manuscript. |
git clone https://github.com/WayScience/NF1_SchwannCell_data_analysis.gitgit submodule update --init --recursiveThere are two different environments used in this analysis:
- environment.yml: A Python-based environment meant for performing model training, evaluation, and image montage generation.
- figure_environment.yml: An R-based environment meant for generating figures.
conda env create -f environment.yml
conda env create -f figure_environment.yml