We have developed a code for learning decision tree classifiers, predicting manually-annotated endodermal morphotypes at 96 hrs, from morphological features (derived from brightfield images) and marker expression features (derived from fluorescence GFP and RFP images).

The code is deposited on both GitHub and Zenodo, and available from the following links:

https://github.com/ChenSchiff/Dev_Cell_paper

https://doi.org/10.5281/zenodo.11181735

The code learns 500 decision trees from the input data set (using a bootstrap train/test split approach). It then visualizes the statistics of one-parameter (top tree node) and three-parameter (first and second tree levels) frequencies, as either bar graph or heatmap, respectively. In computing these frequencies, only learned trees with test-set accuracy above a set threshold are taken into consideration.

The code is at the basis of the machine learning section of our preprint Farag et. al. , (see Figure 2 in the preprint)

https://www.biorxiv.org/content/10.1101/2023.02.07.527329v1

The morphological and expression features used by the code were manually curated from the time series dataset.