Neural network approaches, including use of topological data analysis, enhance classification of human induced pluripotent stem cell colonies by treatment condition
Alexander Ruys de Perez, Paul E. Anderson, Elena S. Dimitrova, Melissa L. Kemp
Abstract
Understanding how stem cells organize to form early tissue layers remains an important open question in developmental biology. Helpful in understanding this process are biomarkers or features that signal when a significant transition or decision occurs. We show such features from the spatial layout of the cells in a colony are sufficient to train neural networks to classify stem cell colonies according to differentiation protocol treatments each colony has received. We use topological data analysis to derive input information about the cells’ positions to a four-layer feedforward neural network.
Introduction
Organoids and microphysiological systems generated from human induced pluripotent stem cells (hiPSCs) hold promise for developing in vitro assays that can be used for evaluating therapeutics, toxicological screening, and regenerative medicine. Furthermore, the way in which stem cells organize themselves into more specialized tissues under various culture conditions provides critical insight into morphogenesis, the dynamic formation of organ systems.
Materials and method
Persistent homology
The mathematical tool we use in preparing the colony data is persistent homology. Here, we will present a simplified version for analysis of 2-dimensional data. Those wishing to learn more about the generalized approach should see Edelsbrunner et al. [14] Informally, the job of persistent homology is to provide quantitative data for properties of data sets that are usually described qualitatively.
Results
Both the topological data-using network and the standard image classifier succeed in classifying images
In order to investigate the potential utility of TDA in detecting changes in stem cell aggregates associated with culturing protocols, we performed a comparative analysis between a standard image classifying neural network (ResNet) and a simple feedforward neural network (TDANet) which used our topologically-derived feature set. Frames from a previously published study [10] of time-lapsed microscopy of iPSC aggregates undergoing differentiation from 5 experimental conditions over 48 hours were analyzed for 0th order and 1st order homology; this information was used to train TDANet as described in Materials & Methods.
Discussion
We investigate the potential for neural networks to accurately classify stem cell fates using morphological data. To do so, we take pluripotent stem cell colonies given one of five differentiation protocol treatments, and ask a neural network to guess the correct protocol. We compare two different network models, each with its own type of information used as input.
Acknowledgments
We thank David Joy and the Todd McDevitt lab for their sharing of the colony images and data. We also thank Curly Zhao for help with questions about PyTorch.
Citation: de Perez AR, Anderson PE, Dimitrova ES, Kemp ML (2025) Neural network approaches, including use of topological data analysis, enhance classification of human induced pluripotent stem cell colonies by treatment condition. PLoS Comput Biol 21(7): e1012801. https://doi.org/10.1371/journal.pcbi.1012801
Editor: Serdar Bozdag, University of North Texas, UNITED STATES OF AMERICA
Received: January 15, 2025; Accepted: June 2, 2025; Published: July 11, 2025
Copyright: © 2025 Perez et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The code used to create and train the neural networks, as well as process the topological data, is available at https://github.com/aruysdeperez/TDANet.git.
Funding: E.D. and M.K. were supported by the NSF-Simons Southeast Center for Mathematics and Biology through the grant DMS1764406 from the National Science Foundation (https://www.nsf.gov/) and the grant SFARI 594594 from the Simons Foundation (https://www.simonsfoundation.org/). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
