(59s) Investigating the Effects of Tunable Experimental Parameters on hiPSC-Cms Maturation Via Clustering Techniques | AIChE

(59s) Investigating the Effects of Tunable Experimental Parameters on hiPSC-Cms Maturation Via Clustering Techniques

Authors 

Cardiovascular disease is the leading cause of death worldwide (Ahmed et al., 2020). It damages cardiomyocytes irreversibly by causing a series of detrimental effects that results in the loss of more than one billion cardiomyocytes (heart cells). Cardiomyocytes (CMs) can be produced in vitro from human-induced pluripotent stem cells (hiPSCs) with high yield and purity (Ahmed et al., 2020) and have great potential for biomedical research. However, the cardiomyocytes from hiPSCs are immature. Mature CMs can be used in regenerative medicine and cardiovascular research. The immaturity of the hiPSC-derived CMs limits their utility in drug development, disease research, and regenerative therapies (Karbassi et al., 2020). Developing protocols to produce mature CMs resembling adult human cardiomyocytes is essential for their applications in cardiac regenerative therapies.

Eight different clustering methods were applied to day 60 hiPSC cardiac differentiation data to determine whether maturity-relevant features of cells could be used to cluster the cells into mature or immature CMs and to identify the relationship between maturity-relevant features and experimental parameters. The CMs were produced through hiPSC hydrogel encapsulation and differentiation within this 3D-engineered tissue microenvironment (Finklea et al., 2021; Kerscher et al., 2017). The data was collected from 10 different batches, each with unique experimental parameters, such as initial axial ratio (AR), cell diameter, and cell concentration. The data used for the clustering consists of morphological features of the cells, such as cell eccentricity, area, elongation, circularity index, and sarcomere properties, such as sarcomere length, organization score, and orientation index.

K-means clustering, Gaussian mixture models (GMM), agglomerative clustering, DBSCAN, OPTICS, affinity propagation, BIRCH, and spectral clustering (Fabian Pedregosa) were considered for clustering. The results of K-means and GMM were chosen for further analysis since they have the highest Silhouette score of 0.22 and 0.21, respectively. The number of clusters was fixed to two as our goal is to group the cells into either mature or immature clusters. The K-means algorithm clustered the cells with an eccentricity of 0 – 0.7 and elongation of 0 – 2 in one cluster and the remaining in another. The GMM yielded similar clusters to K-means results. We labeled the cluster with an eccentricity of 0.7 – 1 and elongation of 2-10 mature and studied the relationship between the percentage of mature cells in each batch and tunable experimental parameters. The cosine similarity between the percentage of mature CMs and the experimental parameters confirms that there is a relationship between them (Fig 1).

The relationship between the percentage of mature CMs and the tunable experimental parameters was found using ALAMO (Automatic Learning of Algebraic MOdels). ALAMO (Sahinidis, 2016) uses a linear summation of non-linear transformations of the input variables to predict the output. The non-linear transformations considered were polynomial, exponential, logarithmic, ratio, and trigonometric functions. The relationship between the percentage of mature CMS in a batch and the experimental parameters yielded by ALAMO is given in Eq. (1). The adjusted R2 for Eq. (1) is 0.845.

y = - 0.45 exp(X2) + 1.3 cos(X3)

where y is the percentage of mature CMs, X2 is cell diameter, and X3 is cell concentration.

Fig – 1: Scatter plot between the percentage of mature hiPSC-CMs in each batch vs. the tunable experimental parameters considered. – cosine similarity between the percentage of mature CMs and the experimental parameters.

In conclusion, the K-means algorithm clustered the cells based on eccentricity and elongation, which suggests the significance of these features in the hiPSC-CMs maturation. ALAMO was used to develop a regression model to predict the percentage of mature cells that can be expected in an experimental batch given cell diameter and concentration. Future studies will focus on the experimental validation of these findings and refining the models.

References:

Ahmed, R. E., Anzai, T., Chanthra, N., & Uosaki, H. (2020). A Brief Review of Current Maturation Methods for Human Induced Pluripotent Stem Cells-Derived Cardiomyocytes. Front Cell Dev Biol, 8, 178.

Fabian Pedregosa, G. V., Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, Édouard Duchesnay;. Scikit-learn: Machine Learning in Python. https://scikit-learn.org/stable/about.html#citing-scikit-learn

Finklea, F. B., Tian, Y., Kerscher, P., Seeto, W. J., Ellis, M. E., & Lipke, E. A. (2021). Engineered cardiac tissue microsphere production through direct differentiation of hydrogel-encapsulated human pluripotent stem cells. Biomaterials, 274, 120818.

Karbassi, E., Fenix, A., Marchiano, S., Muraoka, N., Nakamura, K., Yang, X., & Murry, C. E. (2020). Cardiomyocyte maturation: advances in knowledge and implications for regenerative medicine. Nat Rev Cardiol, 17(6), 341-359.

Kerscher, P., Kaczmarek, J. A., Head, S. E., Ellis, M. E., Seeto, W. J., Kim, J., Bhattacharya, S., Suppiramaniam, V., & Lipke, E. A. (2017). Direct Production of Human Cardiac Tissues by Pluripotent Stem Cell Encapsulation in Gelatin Methacryloyl. ACS Biomater Sci Eng, 3(8), 1499-1509.

Sahinidis, N. (2016). The ALAMO approach to machine learning. In (pp. 2410). Elsevier.