Design Optimization of Functional Cardiac Organoids Enabled by Machine Learning Techniques

Introduction: Stem cell derived cardiac organoids have been proven to be a robust animal free model for disease modeling, regenerative medicine and drug screening. These cardiac organoids demonstrated great potential in recapitulating both morphogenic events and functional outputs. However, there is still a lack of optimization pipeline that can evaluate the design of organoid systems under controllable engineering parameters. Therefore, we need to create reproducible methodologies for understanding and quantifying how engineering inputs can be used to guide and alter the physiological properties of cardiac organoids. Understanding organoid design principle necessitates a comprehensive analysis of multiple variables concurrently, thereby elevating the dimensionality of data analysis and visualization. Utilizing artificial intelligence and machine learning techniques, we can quickly analyze the total functional output of a single organoid and evaluate the impact of engineering inputs on organoid physiological outputs. In this work, we used micropatterning technique to create a cardiac organoid library (230 cardiac organoids) with 7 geometric designs from circles, rectangles, to stars with varied aspect ratios. We introduced a ML-based workflow to analyze this organoid library with 10 unique cardiac function variables from contractile motion and calcium transient analysis. Different AI-driven approaches, including manifold learning, unsupervised data clustering, and ensemble learning, were applied to not only understand functional heterogeneity from single cardiac organoids, but also provide mechanistic insights of physiological enhancement associated with different geometric designs.

Materials and

Methods: Utilizing PEG-Based micropatterning, we generate a cardiac organoid library of 230 organoids with 7 unique geometric designs (Circle 200, Circle 600, Circle 1000, Rectangle 1:1, Rectangle 1:4, Star 1:1, and Star 1:4). Contractile motion and calcium transient were recorded and analyzed to generate 10 quantifiable funcitonal parameters for each organoid, including beat rate, contraction velocity (Cont_Vel), relaxation velocity (Relax_Vel), contraction-relaxation interval (Cont-Relax Interval), fluorescent amplitude (Ca Amp), calcium rising time (t0), calcium decay time (t50 and t75), pulse duration (FWHM), and ratio between the projected area of GCaMP fluorescence and the area of original 2D pattern design (Area Ratio). The unsupervised manifold learning techniques t-SNE, UMAP, PACMAP and TriMAP were utilized to embed all 10 parameters into a 2D space. This allowed for all organoids and their functional parameters to be visualized simultaneously for a systematic comparison. Next, k-means clustering was conducted to regroup the organoids based on their functional similarity for each dimensionality reduction 2D embedding. Furthermore, we developed a network-based organoid refinement process that can reduce the organoid heterogeneity and feature the organoid representatives for each geometric design. Last, supervised ensemble learning techniques of extreme gradient boosting and random forest were then utilized to investigate which physiological parameters were in fact causing the organoid clustering.

Results, Conclusions, and Discussions: The unsupervised learning techniques proved to be rather effective in grouping similar organoids together based on their geometric design. Distinct separation was observed from small circle-200 organoids from the other larger organoids. There is an obvious separation of star (1:1 and 1:4) organoids away from circle (600 and 1000) and rectangle 1:4 organoids. PaCMAP algorithm that better preserves global higher dimensional data structures illustrated a gradient change of organoid functions cross different geometries: starting from the circle-200 organoids to the star 1:1 organoids, and then rectangle 1:4 organoids, and branching into circle-1000 and circle-600 organoids. For each 2D manifold embedding, we clustered the organoids into 7 clusters using the k-means clustering algorithm, which generated a total of 28 organoid clusters from 4 different algorithms that re-grouped shape-patterned cardiac organoids based on functional similarity. Next, we performed a network analysis to identify co-presenting organoids across the clusters and fast greedy optimization to gather the cluster nodes into the organoid communities. From these organoid cluster communities, we refined our organoid population by only retaining the organoids presented within its own community and meanwhile pruning the organoids shared by two or more communities, which elevated the core organoid representatives for each geometric design. From refinement results, we found that rectangle 1:4 organoids were significantly enriched in the organoid population during the organoid refinement process due to enhanced calcium handling capability. We employed two ensemble machine learning techniques: Random Forest (RF) and Extreme Gradient Boosting (XGB) to classify cardiac organoids based on different geometric shapes (Accuracy: 79% for RF and 76% for XGB). We found that both RF and XGB classifiers used Ca2+ rising time (t0) as the most important feature at ~20% of the model to make the prediction of geometric designs and organoid clusters, followed by contraction velocity and area ratio as second and third features.

The integration of hiPSC technology, organoid engineering, and AI opens the possibilities for uncovering the relationships between organoid geometric designs and their physiological functions that traditional analytical methods might overlook, paving the way for more controlled and optimized organoid design.

Acknowledgements (Optional):