Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session B

Poster A5 - Data-Driven Approaches to Investigate Cardiotoxicity Using Human Induced Pluripotent Stem Cell-Based High-Throughput Drug Screening

Thursday, October 24, 2024
2:45 PM – 3:45 PM EST
Location: Exhibit Hall E, F & G

Introduction:  Recent advances in human induced pluripotent stem cell (hiPSC) technology have enabled the development of high-throughput models for cardiotoxicity drug screening. However, traditional analytical tools pose as a bottleneck for rapid implementation due to the time-intensive nature of their setup. Through the usage of modern data analytical techniques and machine learning (ML) algorithms, we can overcome these challenges and develop automated, unbiased models for analyzing these ever-expanding complex biological datasets.

In this study, we utilized calcium fluorescence signals from human induced pluripotent stem cells derived cardiomyocytes (hiPSC-CMs) collected through a high-throughput imaging system. We have developed a comprehensive analytical pipeline that incorporates machine learning, nonlinear analysis, and dimensionality reduction to quantitatively investigate the intricate biological signals of hiPSC-CMs. This innovative methodology effectively addresses the analytical bottleneck by streamlining data processing and enhancing the accuracy of our models. We applied our pipeline to investigate cardiotoxicity from drug exposure to multiple donor iPSC lines, demonstrating its efficacy in exploring and quantifying complex responses with donor variations.

Materials and

Methods:  All hiPSC-CMs were differentiated and cultured within 384-well plates utilizing standard procedures. Cells were treated with a calcium transient dye and imaged utilizing a FLIPR-Tetra Cellular Screening System. Measurements were taken at baseline and 90 minutes after exposure to 4 dosages (0.1,1,10,100μM) of 260 chemicals across 4 iPSC lines from consenting donors. This process resulted in a signal library with 9,476 signals reflecting varying degrees of drug exposure. Traditional calcium transient signal analysis was conducted utilizing ScreenWorks PeakPro, while nonlinear analysis was conducted through phase space reconstruction using available open-source algorithms, resulting in a total of 8 parameters for investigative analysis. Machine learning and statistical analysis were carried out in python using packages such as Numpy, Pandas, Matplotlib, Seaborn, Scikit-Learn, and Scanpy. Additionally, a sample group blind-labeled a subset of corresponding signals and nonlinear phase spaces to gauge the severity of irregularities present in signals. This enabled a deeper investigation into the feasibility of quantifying cardiotoxicity-induced irregularities in hiPSC-CM’s physiological function. The ComBat algorithm was applied to correct for donor-specific batch effects, and Uniform Manifold Approximation and Projection (UMAP), an unsupervised machine learning technique, was used for mining the trends within our data and identifying key factors in classification and irregularities. UMAP embeddings were also incorporated into our model as additional features.

Results, Conclusions, and Discussions: In our investigation of classifying hiPSC-CMs exposed to cardioactive drugs, we found that we could boost our top-performing algorithm, a neural network from 67% to 71% classification accuracy with the addition of nonlinear parameters (Figure 1a). UMAP visualization on our extracted parameter dataset revealed that data tends to cluster based on donors (Figure 1b). To address this, we decided to use the ComBat algorithm to correct the donor effects. Furthermore, we also used UMAP dimension reduction to reduce the raw calcium fluorescence signals with 800 data points into 2 dimensional embeddings (Figure 1c). We found that training the neural network model using both the corrected parameters and UMAP embedding boosted classification accuracy to 84%. This improvement likely results from the UMAP’s ability to preserve relationships between data points, capture nonlinearities, and reduce dimensions, leading to better generalization. Similar trends were observed using Support Vector Machine (SVM) and Random Forest (RF) algorithms, and the performance improvements can be further observed by the Receiver Operating Curve (ROC) and Area Under the Curve (AUC) calculations.

To conclude, our analytical model allowed us to quickly investigate a large semi-unstructured dataset with supervised and unsupervised ML algorithms to investigate driving trends to the functioning of hiPSC-CMs due to exposure to cardiotoxic drugs. This collaborative effort proved the feasibility of quickly evaluating drug cardiotoxicity effects using machine learning as an unbiased analytical tool. Nonlinear analysis and unsupervised machine learning also proved useful in improving our investigative analysis and visualizing the effects of cardiotoxicity.

Acknowledgements (Optional):