Deriving waveform parameters from calcium transients in human iPSC-derived cardiomyocytes to predict cardiac activity with machine learning

Abstract

Human induced pluripotent stem cell-derived cardiomyocytes have been established to detect dynamic calcium transients by fast kinetic fluorescence assays that provide insights into specific aspects of clinical cardiac activity. However, the precise derivation and use of waveform parameters to predict cardiac activity merit deeper investigation. In this study, we derived, evaluated, and applied 38 waveform parameters in a novel Python framework, including (among others) peak frequency, peak amplitude, peak widths, and a novel parameter, shoulder-tail ratio. We then trained a random forest model to predict cardiac activity based on the 25 parameters selected by correlation analysis. The area under the curve (AUC) obtained for leave-one-compound-out cross-validation was 0.86, thereby replicating the predictions of conventional methods and outperforming fingerprint-based methods by a large margin. This work demonstrates that machine learning is able to automate the assessment of cardiovascular liability from waveform data, reducing any risk of user-to-user variability and bias.

Keywords: calcium transients; cardiac activity; cardiomyocytes; cardiotoxicity; cardiovascular liability; hiPSC-CMs; machine learning; random forest; waveform parameters.

Graphical abstract

Figure 1

Frequency-, amplitude-, and shoulder-related parameters derived from each cycle in a waveform (A) The definitions of rise time, peak to end, decay time, tail duration, and peak space. PW10 is the peak width at 10% of prominence from top to bottom (analogous to other peak-width measures). (B) Definitions of amplitude, intensity, and valley. Amplitude equals the difference between intensity and valley. (C) An example showing the shoulder/tail. (D) The estimated density of amplitude of the waveform in (C). (E) Synthetic waveform data with the same peak space but different shoulder and tail durations. It can be seen that the parameter added here is able to distinguish numerically between the different shapes shown.

Figure 2

Comparison of the peak frequency and amplitude calculated in this study with those calculated by ScreenWorks The “r” is the Pearson correlation coefficient of the samples excluding those with multi-peaks (black circles) removing zero points. “Within 10%” is the ratio of points where parameters derived by our toolkit are within 10% of the parameters derived from ScreenWorks.

Figure 3

Parameter visualization and correlation to cardiac activity (A) An illustration of how the deviation of peak space and average intensity is able to distinguish between more regular and less regular waveforms. The four examples are for visualization purposes only. (B) The absolute point-biserial correlation coefficient |r_pb| between parameters of waveforms at different concentrations and the cardiac activities of the compounds.

Figure 4

Principal-component analysis of the parameters (A and B) Data points are sized by concentration. (A) shows the center of origin in (B) at increased resolution. The numbers on the axes of (B) are the percentage of variance explained by the first (x axis) and the second (y axis) principal components.

Figure 5

Receiver operating characteristic (ROC) curve of leave-one-compound-out random forest models The final model was built with 38 waveform parameters, while the simple model used only two: number of peaks and average peak amplitude. All three models were evaluated compound-wise.

Figure 6

Predicted class probabilities of cardiac activity of compounds (as established via a leave-one-compound-out validation) Multiple predictions for the same compound correspond to different replicates, where the gray points are short-term (30–60 min) samples and black points are long-term (72 h) samples. The colored points in the lines are the average probability of the compound being active. Orange means the compound is labeled by the FDA as cardiac active, while blue is cardiac inactive. Probability lower than 50% means the compound is predicted to be cardiac inactive by the machine learning model.