Quantitative cross-species translators of cardiac myocyte electrophysiology: Model training, experimental validation, and applications

Abstract

Animal experimentation is key in the evaluation of cardiac efficacy and safety of novel therapeutic compounds. However, interspecies differences in the mechanisms regulating excitation-contraction coupling can limit the translation of experimental findings from animal models to human physiology and undermine the assessment of drugs’ efficacy and safety. Here, we built a suite of translators for quantitatively mapping electrophysiological responses in ventricular myocytes across species. We trained these statistical operators using a broad dataset obtained by simulating populations of our biophysically detailed computational models of action potential and Ca²⁺ transient in mouse, rabbit, and human. We then tested our translators against experimental data describing the response to stimuli, such as ion channel block, change in beating rate, and β-adrenergic challenge. We demonstrate that this approach is well suited to predicting the effects of perturbations across different species or experimental conditions and suggest its integration into mechanistic studies and drug development pipelines.

Fig. 1.. Interspecies differences in APD and APD sensitivity to changes in model parameters.

(A) Simulated AP and CaT traces elicited by stimulating the baseline mouse, rabbit, and human models at 1 Hz in control condition. (B) AP and CaT traces obtained in 50 representative variants sampled from mouse, rabbit, and human model populations (1-Hz pacing, control). (C) Schematic illustrating the approach adopted to perform parameter sensitivity analysis using the population-level data. The matrix X contains the randomly generated scaling factors (represented with circles) used to perturb the values of selected parameters in the baseline mouse, rabbit, and human models. In each population, AP and CaT features (squares) are estimated at steady state in each model variant, and their values are collected in the matrix F. Multivariable regression analysis between the matrix of scaling factors (X) and the matrix of features (F) is performed to assess the sensitivity of AP and CaT features to changes in model parameters in each species (34). The result of this process is the regression matrix B_SA, which coefficients (triangles) quantify the sensitivity of model features to parameter perturbations. (D) Regression coefficients illustrate the sensitivity of APD₉₀ and APD₅₀ to changes in the maximal conductance of repolarizing K⁺ channels in the three species (1-Hz pacing, control).

Fig. 2.. Overview of the cross-species translation workflow.

We created cross-species translators of electrophysiological response using the methodology proposed by Gong and Sobie (27). Our goal is to translate the drug-induced effects experimentally observed in myocytes from an animal model (mouse or rabbit) to predict the effects that these perturbations would cause in human. Specifically, given an experimental dataset consisting of AP and CaT features (APD₉₀, APD₅₀, and CaT_amp) assessed before (control) and after drug administration, we seek to predict the drug-induced effect on the same features in human myocytes. (A) To build the cross-species translator, we first collect the steady-state values of the biomarkers of interest (squares) in two populations of n_c models of control animal (mouse or rabbit) and human myocytes generated applying the same parameters perturbations (circles). Then, we generate B_cross by applying multivariable linear regression to the matrices of log-transformed animal (mouse or rabbit) and human features (F_animal and F_human). (B) Regression coefficients in B_cross (triangles) can be used to predict AP and CaT features in the output species (array “predicted f_human”), given the values observed in the input species (array “actual f_animal”). The process can be generalized to translate across conditions (e.g., changes in pacing rates) within the same species.

Fig. 3.. Development and validation of cross-species translators of electrophysiological response.

Mouse-to-human (A) and rabbit-to-human (B) translation matrices B_cross are built using mouse (or rabbit) and human features (APD₉₀, APD₅₀, CaT_t50, and CaT_tau) assessed in populations of models paced at 1 Hz (control condition). Scatter plots on the right in (A) and (B) show the result of validation performed with an independent set of simulated data, obtained in different populations counting 400 models. For each feature, we plot predicted human features (obtained by applying B_cross to the data produced simulating the input species, ordinate) against the actual values from human simulations (abscissa) and indicate the coefficient of determination R². (C) Bar graphs show the mean R² values obtained averaging the R² values estimated for each feature by predictors built using a varying number n_f of both input and output features (as illustrated by the schematic on the right). Circles correspond to the R² values for each feature, and error bars indicate SD of the mean.

Fig. 4.. Validation of cross-species prediction against data simulating the effects of selective ion channel block.

(A) Simulated AP and CaT traces elicited when pacing the mouse, rabbit, and human models at 1 Hz in control condition or upon selective block (50%) of several ion currents. Quantification of block-induced effects on AP and CaT is reported in fig. S3B. (B to F) Bar graphs show AP and CaT features obtained in mouse (M, green), rabbit (R, orange), and human (H, blue) simulations and predicted human data (white) obtained by applying the mouse-to-human (left) or rabbit-to-human (right) predictors built using a different number of features n_f (as described in Fig. 3C). Note that predictors built with two features can only predict APD₉₀ and APD₅₀ data (n_f = 2), and predictors built with four features can only predict APD₉₀, APD₅₀, CaT_tau, and CaT_t50 data (n_f = 4). Green and orange dashes in (B) to (F) represent the values predicted using as input the control mouse and rabbit data, respectively (see fig. S3A).

Fig. 5.. Experimental validation of cross-species prediction against experimental ion channel block data.

(A) Top: Experimental observations of the effect of the I_NaL blocker GS-967 on AP waveform in mouse (0.3 μM) (35), rabbit (1 μM) (36), and human (0.5 μM) (37) ventricular myocytes. The bar graphs on the bottom display the validation of mouse-to-human and rabbit-to-human translations: Mouse, rabbit, and human data are reported in green, orange, and blue, respectively, while predictions of human response are reported in white. Experimentally measured effects of the (B) I_Kr blocker E-4031 (1 μM) on AP waveform in rabbit (38) and human ventricular myocytes [experimental data from (39), estimated as in (40)] and (C) I_Kr blocker Sotalol on AP waveform in rabbit (52 μM) and human (30 μM) ventricular myocytes (41) at left are compared with the results of rabbit-to-human translation at right: Rabbit and human data are reported in orange and blue, respectively, while predictions of human response are reported in white. All predictions shown here were obtained using the mouse-to-human or rabbit-to-human translators built using only APD₉₀ and APD₅₀ data (n_f = 2 in Fig. 3C). Green and orange dashes near white bars represent the values predicted using as input the control mouse and rabbit data, respectively (see fig. S3A). Translators’ inputs and validation values are obtained using experimental data to scale the simulated control features as described in Methods and fig. S5. Error bars indicate SD of the mean.

Fig. 6.. Assessment of input features’ informative score.

Average R² values characterizing the overall performance of mouse-to-human (A) and rabbit-to-human translators (B) built using a fixed number of output features (APD₉₀, APD₅₀, CaT_amp, and CaT_tau) and a variable number of input features n_f. Starting from a set of 10 input features, at each iteration, an automatic recursive features elimination routine excluded the least informative input feature. Elimination order is reported on the right, where the features highlighted in blue are those produced in output for human. Error bars indicate SD of the mean.

Fig. 7.. Experimental validation of cross-frequency prediction against data describing effect of ion channel block in rabbit myocytes.

Left: Changes in APD₉₀ and APD₅₀ induced by block of I_NaL [1 μM GS-967 (A)], I_to [5 mM 4-AP (B)], I_Kr [1 μM E-4031 (C)], I_Ks [1 μM HMR-1556 (D)], and I_K1 [200 nM PA-6 (E)] in rabbit myocytes paced at 0.5, 1, 2, and 3 Hz (38). Individual data points are reported in fig. S9. Right: Validation of cross-frequency translation of APD values after drug administration: 1-Hz data are reported in gray, while 0.5-, 2-, and 3-Hz data are in orange, and their predictions from 1-Hz data are in white. The predictors used here were built with APD₉₀ and APD₅₀ data (n_f = 2) obtained stimulating the population of rabbit models in control condition at the different pacing rates. Translators’ inputs and validation values are obtained using experimental data to scale the simulated control features as described in Methods and fig. S5. Error bars indicate SD of the mean.

Fig. 8.. Experimental validation of prediction of ISO administration effects across pacing frequencies or species.

(A) Left: Changes in APD₉₀ and APD₅₀ induced by ISO administration (30 nM) in rabbit myocytes paced at different frequencies (38). Individual data points are reported in fig. S9. Right: Validation of cross-frequency translation of ISO effect: 1-Hz data are reported in gray, while 0.5-, 2-, and 3-Hz data are in orange, and their predictions from 1-Hz data are in white. The predictors used here were those described in Fig. 7. (B) Experimental observation of the effect of ISO administration in mouse (100 nM) (43) and rabbit (30 nM) (38) ventricular myocytes. a.u., arbitrary units. (C) Mouse and rabbit data are reported in green and orange, respectively, while white bars represent prediction of rabbit response obtained with two mouse-to-rabbit translators built using different number of input features. The translator built with two features uses mouse APD₉₀ and APD₅₀ data only (n_f = 2), while the translator built with four features also uses mouse CaT_amp and CaT_tau data (n_f = 4). (D) Actual and predicted ISO-induced APD changes estimated normalizing the values in (C) to the corresponding APD value assessed in the absence of ISO (i.e., control). Translators’ inputs and validation values are obtained using experimental data to scale the simulated control features as described in Methods and fig. S5. Error bars indicate SD of the mean.

Fig. 9.. Experimental validation of cross-species prediction of sympathetic stimulation effect in quasi-physiological conditions.

(A) Experimental observation of the effect on heart rate (HR) and duration (D) of AP and CaT induced by SNS in mouse and rabbit innervated whole-heart preparations (21). (B) Experimental validation of mouse-to-rabbit translation of SNS-induced relative APD and CaTD changes. Experimental mouse and rabbit data are reported in green and orange, respectively. Predicted rabbit response (in white) is obtained by applying a predictor built using relative changes in APD and CaTD estimated from simulations in mouse and rabbit populations that mimic the conditions observed in experiments (see fig. S10C). (C) The relative SNS-induced effect in rabbit predicted from mouse data is used to estimate AP and CaTD during SNS from control rabbit data. Error bars indicate SD of the mean. Open circles in (A) indicate results of individual experiments.