Logic-based network model calibrated with machine-learning predicts the effect of drug therapies on cardiac hypertrophy induced by multiple different interventions

Introduction: : Primary mitral regurgitation (MR) is a heart disease that alters mechanical loading on the left ventricle, triggers neurohormonal activation and induces eccentric hypertrophy. Mitral valve repair (MVr) is the definitive treatment for MR, alleviating symptoms and offering chances of reverse cardiac remodeling, but at least 20% of patients do not fully recover. Furthermore, many patients undergoing MVr are receiving treatment with at least one heart failure (HF) drug that can independently influence remodeling. Individual animal studies of volume overload (VO) have provided snapshots of the complex underlying mechanobiology through non-invasive imaging, histopathologic assays, and gene microarray analyses. Separately, these studies only offer a partial picture of a complex process, so to produce a holistic interpretation of these data, we employ a biology-informed machine learning (ML) approach to combine the wealth of available data on experimental VO in dogs and rats using a logic-based network model of hypertrophic signaling in myocytes. When tested against data from independent studies not used in its calibration, the model reproduced the effects of agonists that stimulate key hypertrophic pathways, alone and in combination with HF drugs, as well as the effect of HF drugs in the context of VO. Using the known structure of the hypertrophic signaling network to integrate information from multiple studies into quantitative predictions of the range of expected drug responses appears to be a promising approach with potential applications in virtual clinical trials, drug testing, and personalized medicine.

Materials and

Methods: : We reviewed and collected data from 37 publications on experimental MR in dogs and 33 on experimental VO in rats. We collected data on serum concentration of norepinephrine (NE), angiotensin II (AngII), endothelin 1 (ET1), and natriuretic peptides; western blot-estimated myocyte activity of FAK, Akt, ERK5, ERK12, ELK1, cGMP, p38, and JNK; abundance of αMHC, βMHC, and SERCA proteins; and cell size. We also collected data on left ventricular end-diastolic volume and mass from dogs, which we used to estimate organ-scale strains. We used this data to calibrate the logic-based network model of cardiomyocyte signaling by Khalilimeybodi et al. (2020), consisting of 192 reactions (Fig 1a). We used the Markov Chain Monte Carlo (MCMC) algorithm with Gibbs sampling and Metropolis-Hasting selection criteria (an ML algorithm) to calibrate the relative influence (weight) on downstream signaling of network inputs NE, AngII, ET1, and stretch, by comparing the likelihood of model predictions to data on FAK, Akt, ERK5, ERK12, ELK1, cGMP, p38, and JNK activity as well as increase in cell size or mass. We then used the calibrated model to generate 1,000 Monte Carlo simulations replicating experimental volume overload or infusion of NE, AngII, or isoproterenol (ISO), alone or in combination with β Blockers (βB), angiotensin receptor blockers (ARB), or endothelin receptor antagonists (ERA). For validation, we counted how many experimental means fell within the model-predicted 50% (CI50) and 90% (CI90) confidence intervals.

Results, Conclusions, and Discussions:: Our model predicted that βB reduces ISO-induced hypertrophy in essentially all simulations, with a mean decrease of 27%, CI50 [-36%, -22%] and CI90 [-47%, -18%]. These predictions compared well to independent studies used for validation, all of which reported a significant reduction; 7 of 8 reported experimental means fell within our CI90, and 5 of 8 within the CI50 (Fig1b). The model predicted much smaller effects of ARB administration on ISO-induced growth and ERA on AngII-induced hypertrophy. In these cases, most simulations predicted a decrease, but some predicted an increase, with a mean predicted change of less than 10% and a CI90 encompassing 0 (Fig1b). Consistent with these predictions, validation studies reported a mix of significant and non-significant effects, with all reported means falling within the model-predicted CI90. Our model predicted that the effect of ERA on NE-infused growth is negligible. The only independent experimental study available for comparison reported a significant 23% decrease, which fell within the model-predicted CI90 (Fig1b).
The simulated effect of β-blockers and ARB treatment during VO was negligible on average. Experimental reports included some significant and some non-significant effects, and all means (7 studies total) fell within the model-predicted CI90 (Fig1c). By contrast, our model predicted that ERA protects against VO-induced hypertrophy in 90% of the cases, with a mean effect of -31%, CI50 [-51%,-13%] and CI90 [-77%,8%]. All four reported means from equivalent experiments fell within our predicted CI90 and half within our CI50 (Fig1c).
In total, 27 out of 28 (96%) means collected from independent studies of agonist infusions with or without pharmacologic blockers or drug administration during VO fell within model-predicted CI90s, confirming the ability of the calibrated model to anticipate the range of effects expected from interventions relevant to cardiac hypertrophy. The model suggests that changes in cell size and activation of the fetal gene program in the chronic stage of VO are most sensitive to Endothelin1 signaling. We propose that the approach presented here could be useful in future simulations of potential treatments for VO.

Acknowledgements and/or References (Optional):: This work is supported by the American Heart Association postdoctoral fellowship grant 23POST1026645 (https://doi.org/10.58275/AHA.23POST1026645.pc.gr.161204) to J.H.B; the National Institutes of Health and National Heart, Lung, and Blood Institute (NIH/NHLBI) grants R01HL159945 to J.W.H, R01HL162925 to J.J.S.