Application of transfer learning to predict drug-induced human in vivo gene expression changes using rat in vitro and in vivo data.

The liver is the primary site for the metabolism and detoxification of many compounds, including pharmaceuticals. Consequently, it is also the primary location for many adverse reactions. As the liver is not readily accessible for sampling in humans; rodent or cell line models are often used to evaluate potential toxic effects of a novel compound or candidate drug. However, relating the results of animal and in vitro studies to relevant clinical outcomes for the human in vivo situation still proves challenging. In this study, we incorporate principles of transfer learning within a deep artificial neural network allowing us to leverage the relative abundance of rat in vitro and in vivo exposure data from the Open TG-GATEs data set to train a model to predict the expected pattern of human in vivo gene expression following an exposure given measured human in vitro gene expression. We show that domain adaptation has been successfully achieved, with the rat and human in vitro data no longer being separable in the common latent space generated by the network. The network produces physiologically plausible predictions of human in vivo gene expression pattern following an exposure to a previously unseen compound. Moreover, we show the integration of the human in vitro data in the training of the domain adaptation network significantly improves the temporal accuracy of the predicted rat in vivo gene expression pattern following an exposure to a previously unseen compound. In this way, we demonstrate the improvements in prediction accuracy that can be achieved by combining data from distinct domains.

Influence of the Tikhonov Regularization Parameter on the Accuracy of the Inverse Problem in Electrocardiography.

The electrocardiogram (ECG) is the standard method in clinical practice to non-invasively analyze the electrical activity of the heart, from electrodes placed on the body's surface. The ECG can provide a cardiologist with relevant information to assess the condition of the heart and the possible presence of cardiac pathology. Nonetheless, the global view of the heart's electrical activity given by the ECG cannot provide fully detailed and localized information about abnormal electrical propagation patterns and corresponding substrates on the surface of the heart. Electrocardiographic imaging, also known as the inverse problem in electrocardiography, tries to overcome these limitations by non-invasively reconstructing the heart surface potentials, starting from the corresponding body surface potentials, and the geometry of the torso and the heart. This problem is ill-posed, and regularization techniques are needed to achieve a stable and accurate solution. The standard approach is to use zero-order Tikhonov regularization and the L-curve approach to choose the optimal value for the regularization parameter. However, different methods have been proposed for computing the optimal value of the regularization parameter. Moreover, regardless of the estimation method used, this may still lead to over-regularization or under-regularization. In order to gain a better understanding of the effects of the choice of regularization parameter value, in this study, we first focused on the regularization parameter itself, and investigated its influence on the accuracy of the reconstruction of heart surface potentials, by assessing the reconstruction accuracy with high-precision simultaneous heart and torso recordings from four dogs. For this, we analyzed a sufficiently large range of parameter values. Secondly, we evaluated the performance of five different methods for the estimation of the regularization parameter, also in view of the results of the first analysis. Thirdly, we investigated the effect of using a fixed value of the regularization parameter across all reconstructed beats. Accuracy was measured in terms of the quality of reconstruction of the heart surface potentials and estimation of the activation and recovery times, when compared with ground truth recordings from the experimental dog data. Results show that values of the regularization parameter in the range (0.01-0.03) provide the best accuracy, and that the three best-performing estimation methods (L-Curve, Zero-Crossing, and CRESO) give values in this range. Moreover, a fixed value of the regularization parameter could achieve very similar performance to the beat-specific parameter values calculated by the different estimation methods. These findings are relevant as they suggest that regularization parameter estimation methods may provide the accurate reconstruction of heart surface potentials only for specific ranges of regularization parameter values, and that using a fixed value of the regularization parameter may represent a valid alternative, especially when computational efficiency or consistency across time is required.

Use of deep learning methods to translate drug-induced gene expression changes from rat to human primary hepatocytes.

In clinical trials, animal and cell line models are often used to evaluate the potential toxic effects of a novel compound or candidate drug before progressing to human trials. However, relating the results of animal and in vitro model exposures to relevant clinical outcomes in the human in vivo system still proves challenging, relying on often putative orthologs. In recent years, multiple studies have demonstrated that the repeated dose rodent bioassay, the current gold standard in the field, lacks sufficient sensitivity and specificity in predicting toxic effects of pharmaceuticals in humans. In this study, we evaluate the potential of deep learning techniques to translate the pattern of gene expression measured following an exposure in rodents to humans, circumventing the current reliance on orthologs, and also from in vitro to in vivo experimental designs. Of the applied deep learning architectures applied in this study the convolutional neural network (CNN) and a deep artificial neural network with bottleneck architecture significantly outperform classical machine learning techniques in predicting the time series of gene expression in primary human hepatocytes given a measured time series of gene expression from primary rat hepatocytes following exposure in vitro to a previously unseen compound across multiple toxicologically relevant gene sets. With a reduction in average mean absolute error across 76 genes that have been shown to be predictive for identifying carcinogenicity from 0.0172 for a random regression forest to 0.0166 for the CNN model (p < 0.05). These deep learning architecture also perform well when applied to predict time series of in vivo gene expression given measured time series of in vitro gene expression for rats.

Adipose tissue in health and disease through the lens of its building blocks.

Understanding adipose tissue cellular heterogeneity and homeostasis is essential to comprehend the cell type dynamics in metabolic diseases. Cellular subpopulations in the adipose tissue have been related to disease development, but efforts towards characterizing the adipose tissue cell type composition are limited. Here, we identify the cell type composition of the adipose tissue by using gene expression deconvolution of large amounts of publicly available transcriptomics level data. The proposed approach allows to present a comprehensive study of adipose tissue cell type composition, determining the relative amounts of 21 different cell types in 1282 adipose tissue samples detailing differences across four adipose tissue depots, between genders, across ranges of BMI and in different stages of type-2 diabetes. We compare our results to previous marker-based studies by conducting a literature review of adipose tissue cell type composition and propose candidate cellular markers to distinguish different cell types within the adipose tissue. This analysis reveals gender-specific differences in CD4+ and CD8+ T cell subsets; identifies adipose tissue as rich source of multipotent stem/stromal cells; and highlights a strongly increased immune cell content in epicardial and pericardial adipose tissue compared to subcutaneous and omental depots. Overall, this systematic analysis provides comprehensive insights into adipose tissue cell-type heterogeneity in health and disease.

A computational model of postprandial adipose tissue lipid metabolism derived using human arteriovenous stable isotope tracer data.

Given the association of disturbances in non-esterified fatty acid (NEFA) metabolism with the development of Type 2 Diabetes and Non-Alcoholic Fatty Liver Disease, computational models of glucose-insulin dynamics have been extended to account for the interplay with NEFA. In this study, we use arteriovenous measurement across the subcutaneous adipose tissue during a mixed meal challenge test to evaluate the performance and underlying assumptions of three existing models of adipose tissue metabolism and construct a new, refined model of adipose tissue metabolism. Our model introduces new terms, explicitly accounting for the conversion of glucose to glyceraldehye-3-phosphate, the postprandial influx of glycerol into the adipose tissue, and several physiologically relevant delays in insulin signalling in order to better describe the measured adipose tissues fluxes. We then applied our refined model to human adipose tissue flux data collected before and after a diet intervention as part of the Yoyo study, to quantify the effects of caloric restriction on postprandial adipose tissue metabolism. Significant increases were observed in the model parameters describing the rate of uptake and release of both glycerol and NEFA. Additionally, decreases in the model's delay in insulin signalling parameters indicates there is an improvement in adipose tissue insulin sensitivity following caloric restriction.

The application of omics-based human liver platforms for investigating the mechanism of drug-induced hepatotoxicity in vitro.

Drug-induced liver injury (DILI) complicates safety assessment for new drugs and poses major threats to both patient health and drug development in the pharmaceutical industry. A number of human liver cell-based in vitro models combined with toxicogenomics methods have been developed as an alternative to animal testing for studying human DILI mechanisms. In this review, we discuss the in vitro human liver systems and their applications in omics-based drug-induced hepatotoxicity studies. We furthermore present bioinformatic approaches that are useful for analyzing toxicogenomic data generated from these models and discuss their current and potential contributions to the understanding of mechanisms of DILI. Human pluripotent stem cells, carrying donor-specific genetic information, hold great potential for advancing the study of individual-specific toxicological responses. When co-cultured with other liver-derived non-parenchymal cells in a microfluidic device, the resulting dynamic platform enables us to study immune-mediated drug hypersensitivity and accelerates personalized drug toxicology studies. A flexible microfluidic platform would also support the assembly of a more advanced organs-on-a-chip device, further bridging gap between in vitro and in vivo conditions. The standard transcriptomic analysis of these cell systems can be complemented with causality-inferring approaches to improve the understanding of DILI mechanisms. These approaches involve statistical techniques capable of elucidating regulatory interactions in parts of these mechanisms. The use of more elaborated human liver models, in harmony with causality-inferring bioinformatic approaches will pave the way for establishing a powerful methodology to systematically assess DILI mechanisms across a wide range of conditions.

Improved quantification of muscle insulin sensitivity using oral glucose tolerance test data: the MISI Calculator.

The Muscle Insulin Sensitivity Index (MISI) has been developed to estimate muscle-specific insulin sensitivity based on oral glucose tolerance test (OGTT) data. To date, the score has been implemented with considerable variation in literature and initial positive evaluations were not reproduced in subsequent studies. In this study, we investigate the computation of MISI on oral OGTT data with differing sampling schedules and aim to standardise and improve its calculation. Seven time point OGTT data for 2631 individuals from the Maastricht Study and seven time point OGTT data combined with a hyperinsulinemic-euglycaemic clamp for 71 individuals from the PRESERVE Study were used to evaluate the performance of MISI. MISI was computed on subsets of OGTT data representing four and five time point sampling schedules to determine minimal requirements for accurate computation of the score. A modified MISI computed on cubic splines of the measured data, resulting in improved identification of glucose peak and nadir, was compared with the original method yielding an increased correlation (ρ = 0.576) with the clamp measurement of peripheral insulin sensitivity as compared to the original method (ρ = 0.513). Finally, a standalone MISI calculator was developed allowing for a standardised method of calculation using both the original and improved methods.

In Vivo Validation of Electrocardiographic Imaging.

OBJECTIVES: The purpose of this study was to evaluate the accuracy of noninvasive reconstructions of epicardial potentials, electrograms, activation and recovery isochrones, and beat origins by simultaneously performing electrocardiographic imaging (ECGI) and invasive epicardial electrography in intact animals. BACKGROUND: Noninvasive imaging of electrical potentials at the epicardium, known as ECGI, is increasingly applied in patients to assess normal and abnormal cardiac electrical activity. METHODS: Body-surface potentials and epicardial potentials were recorded in normal anesthetized dogs. Computed tomography scanning provided a torso-heart geometry that was used to reconstruct epicardial potentials from body-surface potentials. RESULTS: Electrogram reconstructions attained a moderate accuracy compared with epicardial recordings (median correlation coefficient: 0.71), but with considerable variation (interquartile range: 0.36 to 0.86). This variation could be explained by a spatial mismatch (overall resolution was <20 mm) that was most apparent in regions with electrographic transition. More accurate derivation of activation times (Pearson R: 0.82), recovery times (R: 0.73), and the origin of paced beats (median error: 10 mm; interquartile range: 7 to 17 mm) was achieved by a spatiotemporal approach that incorporates the characteristics of the respective electrogram and neighboring electrograms. Reconstruction of beats from repeated single-site pacing showed a stable localization of origin. Cardiac motion, currently ignored in ECGI, correlates negatively with reconstruction accuracy. CONCLUSIONS: ECGI shows a decent median accuracy, but variability in electrogram reconstruction can be sizable. At present, therefore, clinical interpretations of ECGI should not be made on the basis of single electrograms only. Incorporating local spatiotemporal characteristics allows for accurate reconstruction of epicardial activation and recovery patterns, and beat origin localization to a 10-mm precision. Even more reliable interpretations are expected when the influences of cardiac motion are accounted for in ECGI.

Physiology-based regularization of the electrocardiographic inverse problem.

The inverse problem of electrocardiography aims at noninvasively reconstructing electrical activity of the heart from recorded body-surface electrocardiograms. A crucial step is regularization, which deals with ill-posedness of the problem by imposing constraints on the possible solutions. We developed a regularization method that includes electrophysiological input. Body-surface potentials are recorded and a computed tomography scan is performed to obtain the torso-heart geometry. Propagating waveforms originating from several positions at the heart are simulated and used to generate a set of basis vectors representing spatial distributions of potentials on the heart surface. The real heart-surface potentials are then reconstructed from the recorded body-surface potentials by finding a sparse representation in terms of this basis. This method, which we named 'physiology-based regularization' (PBR), was compared to traditional Tikhonov regularization and validated using in vivo recordings in dogs. PBR recovered details of heart-surface electrograms that were lost with traditional regularization, attained higher correlation coefficients and led to improved estimation of recovery times. The best results were obtained by including approximate knowledge about the beat origin in the PBR basis.

Wavelet-sparsity based regularization over time in the inverse problem of electrocardiography.

Noninvasive, detailed assessment of electrical cardiac activity at the level of the heart surface has the potential to revolutionize diagnostics and therapy of cardiac pathologies. Due to the requirement of noninvasiveness, body-surface potentials are measured and have to be projected back to the heart surface, yielding an ill-posed inverse problem. Ill-posedness ensures that there are non-unique solutions to this problem, resulting in a problem of choice. In the current paper, it is proposed to restrict this choice by requiring that the time series of reconstructed heart-surface potentials is sparse in the wavelet domain. A local search technique is introduced that pursues a sparse solution, using an orthogonal wavelet transform. Epicardial potentials reconstructed from this method are compared to those from existing methods, and validated with actual intracardiac recordings. The new technique improves the reconstructions in terms of smoothness and recovers physiologically meaningful details. Additionally, reconstruction of activation timing seems to be improved when pursuing sparsity of the reconstructed signals in the wavelet domain.

Determinants of beat-to-beat variability of repolarization duration in the canine ventricular myocyte: a computational analysis.

Beat-to-beat variability of repolarization duration (BVR) is an intrinsic characteristic of cardiac function and a better marker of proarrhythmia than repolarization prolongation alone. The ionic mechanisms underlying baseline BVR in physiological conditions, its rate dependence, and the factors contributing to increased BVR in pathologies remain incompletely understood. Here, we employed computer modeling to provide novel insights into the subcellular mechanisms of BVR under physiological conditions and during simulated drug-induced repolarization prolongation, mimicking long-QT syndromes type 1, 2, and 3. We developed stochastic implementations of 13 major ionic currents and fluxes in a model of canine ventricular-myocyte electrophysiology. Combined stochastic gating of these components resulted in short- and long-term variability, consistent with experimental data from isolated canine ventricular myocytes. The model indicated that the magnitude of stochastic fluctuations is rate dependent due to the rate dependence of action-potential (AP) duration (APD). This process (the "active" component) and the intrinsic nonlinear relationship between membrane current and APD ("intrinsic component") contribute to the rate dependence of BVR. We identified a major role in physiological BVR for stochastic gating of the persistent Na(+) current (INa) and rapidly activating delayed-rectifier K(+) current (IKr). Inhibition of IKr or augmentation of INa significantly increased BVR, whereas subsequent ß-adrenergic receptor stimulation reduced it, similar to experimental findings in isolated myocytes. In contrast, ß-adrenergic stimulation increased BVR in simulated long-QT syndrome type 1. In addition to stochastic channel gating, AP morphology, APD, and beat-to-beat variations in Ca(2+) were found to modulate single-cell BVR. Cell-to-cell coupling decreased BVR and this was more pronounced when a model cell with increased BVR was coupled to a model cell with normal BVR. In conclusion, our results provide new insights into the ionic mechanisms underlying BVR and suggest that BVR reflects multiple potentially proarrhythmic parameters, including increased ion-channel stochasticity, prolonged APD, and abnormal Ca(2+) handling.

Realistic training data improve noninvasive reconstruction of heart-surface potentials.

The inverse problem of electrocardiography is to noninvasively reconstruct electrical heart activity from body-surface electrocardiograms. Solving this problem is beneficial to clinical practice. However, reconstructions cannot be obtained straightforwardly due to the ill-posed nature of this problem. Therefore, regularization schemes are necessary to arrive at realistic solutions. To date, no electrophysiological data have been used in reconstruction methods and regularization schemes. In this study, we used a training set of simulated heart-surface potentials to create a realistic basis for reconstructions of electrical cardiac activity. We tested this method in computer simulations and in one patient. The quality of reconstruction improved significantly after projection of the results of traditional regularization methods on this new basis, both in silico (p<0.01) and in vivo (p<0.05). Thus, we demonstrate that the novel concept of applying electrophysiological data might be useful to improve noninvasive reconstruction of electrical heart activity.