In clinical trials animal and cell line models are often used to evaluate the toxicity of a novel or existent chemical. However, relating the results of animal and in vitro model exposures to the human in vivo state presents a challenge, particularly given the high dimensionality of gene expression data.

In this study, we evaluate the potential of deep learning techniques to relate time series of human in vitro and rat in vivo gene expression to rat in vitro gene expression following exposure to a compound with the ultimate goal of predicting human in vivo gene expression following an exposure. Gene expression profiles from in vivo rat and in vitro primary rat and human hepatocytes following exposure to a subset of 47 compounds form the TG Gates data set was use.

Initially several simpler auto-encoder architectures along with a convolutional neural network are utilised to predict time series of human in vitro and rat in vivo gene expression given rat in vitro gene expression following an exposure to a novel compound. While the deep learning networks outperformed more traditional machine learning techniques used for benchmarking, such as random regression forest and k-nearest neighbours, due to the unavailability of human in vivo gene expression data upon which to train the networks it was not possible to make this final leap to human in vivo predictions. To this end Unsupervised Domain Adaptation (UDA) was employed.

Unsupervised Domain Adaptation permits the training of a predictor model for the human in vitro to in vivo problem, for which we have no data, given the similarly distributed data available for the rat in vitro to rat in vivo using transfer learning. The UDA model generated predictions of time series of human in vivo gene expression following exposure to a compound. The model’s rat in vivo predictions for a novel compound were more accurate than the both the auto-encoders and CNNs. Moreover, cursory exploration of the common latent space generated by UDA suggest a promising new method for the classification of a compound’s toxicity.

Adipose tissue is a dynamic organ that stores excess energy but also acts as an endocrine organ able to regulate energy homeostasis. While it mainly consists of adipocytes, other cell types are also present (e.g. immune cells).

The cell type composition of the tissue plays an important role in the function of adipose tissue as an endocrine organ and energy depot. Common experimental methods for studying cell heterogeneity including immunohistochemistry and flow cytometry are time consuming and expensive. Differences in cell type composition often lead to differences in gene expression in the tissue, thus ignoring cell type composition could influence the differences that one is interested in (e.g. between two experimental conditions). Therefore, computational methods have been developed to predict cell composition of tissue samples from gene expression data. The aim of this study is to investigate tissue gene expression corrected for differences in cell type composition.

Spontaneous Ca2+-release events (SCaEs) from the sarcoplasmic reticulum play crucial roles in the initiation of cardiac arrhythmias by promoting triggered activity. Recent publications have suggested that the subcellular distribution of RyR2 and L-type Ca2+ channels (LTCC) in cardiomyocytes modulates Ca2+ handling, but it is experimentally challenging to directly study the impact of Ca2+-handling protein distributions and Ca2+ handling in the same cell.

Here, we employ computational modeling to provide an in-depth analysis of the impact of variations in subcellular RyR2 and LTCC distributions on Ca2+-transient properties and SCaEs in a human atrial cardiomyocyte model. We incorporate experimentally observed RyR2 expression patterns and various configurations of axial tubules in a previously published model of the human atrial cardiomyocyte.

We identify an increased SCaE incidence for larger heterogeneity in RyR2 expression, in which SCaEs preferentially arise from regions of high local RyR2 expression. We show that incorporation of axial tubules in various amounts and locations reduces Ca2+-transient time to peak. Moreover, implementation of selective hyperphosphorylation of RyR2 around axial tubules increases the number of SCaEs.

In this talk I will compare septal and global wasted work rate in their sensitivity to changes in left ventricular afterload.

Wasted work ratio is a promising imaging (ultrasound/MRI) based indexes that, based on small retrospective clinical studies, can predicting response to cardiac resynchronization therapy. From previous computer modeling, animal and small clinical studies we know that afterload can change cardiac function under the same electrical substrate. The CircAdapt model has been used extensively in the context of to cardiac resynchronization therapy.

In this talk I will present new CircAdapt simulation results that demonstrate differences between the wasted work ratios and its relevance for clinical decision making.

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia, occurring in 1-2% of the population. Despite the development of sophisticated techniques for mapping of atrial electric activity, the electrophysiological mechanisms underlying AF maintenance are still incompletely understood.

Our work aims to describe the differences in the spatial dynamics of depolarization and repolarization waves in both sinus rhythm and AF. We detect local depolarization and repolarization time points in optical mapping recordings of perfused goat hearts with a history of AF. The qualitative assessment reveals a presence of different interaction patterns between the depolarization and repolarization waves during AF and sinus rhythm.