Recurrent neural networks for classification of human embryonic stem cell-derived cardiomyocytes

hESC-CMs;recurrent neural networks;semi-supervised learning;domain adaptation;Biomedical Engineering

Pacheco, Carolina A.Vidal, Rene ;

Johns Hopkins University

关键词: hESC-CMs;    recurrent neural networks;    semi-supervised learning;    domain adaptation;    Biomedical Engineering;   

Others  :  https://jscholarship.library.jhu.edu/bitstream/handle/1774.2/60153/PACHECO-THESIS-2018.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: JOHNS HOPKINS DSpace Repository

【 摘 要 】

Classification of human embryonic stem cell-derived cardiomyocytes (hESC-CMs) into phenotypes such as atrial-like or ventricular-like is important for applications in cardiac regenerative medicine and drug screening. However, a key challenge is the lack of ground truth labels for the phenotype of hESC-CMs: Whereas adult phenotypes are well-characterized in terms of the shape of their action potentials (APs), the understanding of how the shape of the AP of immature CMs relates to that of adult CMs remains limited. Recently, a new metamorphosis distance has been proposed to determine if a query immature AP is closer to a particular adult AP phenotype. However, the metamorphosis distance is difficult to compute making it unsuitable for classifying a large number of CMs.This thesis proposes two recurrent neural networks (RNNs) with long short-term memory (LSTM) units for classifying hESC-CMs. The first network is trained using a semi-supervised approach, in which the parameters of the network are learned by minimizing a loss function consisting of two terms: a supervised term that uses labeled data obtained from computational models of adult CMs, and an unsupervised term that uses a contrastive loss to encourage the labels of similar APs (as measured by the metamorphosis distance) to be the same. The second network is trained using a domain adaptation approach that captures the domain shift between immature and adult cells by adding a term to the loss function that penalizes their maximum mean discrepancy (MMD) in feature space.Experiments confirm the benefit of integrating information from both adult and stem cell-derived domains in the learning scheme and show that the proposed semi-supervised method generates results similar to the state of the art (94.73%) with clear computational advantages when applied to new samples. Experimental results on the domain adapted learning approach confirm that it not only is more computational efficient but also outperforms the state of the art in terms of clustering quality.In summary, the main contributions of this thesis are to formulate the classification of hESC-CM APs in the framework of artificial neural networks and to show that this new formulation improves with respect to the state of the art for this task in terms of both performance and computational efficiency.