User: Guest  Login
Original title:
Modeling single-cell perturbations using deep learning 
Translated title:
Modellierung von Störungen in einzelnen Zellen durch tiefes Lernen 
Document type:
TUM School of Life Sciences 
Theis, Fabian (Prof. Dr. Dr.) 
Theis, Fabian (Prof. Dr. Dr.); Peér, Dana (Prof. Dr.); Colomé-Tatché, Maria (Prof. Dr.) 
Subject group:
BIO Biowissenschaften 
deep learning, single-cell, perturbation 
Translated keywords:
einzelnen Zellen, deep learning 
TUM classification:
BIO 110 
Single-cell genomics has revolutionized the understanding of heterogeneity in both health and disease and empowered profiling millions of cells across different tissues to construct reference atlases. The ultimate goal of a single-cell reference atlas is to facilitate understanding cellular perturbations by comparing them to a healthy reference. Perturbation is defined as any intervention changing the cell state from normal to perturbed state. The intervention can be caused by disease or a treatment such as drugs. In this cumulative thesis, the goal was to develop deep learning algorithms to analyze single-cell perturbation studies. To pursue this, we first need to map the newly acquired datasets (i.e., query) like perturbation studies into healthy reference atlases built by consortia such as Human Cell Atlas (HCA). However, the usability of public reference atlases to analyze the query data is hindered by technical variations between the query and reference atlas, computational complexities, resource limitations, and raw data sharing policies. To address these issues, I developed a deep learning algorithm called single-cell architecture surgery (scArches). scArches allows fast, efficient, and accurate integration of perturbation datasets into the reference atlas while preserving the perturbation heterogeneity enabling the discovery of novel cell states. The integration of the perturbation dataset into the reference atlas transforms it into a perturbation atlas. Yet, the space of potential outcomes is vast and experimentally infeasible to measure all possible perturbations such as drugs or gene knockouts. Therefore, computational tools are required to predict the response to the stimuli for unseen phenomena not observed in the initial atlas for an efficient experimental design and novel biological discovery. This motivates the second aim of this thesis, which is to design models to predict the transcriptomic responses to a perturbation at the single-cell level. To accomplish this goal, I developed deep learning algorithms to learn and predict perturbation responses. These methods demonstrated the ability to predict the response to drugs, genetic knock-outs, and diseases. I envision the strategies presented in this thesis would facilitate efficient experimental design and thus hypothesis generation using single-cell genomics. 
Translated abstract:
Die Einzelzellgenomik hat das Verständnis der Heterogenität in Gesundheit undKrankheit revolutioniert und die Erstellung von Profilen von Millionen von Zel-len in verschiedenen Geweben ermöglicht, um Referenzatlanten zu erstellen. Dasultimative Ziel eines Einzelzell-Referenzatlasses ist es, das Verständnis zellulä-rer Störungen zu erleichtern, indem sie mit einer gesunden Referenz verglichenwerden. Als Störung wird jeder Stimulus definiert, der den Zellzustand von ei-nem normalen zu einem gestör...    »
Oral examination:
File size:
45935548 bytes 
Last change: