He and his team have now developed an algorithm that can predict the decision in advance. So-called Deep Learning is the key. "Deep Neural Networks play a major role in our method", stated Dr. Marr. "Our algorithm classifies light microscopic images and videos of individual cells by comparing these data with past experience from the development of such cells. In this way, the algorithm 'learns' how certain cells behave."
Specifically, the researchers examined hematopoietic stem cells that were filmed under the microscope in the lab of Timm Schroeder at ETH Zurich. Using the information on appearance and speed, the software was able to 'memorize' the corresponding behaviour patterns and then make its prediction. "Compared to conventional methods, such as fluorescent antibodies against certain surface proteins, we know how the cells will decide three cell generations earlier", reported ICB scientist Dr. Felix Buggenthin, joint first author of the study together with Dr. Florian Büttner.
But what is the benefit of this look into the future? As study leader Dr. Marr explained: "Since we now know which cells will develop in which way, we can isolate them earlier than before and examine how they differ at a molecular level. We want to use this information to understand how the choices are made for particular developmental traits."
In the future, the focus will expand beyond hematopoietic stem cells. "We are using Deep Learning for very different problems with sufficiently large data records", explained Prof. Dr. Dr. Fabian Theis, ICB director and holder of the Mathematical Modelling of Biological Systems Chair at the Technical University of Munich (TUM), who led the study together with Carsten Marr. "For example, we use very similar algorithms to analyse disease-associated patterns in the genome and identify biomarkers in clinical cell screens."
The study is the latest result of a close cooperation between the ICB scientists and Prof. Dr. Timm Schroeder from the Department of Biosystems Science and Engineering at ETH Zurich in Basel, who previously worked at the Helmholtz Zentrum in Munich. In July 2016, the scientists jointly introduced a software inNature Biotechnolgythat allows to observe individual cells over many days and simultaneously measure their molecular properties. They also published a study inNaturethat already dealt with the development of hematopoietic stem cells. Using time-lapse microscopy, the researchers were able to observe the maturation of living hematopoietic stem cells with high precision while also quantifying certain proteins.
Deep Learning algorithms simulate the learning processes in people using artificial neural networks. The principle functions particularly well when large quantities of data (Big Data) are available for training. Image recognition is one of Deep Learning's strengths. More decision layers are placed between the input - here, the cell image data - and the output - here, the prediction of the cell development - than usually found in neuronal networks, which is why the term "deep" is used.
The study titled " Prospective identification of hematopoietic lineage choice by deep learning " is authored by Buggenthin F. et al. and appeared inNature Methods.