Intelligent Imaging
Like from the latin word intellegere = recognize, we would like to recognize from imaging how molecular biology in cells is unfolded. To interrogate genomic expression, single cell sequencing is currently the action of choice, while for 3D fluorescence, novel light sheet microscopy offers unprecedent imaging speed at lowest phototoxicity. We use these technologies in conjunction to explore the phenotype-genotype domain space and ultimately model different gene expression of patient organoids by imaging only. To correlate these big data matrices, deep learning classification became indispensable and finally drive our understanding in different aspects of therapy research and precision medicine at the Charité/ BIH.
Research topics
Automated light sheet microscopy
Advanced automated light sheet microscopy and single cell sequencing are primed in the intelligent imaging lab to compare morphologies with different expression. With dual top objective geometries we apply stage-scanning modes for fast 3D image screening or acquisition of organoids in hydrogel droplets while further subsequent confocal imaging adds resolution. We are determined to automated the whole process of 3D spotting of organoids to droplet respiration, followed by single cell acid nucleic library generations. Especially, massive imaging over time demands for intelligent solution, namely deep learning algorithms, to selectively save and identify relevant information from those volumetric heterogenous data.
Sample analytics by single cell genomics
We engineered single cell as well as nuclei RNA/ATAC sequencing libraries from different tissues derived from patients material directly or from their derived organoids. Here, we specialized on protocols for lung and pancreas biopsies adapted to high autolytic properties after tissue resection. Beside these wet lab challenges, the intelligent imaging group studies mophological-cellular features correlating with the different gene expression profiles. Ideally, assuming a sufficient collection size that a phenotypic picture of tissues or organoids holistically inform us about the genetic makeup. In essence, we are employing single cell sequencing to understand different disease entities (cancer) on cellular scale.
Deep tissue learning
We do utilize the way artificial deep neural networks learn to recognize and reconstruct patterns in input data. This approach to single cell genomics datasets allows the de novo identification of functional gene sets, master regulator genes and housekeeping genes from any kind of tissue origin. Precise partitioning into cell types or sub clones in cancer tissues can be improve by introducing class-specific filters of measured modalities (images or clinical diagnostic parameters). Ultimately, tissue or organoid images should contain all features learnt in deep neural networks to enable seamless disease prediction and therapy management.
Publications
*these authors contributed equally
§corresponding author