To complement the manual, expert-driven analysis of flow cytometry data performed in our labs, we run automated gating algorithms to identify cell populations in a hypothesis-free way. This approach simplifies the (automated) analysis of hundreds or thousands of samples while at the same time providing novel insights into rare cell-populations that would have otherwise gone unnoticed.

We perform differential expression analysis, gene set testing, and co-expression network analysis on RNAseq and Proteomics data to generate the most insights from valuable data. Our routine analysis combines standard pipelines and exploratory components that involve frequent feedback cycles between the data science team and the researcher who knows their data best.

Our pipelines are designed to operate both on targeted and untargeted metabolomics data. This includes mapping of raw machine readouts to metabolome databases, feature selection, and downstream statistical analysis to obtain interpretable, actionable results.

Our labs produce histologic images using H&E, chromogenic and fluorescence multiplex immunohistochemistry images. That’s why our digital pathology pipelines can process all of them by respecting each methods unique strengths and weaknesses. For all images, we perform tissue segmentation, cell segmentation, and phenotyping of identified cells. This enables us to perform basic cell quantification for highly interpretable features, and the extraction spatial features, like nearest neighbor analysis or the investigation of touching-cells. If needed, custom object detection and feature extraction pipelines can be setup according to researcher specifications.

To tap into the vast potential of manually written, unstructured clinical health records we employ natural language processing (NLP) pipelines to identify relevant clinical concepts in the text and map them to well-defined clinical ontologies (ICD-10, SNOMED CT). This enables downstream search and automated processing, including the correlation of features with lab parameters in an automated and interpretable way.

When predictive models are desired, we combine multiple data sets to get a holistic view of the data. To avoid that black-box machine learning models make good predictions that cannot be interpreted by humans, we employ both feature extraction and feature selection on each data set. Small, interpretable models combined with statistical analysis of the most relevant features enable us to combine the power of machine learning with the necessity of understanding every detail of our data.