PhD Position

Your position

Data-driven analysis with deep learning offers an enormous potential in today's life-science sector. Trained end-to-end, deep learning models exceed human-level performance. However, controlling for confounding factors is not straight forward. Is it a bias in an acquisition batch or other hidden correlations of the input data and the target, confounding variables need to be controlled to make meaningful predictions. The project will cover a multitude of data modalities from large biological and chemical data sets of high-content screenings, whole-slide images, and genomics.
Acquisition Batches in Biological Activity Prediction

Capturing the morphology of a cell from microscopy images offers great potential for predicting its biological activity. Exploiting predictive features, so called image-based compound fingerprints, may increases hit rates by orders of magnitude. Despite their potential, in practice, the biological activity severally correlates with acquisition batches which need to be considered as confounds to make accurate predictions. You are going to be working with public data sets such as the Compound Profiling Data Set BBBC021 and the 1008 Tales Data Set , and the “Phenoprint” data set acquired by our industrial partner which is based on https://doi.org/10.1016/j.chembiol.2018.01.015 .

Histological Whole-Slide Images

The automatic analysis of histological whole-slide images is ever more relevant in digital pathology. In a wide variety of challenges ranging from tissue detection and characterization and understanding of cell interactions, confounds ought to be identified and their influence alleviated to reach high-quality estimates. You will handle and analyze giga-pixel images from several studies which are conducted in-house and heterogeneous sets of oncological data of human trials acquired on multiple sites.