Spatial transcriptomics and imaging

LODE- Longitudinal deep learning for clinical diagnosis, feature representation and interventional modeling

Longitudinal data capture changes in patient phenotypes (over time and under clinical intervention) and such data play a key role in clinical medicine. Our specific use-case is in ophthalmology, with very large-scale and clinically-annotated 2D and 3D eye imaging data combined with clinical interventions from the University Eye Hospital Munich. In the LODE project, we aim to (1) to develop new DL methods for longitudinal image-based clinical data, with an emphasis on interpretable methodology applicable to large-scale, routine clinical observations, and (2) to integrate causal and network modeling concepts with DL in this application. In recent work, we propose a self-supervised learning framework to impute modalities by cross-modal prediction. Specifically, we consider learning to predict retinal thickness on infrared fundus images to pre-train Deep learning models for classify diabetic retinopathy classes on fundus images.



Analysis of spatial transcriptomics data and integration with scRNA-sequencing

Spatial molecular profiling techniques allow us to assay gene expression at protein or RNA level in situ. These experimental techniques provide a toolkit to investigate tissue biology at an unprecedented resolution. Computational and analytical tools are needed for the analysis of such data. To this end we are developing analysis tools and data infrastructure for spatial molecular data, as well as modeling approaches to disentangle spatial components of cellular and tissue variation.


Scanpy Tutorials