For ELLIS Munich, we align expertise and substantial new funding to setup a thriving hub for biomedical AI.

Research in biomedicine is generating data on large scales, with more than 100M genomes expected in 2030, electronic health records finally maturing in many countries and standardized imaging of organs or smaller tissues being acquired, digitized and stored in often longitudinal fashion. Preliminary analysis of this data, disease diagnosis and subsequent treatment are often left to human experts who struggle with the volume and complexity of the data. ML can leverage these data sets to diagnose and prevent disease, identify mechanisms and drug targets, and ultimately support clinical decisions.

Research topics

Human genomics

We aim to model molecular variation onto populations of cells and individuals. Predicting and understanding variation across populations and under perturbation, e.g., diseases or drug treatment, will be addressed with both supervised and unsupervised methods.

Biomedical imaging

Our goal is to provide AI solutions for image-based diagnostics in cooperation with the computer vision topic, leveraging existing imaging data and data from novel technologies. We will integratively apply statistical and mechanistic models to, e.g., reconstruct raw imaging data, and to parametrize disease models with single-cell measurements for predicting patient-specific dynamics. This will enable early diagnosis, standardized and rapid disease classification, and improved individualized prediction.

Drug research

We aim for a learning system that integrates available experimental data towards a virtual screening system, proposing chemical synthesis and experimental assays for automated design of small molecules for drug discovery projects. We will develop immunotherapy pipelines by integrating uncertainty quantification for risk-averse designs and contribute to DL-based representations of biotherapeutics to accelerate protein engineering.

Electronic health records and patient cohorts

We will focus on advancing methods from explainable AI to allow confident use of ML as support system, e.g., in clinical decision making. This will build upon our work on network inference and knowledge graphs and allow human interpretability, e.g., by adding constraints. We will combine this with Bayesian modeling to derive probabilistic statements about the plausibility of facts and apply this to learning representations of unstructured data in combination with control for confounding variables (causality), the validation of decision recommendations, and analyze issues with missing data.