Counterfactual domain-adaptive machine learning for personalized drug recommendations in cancer treatment

Funded by: DFG: Deutsche Forschungsgemeinschaft

Duration: 10/2025 - 09/2028 (36 months)

Aim of the project

High-throughput perturbation screenings provide a powerful tool to decipher the causal effects of chemical and genetic perturbations on cancer cell lines. Their ability to evaluate a wide spectrum of interventions, from single drugs to intricate drug combinations and CRISPR-interference, has established them as an invaluable resource for the development of novel therapeutic approaches. Nevertheless, (1) the combinatorial complexity of potential interventions makes a comprehensive exploration intractable, and (2) model-derived insights frequently fail to translate to in vivo tumors.
In this project, our aim is to leverage insights from causal modeling, transfer learning, and data augmentation for machine learning-based drug re- purposing. Key to this project will be to aggregate knowledge from multiple drug perturbation data resources with the overall goal to extrapolate the space of interventions to unseen drugs, drug combinations, and dosages thereof. This task will be pursued using single-cell and bulk transcriptomics data of in vitro cancer cell lines. Our models will be further adapted to real tumor measurements for personalized drug-response predictions. To achieve this goal, we will (1) develop algorithms for causal representation learning in order to predict downstream effects of new drugs, drug combinations, and dosages thereof, (2) develop transfer learning approaches leveraging insights from different perturbation experiments, (3) model systematic differences between cancer cell lines and real tumor specimens, and (4) to incorporate compound information into prediction models. Finally, we will combine all these ingredients to develop a general framework to predict the efficacy of novel drug-drug interventions in vitro and in vivo, and will validate our findings in comprehensive experiments.

For more information, please see here.

Workpackages

Background and Workpackages

In this project, we will leverage insights from causal modeling, transfer learning, and data augmentation for ML-based drug re-purposing. Key to this project will be to aggregate knowledge from multiple drug perturbation data resources with the overall goal to extrapolate the space of interventions to unseen drugs, drug combinations, and dosages thereof. This task will be pursued using single cell and bulk transcriptomics data of in vitro cancer cell lines. Our models will be further adapted to real tumor measurements for personalized drug-response predictions. This will be achieved by model invariances, compensating for cellular differences between tumor bulks and cell lines.

Contribution of the Department of Medical Bioinformatics

Our department coordinates subproject WP3, WP4 and together with Prof. Anne-Christin Hausschild WP5.

WP3: Cell-composition invariant drug repurposing

WP4: Incorporate compound information

WP5: Combined XAI model: causal representation learning + transfer learning + cell-composition invariances + compound information

Project Partners

Project partners:

Prof. Dr. Michael Altenbuchinger, 
AG Medical Data Science, Department of Medical Bioninformatics,
University Medical Center Göttingen, Göttingen

Prof. Dr. Anne-Christin Hausschild,
Institut for Predictive Deep Learning for Medicine and Healthcare
Justus Liebig University Giessen, Giessen

Prof. Dr. Günter Schneider,
AG Schneider, Klinik für Allgemein-, Viszeral- und Kinderchirurgie
University Medical Center Göttingen, Göttingen