Our research is situated at the intersection of computer science and biomedical research, with a strong emphasis on the design an application of data mining and machine learning techniques. We perform fundamental research on robustness and interpretability of machine learning techniques, and focus on applications in biomedicine, translating machine learning applications to the clinic. Many of our applications deal with single-cell technologies, including flow/mass cytometry data, single-cell (multi)omics, and single-cell imaging data types. Clinical applications include allergies and asthma, rheumatology, cancer (lung cancer and leukemia) and primary immune deficiencies (PID).
The core research focus of our group is the development and application of machine learning techniques for single cell bioinformatics. Machine learning techniques are a class of data mining techniques that focus on automated model building in a data-driven fashion, leading to either descriptive or predictive models. Our aim is to design models that are generally applicable to a wide variety of single cell technologies, including classical flow and mass cytometry, imaging flow cytometry, high-content image based screens and single cell “omics” technologies. The following research lines are currently explored:
- Robustness and interpretability of machine learning techniques
An important aspect when applying machine learning techniques to biomedical data is gaining trust in the models and being able to explain them. Our lab is investigating fundamental issues of robustness and interpretability of machine learning models. Especially current techniques like deep learning, which have been shown to perform excellent for e.g. image applications, suffer from problems of robustness (e.g. to adversarial attacks) and interpretability (black box models). Our research aims to characterize and improve robustness of machine learning models and ways to interpret them. From the application point of view, we heavily focus on microscopy image analysis, including challenging data like 3D-electron microscopy, where often labeled data is scarce.
- Single-cell technologies
Our lab develops novel machine learning techniques to interpret large and high-dimensional single-cell data, including flow/mass cytometry, imaging flow cytometry, high-content screening, imaging and single-cell (multi)omics data types. We also pioneer novel single-cell technologies in our own wet lab, providing direct access to large and challenging data types. Our research focuses on developing automated pipelines for quality and preprocessing, visualization and biomarker discovery from high-throughput single-cell data.
- Systems immunology
To unravel the dynamics of the immune system, we develop novel machine learning methods to model dynamic processes, including cell developmental dynamics, spatial organization of tissues and intercellular communication. For all these problems, novel machine learning methods that integrate several data types are developed, and large-scale benchmarking studies are performed to get global insights into the strengths and weaknesses of particular machine learning models for specific types of data.
- Clinical applications
To maximize the benefit of our research to the broad society, we take a double translational approach to machine learning: from theoretical model to application, and further to the clinic. Through collaborations with various national and international medical centers and hospitals, we go the extra mile of implementing our methods into the clinic, evaluating the real impact of the models in the clinical setting. Current clinical applications include leukemia, primary immune deficiencies (PID) and myelodysplastic syndromes (MDS).
Technology Transfer Potential
- Development of novel computational tools to analyze high-throughput data
- Automated hypothesis generation using integrated –omics technologies
- Biomarker discovery
This research is part of the VIB Grand Challenges Program.