1. Research Background

Emergence of deep learning methods has opened new opportunities in extracting quantitative information from biomedical image data at a large scale. In addition to images, these methods allow to quantify patterns in temporal data such as videos and motion capture. Both images and motion recordings have huge variability and noise, deep learning offers the capacity to extract from these data signals that are meaningful for a given biological condition while discarding artifacts and batch effects.

2. Research questions addressed by the group:

In Bozek Lab we develop and apply deep learning methods to quantify visual information in images as well as in video data in biomedical research. Our interests lie at the intersection of machine learning and biomedical applications. Projects in the lab span several branches of medical research: pathology, nephrology, ophthalmology, neurology, as well as basic biology. We are additionally keenly interested in analysis of animal behavior with respect to their genetic and other molecular phenotypes.

3. Possible project(s):

1. Quantification of ageing phenotype of C. elegans
   
   Multitude of genetic mutants of C. elegans have been established, several with a fast-ageing phenotype. Molecular phenotypes of these mutants have been extensively studied, however very little is known about their motion and behavior. In our previous work [1,2] we have established methods for C. elegans segmentation, tracking, and its motion representation. In this project we will apply these methods to study the ageing phenotype of C. elegans. We will investigate whether the biological age of the worm can be established based on its motion patterns. This analysis will be based on the extensive publicly available video data of moving worm (Open Worm Movement Database) as well as collaborations with labs in CECAD. By adapting and applying our previously developed methods we will extract motion patterns of young vs. ageing C. elegans. We will explore not only self-supervised representation learning methods but also supervised approaches aimed at inferring the worm age based on its motion. By implementing interpretability mechanisms into the supervised methods, we will disentangle which movements are specific to the aged vs young individuals. Finally, we will match motion patterns with the underlying genotypes to inspect whether particular mutations have their own specific effects on C. elegans motion.
    
2. Quantification of animal motion with respect to their molecular background
   
   The focus of this project is development of Transformer-based methods for quantification of animal behavior across biological contexts. We have previously demonstrated [2,3] that Transformers applied to either video or keypoint trajectory data and trained in a self-supervised manner are suitable for capturing key aspects of animal motion, including e.g. its directionality, movement amplitude, and asymmetry. In this project we will advance these methods with application to specific biological systems, e.g. mouse with modulated endocannabinoid signaling (in collaboration with Okinawa Institute of Science and Technology), various genetic mutans of zebrafish larvae (in collaboration with Paris Brain Institute), and dominance behavior in zebrafish (in collaboration with Free University Amsterdam). The goal of the analysis is not only to distinguish phenotypes based on their behaviors but also employ interpretability mechanisms that allow to identify motion patterns that are characteristic and unique to the phenotypes. The interpretation of deep learning methods will involve additional quantification of recognizable animal motion features, such as its speed, directionality, variability, in order to gain understandable insights into the behavioral differences.

4. Applied Methods and model organisms:

Deep learning, CNN and Transformer models, interpretability mechanisms such as Grad-CAM or attention rollout. Video and motion capture data from mouse, C. elegans, zebrafish in 2D and 3D.

5. Desirable skills and qualifications:

Good knowledge of python, some experience with deep learning and the use of high-performance computing (HPC)

6. References and key publications:

[1] Deserno M, Bozek K. WormSwin: Instance segmentation of C. elegans using vision transformer. Sci Rep (2023) 13, 11021. https://doi.org/10.1038/s41598-023-38213-7

[2] Deserno M.,  Bozek K., Unsupervised Representation Learning of C. elegans Poses and Behavior Sequences From Microscope Video Recordings (2025) eLife14:RP106593 doi: 10.7554/eLife.106593.1

[3] Rose F, Michaluk M, Blindauer T, Ignatowska-Jankowska BM, O’Shaughnessy L, Stephens GJ, Pereira TD, Uusisaari MY, Bozek K Deep Imputation for Skeleton Data (DISK) for Behavioral Science (2025), https://www.biorxiv.org/content/10.1101/2024.05.03.592173v1 Nature Methods, accepted

[4] Pisula JI, Helbig D, Sancéré L, Persa O-D, Bürger C, Fröhlich A, Lorenz C, Bingmann S, Niebel D, Drexler K, Landsberg J, Thomas R, Bozek K*

, Brägelmann J* (2025) Explainable, federated deep learning model predicts disease progression risk of cutaneous squamous cell carcinoma Nature (2025) njp Precision Oncology, 9, 205. doi.org/10.1038/s41698-025-00997-4

[5] Pisula JI, Bozek K. Efficient WSI classification with sequence reduction and transformers pretrained on text. Sci Rep 15, 5612 (2025). https://doi.org/10.1038/s41598-025-88139-5

[6] Sergei G, Unnersjö-Jess D, Butt L, Benzing T, Bozek K (2024) Self-supervised representation learning of filtration barrier in kidney. Front. Imaging. 3:1339770. doi: 10.3389/fimag.2024.1339770