Interpretable Deep Learning for 3D and 4D Cell Shape Profiling
Loading...
Embargo End Date
2025-09-07
ICR Authors
Authors
De Vries, M
Document Type
Thesis or Dissertation
Date
2025-03-07
Date Accepted
Abstract
A central principle in biology is the relationship between form and function. This connection, evident across all biological scales,
from molecules to whole organisms, underpins the study of cellular morphology, where deviations in shape often indicate disease
states or cellular dysfunction. Cell shape has long been recognised as a key readout for diagnosing conditions and understanding
cellular behaviours. Despite advances in microscopy, which now allow for detailed three-dimensional (3D) and even
four-dimensional (4D) imaging, most profiling efforts remain limited by static, two-dimensional (2D) data that fail to capture the full
complexity of cell structures and their dynamic responses to environmental stimuli.
This thesis aims to bridge that gap by developing interpretable deep learning techniques for 3D and 4D cell shape profiling, focusing
on improving the automation and accuracy of profiling methods while accounting for cell-to-cell heterogeneity. Through the use of
geometric deep learning and multiple instance learning (MIL), the models proposed here are capable of capturing subtle
morphological features that vary across individual cells or evolve over time, offering a more detailed view of cellular states than
traditional methods. In particular, these approaches are applied in the context of drug discovery, where accurate phenotypic
profiling can lead to more predictive models and ultimately accelerate the identification of therapeutic targets.
The contributions of this work include robust models for 3D and 4D cell shape analysis, the incorporation of interpretability into deep
learning predictions, and the application of these models to heterogeneous populations of perturbed cells. The methods developed
here offer new insights into the relationship between cell shape and biological function, demonstrating the potential for deep learning to revolutionise the field of morphological profiling and drug discovery.
Citation
2025
DOI
Source Title
Publisher
Institute of Cancer Research (University Of London)
ISSN
eISSN
Collections
Research Team
Dynamical Cell Systems