PhD Defense by Albert Alonso

Title: Mind the Gradient: Differentiable Computational Methods in Microorganism Behaviour Studies

Abstract: At the microscopic scale, where viscosity dominates motion and sensing is constrained by physical limits, microorganisms rely on strategies highly adapted to their environments. This thesis uses differentiable programming techniques to develop computational methods and mathematical models to explore navigation, sensory integration, and behavioural adaptations under such physical constraints.

The first project addresses the challenge of tracking overlapping organisms in microscopy images -- a critical step for understanding behavioural changes. I present a deep learning-based method capable of assigning identities during occlusions and detecting thousands of organisms simultaneously. Trained on synthetic data of simulated nematode motion in viscous environments, the model enables the analysis of massively dense populations.

The second and third studies focus on decision-making during chemotaxis, where physical constraints shape optimal sensing strategies. Using reinforcement learning, I find a continuous transition between temporal and spatial sensing, identifying a regime where integrating both results in a more efficient navigation. Furthermore, spatial sensing in amoebas is modelled as a finite-resource competition between protrusions, revealing how persistence in motion enables cells to overcome the physical limitations in sensing, in agreement with experimental observations.

The fourth work explores the optimal placement of cell-surface receptors for gradient estimation, leading to clustering at the tip of symmetry-breaking protrusions. These findings suggest an evolutionary link between physical mechanisms and functional advantages. For the fifth and last project, the adaptation of biological transport networks is modelled, showing how nodes optimally distribute based on resource delivery costs and environmental constraints.

This thesis showcases how the integration of differentiable computational methods and physical models can advance our understanding of microbial behaviour, and offers a foundation for future applications.