B10 - Bayesian deep learning to study non-Gaussianity, correlations, and change-points in cell-driven transport

This experiment-theory project focuses on the random motion of colloidal particles in a bath of
active nonequilibrium agents. As active agents, we use motile Dictyostelium discoideum (D. dis-
coideum) amoeboid cells. They form a ”living carpet”, on which the colloidal cargo particles
are transported. This composite bio-hybrid system can be seen as a versatile laboratory model
to study how foreign bodies interact with a dynamic tissue environment — a key aspect in
many medically relevant settings, such as oral vaccination strategies or the accumulation of
environmental microplastics in the body. In this project, we will concentrate on basic interaction
mechanisms between the cells and the colloidal particles and aim at a quantitative understanding
of the statistical properties of the resulting cell-driven transport process.
Experimentally, we will conduct extensive microscopy recordings using single particle tracking
(SPT), together with a monitoring of the instantaneous numbers of cell-cargo contacts. Three
regimes for this transport can be distinguished according to the cell density in the active carpet:
(i) at low cell densities, a strongly intermittent cargo motion is observed: the cargo remains
immobile when no cell is attached to it, and is mobile with randomly varying directions when a
moving cell is pulling on the cargo; (ii) at intermediate cell densities, two cells may attach to the
cargo at the same time, resulting in a tug-of-war scenario; also here, intermittency of the cargo
motion is expected as transitions between one or two driving cells occur; (iii) at high densities,
when larger numbers of cells simultaneously interact with the cargo particle, the motion will
become less intermittent and more continuous and thus, change points will be more difficult
to detect. For this regime we previously observed that cargo motion is superdiffusive, with a
crossover to normal diffusion. Concurrently, the displacement PDF is non-Gaussian, even in
the long time limit. This behaviour contrasts standard observations of non-Gaussianity in many
biological and soft-matter systems, in which a Gaussian emerges beyond some correlation time.
Our newly recorded data will provide a basis to explore the mechanistic origins of the non-
Gaussianity and the anomalous diffusion by investigating the dynamics of the cell-particle
interactions in a density-dependent manner. 
For the dedicated analysis of the measured data and those from stochastic simulations, we
will evaluate statistical observables such as mean-squared displacement (MSD) or autocovariance
function (ACVF) for the dynamics. Moreover, we will extend our Bayesian-Deep Learning (BDL)
algorithm to include more general stochastic models, plus measurement noises (static and
dynamic). The BDL approach will guide the more detailed modelling, iteratively extending the
starting FLM model and help extracting best values for the model parameters. This BDL suite
will also be the basis for an automated change-point analysis of the experimental measurements,
in particular, to see whether a significant intermittency can be deduced in the high-density limit.
Taken together, the project will shed new light on cell driven transport of colloidal cargo at
different cell densities in the active carpet and lead to a new stochastic model for the description of
the observed motion. Both will inform new research in active transport and stochastic modelling.