Microscopic organisms such as bacteria and algae adapt their motion to environmental signals through run-and-tumble behavior, alternating between directed swimming and reorientation. In this work, we develop a data-driven framework that learns individual-level motility rules from observable input-output responses, without requiring direct knowledge of intracellular biochemical pathways. A neural network is trained to predict how single cells regulate their movement under external stimuli and is validated using bacterial chemotaxis simulations and experimental data from Euglena gracilis.

We then connect this learned microscopic description to population-scale dynamics through a kinetic model. The neural-network-inferred motility law determines the transport structure at the individual level, from which we derive a Keller–Segel-type macroscopic model. In this population model, the effective diffusion and advection coefficients are determined directly by the trained neural network rather than prescribed empirically. This establishes a systematic bridge from single-cell behavioral data to population-level equations, enabling prediction of collective dynamics from learned individual responses.