Reinforcement learning mimicking sperm chemotaxis

Motile biological cells develop versatile taxis strategies to navigate their living environments. Although various biophysical models have been proposed to capture these taxis strategies and explain the underlying biophysical mechanisms, it remains a challenging research question to identify explicit stimulus-response functions that modulate key biophysical parameters of motile cells for robust taxis strategies. Here, we introduce an alternative approach to develop navigation strategies of sperm cells based on reinforcement learning. The learning algorithm obtains effective navigation strategies via tuning the curvature of sperm cells in response to the detected chemical concentration without the need of any explicit stimulus-response functions. The navigation strategies obtained by the reinforcement learning approach are reminiscent to sperm chemotaxis observed in experiments. We also extend our algorithm to consider more complex situations such as the presence of signal noises and fluid shear. This approach opens an alternative avenue to investigate possible variations of biophysical parameters and the necessary complexity of the sensory system required for the development of biologically relevant navigation strategies.