Parsimonious Representation of Nonlinear Dynamical Systems through Manifold Learning: a Chemotaxis Case Study