Learning partial differential equations for biological transport models from noisy spatio-temporal data