Modeling positional effects of regulatory sequences with spline transformations increases prediction accuracy of deep neural networks