Machine-learned interatomic potentials have emerged as alternatives to molecular mechanical force fields. These models are trained on quantum mechanical data and can, in principle, match the accuracy of QM methods while remaining sufficiently fast to perform long-timescale MD simulations. To date, popular models have been effective at describing molecular species and single-component liquids. However, they are not always transferable to systems and configurations outside their training domain, and long-range dispersion and electrostatic interactions are not consistently included. Our group has developed a machine-learning implementation of the exchange-hole dipole moment model (MLXDM) to account for dispersion interactions within an MLIP. We have also implemented a fourth-generation, transferable neural network potential for bioorganic molecules composed of C, N, O, and H, incorporating long-range electrostatics via a machine-learning adaptation of the QEq charge-equilibration method. We compare this approach to a delta-learning strategy, in which a neural network corrects a tight-binding DFT method to match results from a full conventional DFT calculation.